betaFeedback

WorksMathematics

Phase transitions and approximations of mean squared error for state-space models with fractional differencing

Abstract

We study trend estimation in state-space models in which the trend has a fractional stochastic difference of order d>0d>0 and the observation errors form a short-range-dependent stationary process. Using finite-sequence fractional summation and differencing operators, we analyze the penalized least-squares estimator obtained by shrinking the fractional differences of the trend. We derive asymptotic mean squared error (MSE) approximations for all d>0d>0 and identify a sharp phase transition at d=1/2d=1/2. When d>1/2d>1/2, the estimator is consistent and its optimally balanced MSE has order n−(2d−1)/(2d)n^{-(2d-1)/(2d)}. At the boundary d=1/2d=1/2, we obtain a refined finite-sample approximation and show that the MSE decreases at the slower order log⁡log⁡n/log⁡n\log\log n/\log n. When 0<d<1/20<d<1/2, the MSE converges to an explicit positive limit, so consistent recovery of the trend is impossible under the considered scaling. We also describe a practical criterion for choosing the penalty parameter and differencing order, and numerical experiments illustrate the MSE approximations and the behavior of the selection procedure.

This paper is dedicated to the memory of our colleague, Professor Robert H. Shumway (1936–2025), whose vision, mentorship, and lifelong contributions continue to inspire us.

Introduction

Estimating the trend of a time series is a fundamental task in many applications like economic and financial forecasting [22, 30], climate modeling [24], biomedical monitoring [27], and public health surveillance [13]. In the time series analysis literature [3, 9, 12, 28], the following state-space model for the observations YtY_t, 1≤t≤n1 \le t \le n, has received significant attention due to its simplicity and practical interpretability

Yt=μt+εt,t=1,…,n.(1)Y_t = \mu_t + \varepsilon_t,\quad t = 1,\ldots,n. \tag*{(1)}

where {μt}\{\mu_t\} denotes the trend component, and {εt}\{\varepsilon_t\} is assumed to be a zero-mean stationary process with variance σε2\sigma_\varepsilon^2 which is independent of {μt}\{\mu_t\}.

Modeling the trend component in equation (1) can be broadly classified into two categories: deterministic μt\mu_t and stochastic μt\mu_t. Deterministic trends can be modeled either parametrically, when appropriate, or nonparametrically using techniques such as kernel smoothing, local polynomial regression, and sieve expansions. For a detailed review from this perspective, we refer readers to the monograph [14]. We point out that that, in the nonparametric modeling, it is typically assumed that the trend μt\mu_t can be expressed as μ(t/n)\mu(t/n), where μ(⋅)\mu(\cdot) is a smooth function defined on [0,1][0,1]. It is well known that if μ\mu is qq times differentiable and its ppth derivative satisfies a Lipschitz condition of order θ\theta, then the optimal mean squared error (MSE) of a nonparametric estimator of μ(⋅)\mu(\cdot) is of order O(n−2p/(2p+1))O(n^{-2p/(2p+1)}), where p=q+θp=q+\theta [29].

Stochastic trend models are widely used by engineers, statisticians, and economists [2, 12, 22, 28] due to their modeling flexibility and ability to capture long-memory behavior in time series data. In particular, μt\mu_t is modeled using fractional differencing, meaning that the ddth difference of the trend consists of mean-zero i.i.d. (or stationary) random variables, where d>0d>0 is a real number representing the order of differencing [5, 19, 21, 25, 26]. To be more specific, one assumes that for some d>0d>0

∇dμt=γt,(2)\nabla^d\mu_t=\gamma_t, \tag*{(2)}

where ∇\nabla is the difference operator and {γt}\{\gamma_t\} are some mean zero i.i.d. random variables with variance σγ2\sigma_\gamma^2. The model (2) and its variants have been widely used in the literature to model stochastic trends. For example, see [5, 7, 8, 12, 15, 18, 20, 21].

We point out that the system defined by (1) and (2) constitutes a state space model. In the literature, a common approach to obtaining a solution μ=(μ1,⋯ ,μn)⊤\mu=(\mu_1,\cdots,\mu_n)^\top is to minimize the following objective:

∑t=1n(Yt−μt)2+ν∑t=rn(∇dμt)2,ν:=σε2σγ2,(3)\sum_{t=1}^{n}(Y_t-\mu_t)^2+\nu\sum_{t=r}^{n}(\nabla^d\mu_t)^2,\qquad\nu:=\frac{\sigma_\varepsilon^2}{\sigma_\gamma^2}, \tag*{(3)}

where r≡r(d)r\equiv r(d) is an integer depending on dd. In particular, when dd is an integer, one typically chooses r=d+1r=d+1, as in [5]. A key technical advantage of minimizing (3)—as will be demonstrated in Section 2—is that it admits a closed-form solution, which is closely related to the Kalman filter [5, 19]. This connection allows us to analyze its asymptotic properties.

In this paper, we aim to investigate the theoretical performance of the solution obtained by minimizing (3) for the model (1) under the setup in (2) for all d>0d>0. Specifically, we study how the asymptotic behavior and phase transitions of the mean squared error (MSE) of the estimates depend on the parameters dd and ν\nu, where ν\nu is closely related to the signal-to-noise ratio; see Section 2 for a more precise formulation.

We now provide an overview of our main results. We show that, under mild conditions on ν\nu, the asymptotic behavior of the mean squared error (MSE) exhibits a phase transition that can be fully characterized by dd for all d>0d>0. First, as shown in part (1) of Theorem 3.1, when d>1/2d>1/2, we obtain a consistent estimator whose MSE can be of order O(n−(2d−1)/2)O\left(n^{-(2d-1)/2}\right). Second, as stated in part (2) of Theorem 3.1, when d=1/2d=1/2, the estimator from (3) remains consistent, but with a much slower convergence rate of order O(log⁡n/log⁡log⁡n)O\left(\log n/\log\log n\right). Finally, when 0<d<1/20<d<1/2, (3) fails to yield a consistent estimator under the scaling considered here. In this case, we explicitly compute the nontrivial limiting value of the MSE, as presented in part (3) of Theorem 3.1. In addition, when consistent estimators can be obtained by minimizing (3) (i.e., d≥1/2d\geq1/2), we also propose a data-driven approach to select the practically unknown parameters dd and ν\nu, as detailed in Section 4. This method demonstrates superior performance in our numerical studies, as shown in Section 5.

We now discuss several related works in the literature. In [5], the authors studied the problem when d≥1d\geq1 is an integer. They showed that for d∈N+d\in\mathbb{N}^+ and divergent ν\nu, the optimization problem in (3) yields a consistent estimator, and the mean squared error (MSE) can converge at a rate of O(n−(2d−1)/2)O\left(n^{-(2d-1)/2}\right). However, several challenges arise when extending their results to general d>0d>0. For example, one must appropriately choose the parameter rr in (3) and define fractional differences for general d>0d>0 on finite sequences. Additionally, when d<12d<\frac{1}{2}, the estimator becomes inconsistent, and the analysis in [5] does not directly apply. Computing the non-vanishing asymptotic MSE in this case requires new tools. On the other hand, in [7, 8, 19, 20, 25], the authors consider a closely related fractionally integrated process instead of (2), and establish consistency of the parameter estimates under various assumptions on εt\varepsilon_t. However, the asymptotic behavior of the mean squared error (MSE) and its phase transitions have not been investigated within their framework.

Before concluding this section, we offer two remarks on possible future generalizations of our results and techniques. First, in the current paper, we study the performance of the estimators by minimizing (3), which is closely related to the likelihood function of (1) under the i.i.d. Gaussian assumption for {εt}\{\varepsilon_t\} [8]. An alternative to (3), especially when {εt}\{\varepsilon_t\} is not i.i.d., is to consider a weighted least squares estimator that incorporates the covariance structure of {εt}\{\varepsilon_t\}. Analogous results on the phase transitions of the asymptotic MSE have been established in [5] to some extent for d∈N+d \in\mathbb{N}^{+}. We believe our techniques can be extended to general d>0d > 0; see Remark 3.3 for further discussion. Second, although we focus on the stationary setup, we believe our results can be extended to cases where {εt}\{\varepsilon_t\} in (1) or {γt}\{\gamma_t\} in (2) is locally stationary, as considered in [10, 11]. We plan to pursue this direction in future work.

The remainder of the paper is organized as follows. Section 2 introduces the basic setup and presents the explicit solution to the optimization problem (3). Section 3 states our main results. In Section 4, we discuss strategies for selecting the practically unknown parameters ν\nu and dd when consistent estimation is feasible. Numerical simulations are presented in Section 5, and technical proofs are provided in Appendices A and B.

Conventions. For two sequences of deterministic positive values {an}\{a_n\} and {bn}\{b_n\}, we write an=O(bn)a_n = O(b_n) if an≤Cbna_n \le Cb_n for some positive constant C>0C > 0. We write an≍bna_n \asymp b_n if an=O(bn)a_n = O(b_n) and bn=O(an)b_n = O(a_n). Moreover, we write an=o(bn)a_n = o(b_n) if an≤cnbna_n \le c_n b_n for some positive sequence cn↓0c_n \downarrow0. We use N\mathbb{N} to denote the set of natural numbers (including zero), and N+\mathbb{N}^{+} to denote the set of positive natural numbers.

Some background, the assumptions and basic setup

In this section, we provide the basic setup, assumptions and some background on (2) and the closed form solution of (3). We start with introducing an important operator from [4]. For d>0d > 0 and any sequence z=(z1,…,zn)⊤∈Rn\boldsymbol{z} = (z_1, \ldots, z_n)^{\top} \in\mathbb{R}^n, we denote the operator Sd:Rn→Rn\mathcal{S}_d : \mathbb{R}^n \to\mathbb{R}^n that for all 1≤t≤n1 \le t \le n

(Sdz)(t)=∑j=1t(−1)t−j(−dt−j)zj(4)(\mathcal{S}_d\boldsymbol{z})(t) = \sum_{j=1}^{t}(-1)^{t-j}\binom{-d}{t-j}z_j \tag*{(4)}

where (−dt−j)\binom{-d}{t-j} is the generalized binomial coefficient. The operator defined in (4) possesses several important properties, as established in [4], Lemma 1, for example, for any real numbers α,β∈R\alpha,\beta\in\mathbb{R},

S0=I,SαSβ=Sα+β(5)\mathcal{S}_0 = \mathcal{I}, \quad\mathcal{S}_{\alpha}\mathcal{S}_{\beta} = \mathcal{S}_{\alpha+\beta} \tag*{(5)}

which implies that for d>0d > 0

S−d=Sd−1.(6)\mathcal{S}_{-d} = \mathcal{S}_d^{-1}. \tag*{(6)}

We point out that when d>0d > 0, S−d\mathcal{S}_{-d} coincides with the difference operator [16]. Moreover, in matrix form, S−dz\mathcal{S}_{-d}\boldsymbol{z} can be rewritten as

S−dz=Sz,\mathcal{S}_{-d}\boldsymbol{z} = S\boldsymbol{z},

where S∈Rn×nS \in\mathbb{R}^{n \times n} is a lower triangular matrix with entries defined by Stj=(−1)t−j(−dt−j)S_{tj} = (-1)^{t-j}\binom{-d}{t-j} for 1≤t≤n1 \le t \le n and j≤tj \le t.

Denote S(r−1)S^{(r-1)} be the (n+1−r)×n(n+1-r) \times n matrix obtained by deleting the first (r−1)(r-1) rows of SS. With the above definitions and the conventions μ=(μ1,…,μn)T\mu=(\mu_1,\ldots,\mu_n)^{\mathsf{T}} and Y=(Y1,…,Yn)TY=(Y_1,\ldots,Y_n)^{\mathsf{T}}, we can rewrite equation (3) as follows

∥Y−μ∥22+ν∑t=rn[(S−dμ)(t)]2=∥Y−μ∥22+νμT(S(r−1))TS(r−1)μ.(7)\lVert Y-\mu\rVert_2^2+\nu\sum_{t=r}^{n}[(S_{-d}\mu)(t)]^2=\lVert Y-\mu\rVert_2^2+\nu\mu^{\mathsf{T}}(S^{(r-1)})^{\mathsf{T}}S^{(r-1)}\mu. \tag*{(7)}

Minimizing the above quantity with respect to μ\mu, we obtain the estimate

μ^=(I+ν(S(r−1))TS(r−1))−1Y,(8)\widehat{\mu}=\left(I+\nu(S^{(r-1)})^{\mathsf{T}}S^{(r-1)}\right)^{-1}Y, \tag*{(8)}

where II is the n×nn \times n identity matrix. For d>0d>0, we set r=⌈d⌉+1r=\lceil d\rceil+1, where ⌈d⌉\lceil d\rceil denotes the ceiling of dd. Therefore, the main objective of this paper is to study E∥μ^−μ∥22\mathbb{E}\lVert\widehat{\mu}-\mu\rVert_2^2.

Before introducing our assumptions, we first provide some discussion on the model (1). Following the discussion in [5], for μt\mu_t, 1≤t≤n1\le t\le n, satisfying (2), one can solve and express it in an alternative form as follows

μt=∑j=0r−2βr−2−j(Sd−1−j1)(t)+(Sdγ)(t),\mu_t=\sum_{j=0}^{r-2}\beta_{r-2-j}(S_{d-1-j}\mathbf{1})(t)+(S_d\gamma)(t),
:=μ1t+μ2t,(9):=\mu_{1t}+\mu_{2t}, \tag*{(9)}

where 1∈Rn\mathbf{1}\in\mathbb{R}^n is the vector with all entries equal to one, γ=(γ1,…,γn)T\gamma=(\gamma_1,\ldots,\gamma_n)^{\mathsf{T}} and {βj}j=0r−2\{\beta_j\}_{j=0}^{r-2} are constants determined by the initial conditions of the fractional-difference equation. Based on the above decomposition, we can proceed to discuss the overall signal-to-noise ratio (SNR) for the model (1), denoted as

SNR⁡:=E∥μ∥22E∥ε∥22,(10)\operatorname{SNR}:=\frac{\mathbb{E}\lVert\mu\rVert_2^2}{\mathbb{E}\lVert\varepsilon\rVert_2^2}, \tag*{(10)}

where μ=(μ1,…,μn)T\mu=(\mu_1,\ldots,\mu_n)^{\mathsf{T}} and ε=(ε1,…,εn)T\varepsilon=(\varepsilon_1,\ldots,\varepsilon_n)^{\mathsf{T}}. Due to stationarity, we have that E∥ε∥22=nσε2\mathbb{E}\lVert\varepsilon\rVert_2^2=n\sigma_\varepsilon^2. Moreover, by a discussion similar to equation (4) of [5], one can readily obtain that

E∥μ1∥22≍∑j=0r−2βr−2−j2n2(d−j−0.5),E∥Sdγ∥22≲{n2dσγ2d>1/2,(nlog⁡n)σγ2d=1/2,nσγ20<d<1/2,(11)\mathbb{E}\lVert\mu_1\rVert_2^2\asymp\sum_{j=0}^{r-2}\beta_{r-2-j}^2n^{2(d-j-0.5)},\qquad \mathbb{E}\lVert S_d\gamma\rVert_2^2\lesssim \begin{cases} n^{2d}\sigma_\gamma^2 & d>1/2,\\ (n\log n)\sigma_\gamma^2 & d=1/2,\\ n\sigma_\gamma^2 & 0<d<1/2, \end{cases} \tag*{(11)}

where μ1∈Rn\mu_1\in\mathbb{R}^n is defined using the μ1t\mu_{1t}, 1≤t≤n1\le t\le n, in (9).

To avoid repetition, we now summarize the main assumptions of this paper as follows.

Assumption 2.1 Throughout the paper, we assume the following holds for the model (1).

  1. We assume that {εt}\{\varepsilon_t\} is a zero-mean stationary process with 0<σε2<∞0<\sigma_\varepsilon^2<\infty. Furthermore, we assume that for some constant δ>0\delta>0, its autocovariance function satisfies ρ(∣h∣)=O(∣h∣−2−δ)\rho(\lvert h\rvert)=O(\lvert h\rvert^{-2-\delta}) as ∣h∣→∞\lvert h\rvert\to\infty.

  1. We assume that {μt}\{\mu_t\} satisfies (2) with some d>0d>0, where {γt}\{\gamma_t\} are i.i.d. mean-zero random variables, independent of {εt}\{\varepsilon_t\}, with variance σγ2≡σγ2(d,n)\sigma_\gamma^2\equiv\sigma_\gamma^2(d,n), and for some constant τ>0\tau>0, we have

σγ2=τ2(n2d−1−1)/(2d−1),d>0.(12)\sigma_\gamma^2=\frac{\tau^2}{(n^{2d-1}-1)/(2d-1)},\qquad d>0. \tag*{(12)}

*For d=1/2d=1/2, the right-hand side of (12) is defined as the limit as d→1/2d\to1/2. Furthermore, we assume that in (11), E∥μ1∥22/n=O(1)\mathbb{E}\lVert\mu_1\rVert_2^2/n=O(1). First, the assumptions imposed on εt\varepsilon_t ensure that the noise component forms a short-range dependent stationary process—a mild and commonly adopted condition in the literature [5, 8, 9, 12, 18, 19].

Second, (12) is simply a reformulation of the second term in (11) using the Box-Cox transform. It ensures that n−1E∥Sdγ∥22≍1n^{-1}\mathbb{E}\lVert S_d\gamma\rVert_2^2 \asymp1, which, without loss of generality, facilitates a nontrivial discussion of the signal-to-noise ratio (SNR). In fact, we emphasize that the assumptions and scalings imposed here are mild and primarily serve to ensure that the signal-to-noise ratio (SNR) defined in (10) lies in a nontrivial regime, i.e., 0<SNR<∞0 < \mathrm{SNR} < \infty. On the one hand, if SNR approaches zero—indicating that the signal from the trend is completely dominated by noise—then reliable estimation of the trend from {Yt}\{Y_t\} becomes unlikely. This scenario arises when σε2≫E∥μ1∥22/n+σγ2\sigma_\varepsilon^2 \gg\mathbb{E\lVert\mu_1\rVert_2^2}/n+\sigma_\gamma^2. In particular, when the latter term is bounded, the assumption σε2=O(1)\sigma_\varepsilon^2=O(1) rules out this degenerate case. On the other hand, if SNR tends to infinity, {Yt}\{Y_t\} becomes nearly identical to the underlying trend, making it sufficient to use {Yt}\{Y_t\} directly as an estimate. This occurs when σε2≪E∥μ1∥22/n+σγ2\sigma_\varepsilon^2 \ll\mathbb{E\lVert\mu_1\rVert_2^2}/n+\sigma_\gamma^2, i.e., either σε2\sigma_\varepsilon^2 approaches zero or E∥μ1∥22/n\mathbb{E\lVert\mu_1\rVert_2^2}/n or σγ2\sigma_\gamma^2 diverges. Our assumptions are designed to exclude such extreme cases.

Finally, we note that the i.i.d. assumption on {γt}\{\gamma_t\} is not essential. It can be readily relaxed to allow {γt}\{\gamma_t\} to follow a stationary process. We will pursue this direction in future work.

Asymptotic properties of the mean squared error

Main results

We study the MSE of the estimator in (8). The results follow from the bias-variance decomposition in (26) and Proposition (16). We first prepare some notations.

Let ff denote the spectral density of {εt}\{\varepsilon_t\} that

f(ω)=∑h∈Zρ(h)eihω.(13)f(\omega)=\sum_{h\in\mathbb{Z}}\rho(h)e^{\mathrm{i}h\omega}. \tag*{(13)}

For d≥1/2d\ge1/2, we define

c1:=f(0)2dπB(12d,2−12d),(14)c_1:=\frac{f(0)}{2d\pi}B\left(\frac{1}{2d},2-\frac{1}{2d}\right), \tag*{(14)}

where B(⋅,⋅)B(\cdot,\cdot) denotes the beta function. For d>1/2d>1/2, also define

c2:=τ2(2d−1)2dπB(1−12d,1+12d).(15)c_2:=\frac{\tau^2(2d-1)}{2d\pi}B\left(1-\frac{1}{2d},1+\frac{1}{2d}\right). \tag*{(15)}

For 1≤j≤n1\le j\le n, set

xj=πn+1−jn+1,(16)x_j=\pi\frac{n+1-j}{n+1}, \tag*{(16)}

and let

s(u)=2(1−cos⁡u).(17)s(u)=2(1-\cos u). \tag*{(17)}

Moreover, when d=1/2d=1/2, we define the refined approximation

Rnref(ν):=τ2ν2nlog⁡n∑j=1ns(xj)1/2(1+νs(xj)1/2)2+1n∑j=1nf(xj)(1+νs(xj)1/2)2.(18)R_n^{\mathrm{ref}}(\nu):=\frac{\tau^2\nu^2}{n\log n}\sum_{j=1}^{n}\frac{s(x_j)^{1/2}}{\left(1+\nu s(x_j)^{1/2}\right)^2}+\frac{1}{n}\sum_{j=1}^{n}\frac{f(x_j)}{\left(1+\nu s(x_j)^{1/2}\right)^2}. \tag*{(18)}

The main result of the paper is summarized as follows.

Theorem 3.1 Suppose Assumption (4) holds and f(0)>0f(0)>0. Recall ν\nu in (3). Then the following statements hold as n→∞n \to\infty.

(1). If d>1/2d>1/2, ν→∞\nu\to\infty, and ν1/(2d)/n→0\nu^{1/(2d)}/n \to0, then

E∥μ^−μ∥22n=[c1ν−1/(2d)+c2(ν1/(2d)/n)2d−1](1+o(1)).(19)\frac{\mathbb{E}\lVert\widehat{\mu}-\mu\rVert_2^2}{n}=\left[c_1\nu^{-1/(2d)}+c_2\left(\nu^{1/(2d)}/n\right)^{2d-1}\right](1+o(1)). \tag*{(19)}

(2). If d=1/2d=1/2, ν→∞\nu\to\infty and ν/n→0\nu/n \to0, then for Rnref(ν)\mathcal{R}^{\mathrm{ref}}_n(\nu) defined in (18)

E∥μ^−μ∥22n=Rnref(ν)+O(n−1+νnlog⁡n).(20)\frac{\mathbb{E}\lVert\widehat{\mu}-\mu\rVert_2^2}{n}=\mathcal{R}^{\mathrm{ref}}_n(\nu)+O\left(n^{-1}+\frac{\nu}{n\log n}\right). \tag*{(20)}

If, in addition, when log⁡n/(νlog⁡ν)→0\log n/(\nu\log\nu)\to0, then

E∥μ^−μ∥22n=τ2πlog⁡νlog⁡n(1+o(1)),(21)\frac{\mathbb{E}\lVert\widehat{\mu}-\mu\rVert_2^2}{n}=\frac{\tau^2}{\pi}\frac{\log\nu}{\log n}(1+o(1)), \tag*{(21)}

where τ\tau is defined in (12).

(3). If 0<d<1/20<d<1/2 and ν≳1\nu\gtrsim1, then

E∥μ^−μ∥22n=(τ2(1−2d)ν2π∫0π(2−2cos⁡u)d[1+ν(2−2cos⁡u)]d du+1π∫0πf(u)[1+ν(2−2cos⁡u)]2 du)[1+o(1)].(22)\frac{\mathbb{E}\lVert\widehat{\mu}-\mu\rVert_2^2}{n}=\left(\frac{\tau^2(1-2d)\nu^2}{\pi}\int_0^\pi\frac{(2-2\cos u)^d}{[1+\nu(2-2\cos u)]^{d}}\,\mathrm{d}u+\frac{1}{\pi}\int_0^\pi\frac{f(u)}{[1+\nu(2-2\cos u)]^2}\,\mathrm{d}u\right)[1+o(1)]. \tag*{(22)}

Theorem (13) establishes the asymptotic results for the MSE of the estimators. Based on these results, we identify several phase transitions. In particular, minimizing (3) yields consistent estimators when d≥1/2d\ge1/2, and inconsistent ones when d<1/2d<1/2. We elaborate on these results as follows.

First, when d>1/2d>1/2, the two leading terms in (19) are balanced by choosing

ν∗=c1(2d−1)c2n2d−1.\nu^*=\frac{c_1}{(2d-1)c_2}n^{2d-1}.

This choice yields an MSE of order O(n−(2d−1)/(2d))O(n^{-(2d-1)/(2d)}). The rate matches the usual nonparametric rate O(n−2p/(2p+1))O(n^{-2p/(2p+1)}) for a deterministic trend of smoothness p=d−1/2p=d-1/2 [29]. Thus, in this asymptotic sense, the random trend in (2) has effective smoothness d−1/2d-1/2. Second, when d=1/2d=1/2, under Assumption (4), we find that ν≍log⁡n\nu\asymp\log n, which satisfies the additional condition in part (2) of Theorem (13). Consequently, the right-hand side of (21) also vanishes, but at a slower rate of O(log⁡log⁡n/log⁡n)O(\log\log n/\log n) compared to the case when d>1/2d>1/2. In addition, when 0<d<1/20<d<1/2, under Assumption (4), we find that ν≍1\nu\asymp1, so the right-hand side of (22) does not vanish, resulting in an inconsistent estimator.

Before concluding this section, we provide two remarks as follows.

Remark 3.2 We highlight a connection with the perspective in [31], which assumes a differential operator of order d∈N+d\in\mathbb{N}^+

μt=∑j=0d−1θjtj+τ∫0t(t−u)d−1 dB(u),(23)\mu_t=\sum_{j=0}^{d-1}\theta_jt^j+\tau\int_0^t(t-u)^{d-1}\,\mathrm{d}B(u), \tag*{(23)}

where B(u)B(u) is a standard Brownian motion and each θj\theta_j is normally distributed. Wahba [31] showed that when dd is a positive integer, one can obtain an efficient estimator by minimizing the following objective function

∑t=1n(Yt−μt)2+ν∫01(∂dμu∂ud)2du.(24)\sum_{t=1}^{n}(Y_t-\mu_t)^2+\nu\int_{0}^{1}\left(\frac{\partial^d\mu_u}{\partial u^d}\right)^2\mathrm{d}u. \tag*{(24)}

It is well known in [31] that minimizing the above objective function yields the smoothing spline estimator when d∈N+d \in\mathbb{N}^{+}.

The connection between penalizing finite-order differences in (3) and finite-order differentials in (24) was first identified in [5] for d∈N+d \in\mathbb{N}^{+}. We note that our techniques developed in this paper can potentially be applied to analyze (23) and extend the results of [31] by studying the MSE properties for d>1/2d > 1/2 without requiring d∈N+d \in\mathbb{N}^{+}, provided the stochastic integral is well defined in the Itô sense. However, the differential formulation becomes invalid in the Itô framework when d≤1/2d \leq1/2, in which case the problem must be addressed using tools from fractional or Malliavin calculus. In contrast, the difference operator remains well defined even at d=1/2d = 1/2, although it yields a much slower convergence rate. Since this is beyond the scope of the current paper, we will explore this direction in future work.

Remark 3.3 We note that, in addition to minimizing (3)—which leads to the estimator in (8)—an alternative approach is to consider the weighted least squares method, as discussed in Section 2.2 of [5]. Specifically, let Rε:=Cov⁡(ε,ε)R_{\varepsilon} := \operatorname{Cov}(\varepsilon,\varepsilon). Instead of minimizing (7), one may consider minimizing the following objective

(Y−μ)⊤Rε−1(Y−μ)+νμ⊤(S(r−1))⊤S(r−1)μ.(Y-\mu)^{\top}R_{\varepsilon}^{-1}(Y-\mu)+\nu\mu^{\top}(S^{(r-1)})^{\top}S^{(r-1)}\mu.

This yields the following estimator

μ^wls=(Rε−1+ν(S(r−1))⊤S(r−1))−1Rε−1Y.(25)\hat{\mu}_{\mathrm{wls}}=\left(R_{\varepsilon}^{-1}+\nu(S^{(r-1)})^{\top}S^{(r-1)}\right)^{-1}R_{\varepsilon}^{-1}Y. \tag*{(25)}

One can follow and slightly generalize the techniques developed in the proof of Theorem 3.13.1 to establish the phase transitions and MSE approximations for μ^wls\hat{\mu}_{\mathrm{wls}}. In fact, under the assumption that εt\varepsilon_t follows an AR process, the consistency and convergence rate of the MSE for μ^wls\hat{\mu}_{\mathrm{wls}} have been established for d≥1∈N+d \geq1 \in\mathbb{N}^{+} in Theorem 4 of [5]. We believe that analogues of Theorem 3.13.1 can also be derived for μ^wls\hat{\mu}_{\mathrm{wls}} for all d>0d > 0, with distinct transitions occurring at 0<d<1/20 < d < 1/2, d=1/2d = 1/2, and d>1/2d > 1/2. We plan to pursue this direction in future work.

Finally, in practice, RεR_{\varepsilon} is unknown and must be estimated using an estimator with a sufficiently fast convergence rate. This can be effectively achieved when {εt}\{\varepsilon_t\} follows an ARMA model; see the discussion in Section 2.2 of [5] for more details.

Bias-variance argument: proof of Theorem 3.1

In this section, we present the proof of Theorem 3.13.1 along with several key technical components. Our analysis is based on the bias-variance decomposition. Let μˉ=E(μ^∣μ)\bar{\mu}=\mathbb{E}(\hat{\mu}\mid\mu). We begin with the following decomposition

E∥μ^−μ∥22n=E∥μˉ−μ∥22n+E∥μ^−μˉ∥22n=:B+V.(26)\frac{\mathbb{E}\lVert\hat{\mu}-\mu\rVert_2^2}{n} = \frac{\mathbb{E}\lVert\bar{\mu}-\mu\rVert_2^2}{n} + \frac{\mathbb{E}\lVert\hat{\mu}-\bar{\mu}\rVert_2^2}{n} =: \mathcal{B}+\mathcal{V}. \tag*{(26)}

The following proposition gives the bias and variance approximations. Its proof will be given in Appendix fig:A.

Proposition 3.4 Under the respective asymptotic regimes in Theorem 3.1, the bias and variance terms in (26) satisfy

(1). For the bias part B\mathcal{B}, we have that

B={c2(ν1/(2d)n)2d−1[1+o(1)],d>12;τ2ν2nlog⁡n∑j=1ns(xj)1/2[1+νs(xj)1/2]2+O(νnlog⁡n),d=12;τ2(1−2d)ν2π∫0π(2−2cos⁡u)d[1+ν(2−2cos⁡u)d]2 du[1+o(1)],0<d<12.\mathcal{B} = \begin{cases} c_2\left(\dfrac{\nu^{1/(2d)}}{n}\right)^{2d-1}[1+o(1)], & d>\dfrac{1}{2};\\ \dfrac{\tau^2\nu^2}{n\log n}\sum_{j=1}^{n}\dfrac{s(x_j)^{1/2}}{[1+\nu s(x_j)^{1/2}]^2}+O\left(\dfrac{\nu}{n\log n}\right), & d=\dfrac{1}{2};\\ \dfrac{\tau^2(1-2d)\nu^2}{\pi}\int_0^\pi\dfrac{(2-2\cos u)^d}{[1+\nu(2-2\cos u)^d]^2}\,du[1+o(1)], & 0<d<\dfrac{1}{2}. \end{cases}

(2). For the variance part V\mathcal{V}, we have that

V={1n∑j=1nf(xj)[1+νs(xj)1/2]2+O(n−1),d=12;c1ν−1/(2d)[1+o(1)],d>12;1π∫0πf(u)[1+ν(2−2cos⁡u)d]2 du[1+o(1)],0<d<12.\mathcal{V} = \begin{cases} \dfrac{1}{n}\sum_{j=1}^{n}\dfrac{f(x_j)}{[1+\nu s(x_j)^{1/2}]^2}+O(n^{-1}), & d=\dfrac{1}{2};\\ c_1\nu^{-1/(2d)}[1+o(1)], & d>\dfrac{1}{2};\\ \dfrac{1}{\pi}\int_0^\pi\dfrac{f(u)}{[1+\nu(2-2\cos u)^d]^2}\,du[1+o(1)], & 0<d<\dfrac{1}{2}. \end{cases}

Proof of Theorem 3.1. We apply Proposition 3.4 separately in the three regimes. Suppose first that d>1/2d>1/2. Proposition 3.4 gives

B=c2(ν1/(2d)n)2d−1[1+o(1)],V=c1ν−1/(2d)[1+o(1)].\mathcal{B}=c_2\left(\dfrac{\nu^{1/(2d)}}{n}\right)^{2d-1}[1+o(1)],\quad\mathcal{V}=c_1\nu^{-1/(2d)}[1+o(1)].

Since both leading terms are nonnegative, we can therefore obtain that

B+V=[c1ν−1/(2d)+c2(ν1/(2d)n)2d−1][1+o(1)],\mathcal{B}+\mathcal{V}=\left[c_1\nu^{-1/(2d)}+c_2\left(\dfrac{\nu^{1/(2d)}}{n}\right)^{2d-1}\right][1+o(1)],

which proves (19).

Next, let d=1/2d=1/2. Proposition 3.4 yields

B=τ2ν2nlog⁡n∑j=1ns(xj)1/2[1+νs(xj)1/2]2+O(νnlog⁡n),V=1n∑j=1nf(xj)[1+νs(xj)1/2]2+O(n−1).\mathcal{B}=\dfrac{\tau^2\nu^2}{n\log n}\sum_{j=1}^{n}\dfrac{s(x_j)^{1/2}}{[1+\nu s(x_j)^{1/2}]^2}+O\left(\dfrac{\nu}{n\log n}\right),\quad\mathcal{V}=\dfrac{1}{n}\sum_{j=1}^{n}\dfrac{f(x_j)}{[1+\nu s(x_j)^{1/2}]^2}+O(n^{-1}).

Adding these two expressions proves (20). To show (21), by Lemma A.3 and Theorem A.1 of the appendix, we have that

B=τ2πlog⁡νlog⁡n[1+o(1)],V=f(0)πν−1[1+o(1)].\mathcal{B}=\dfrac{\tau^2}{\pi}\dfrac{\log\nu}{\log n}[1+o(1)],\quad\mathcal{V}=\dfrac{f(0)}{\pi}\nu^{-1}[1+o(1)].

Moreover,

VB=O(log⁡nνlog⁡ν)=o(1),\dfrac{\mathcal{V}}{\mathcal{B}}=O\left(\dfrac{\log n}{\nu\log\nu}\right)=o(1),

under the additional condition log⁡n/(νlog⁡ν)→0\log n/(\nu\log\nu)\to0. The remainder in (20) is also negligible relative to log⁡ν/log⁡n\log\nu/\log n, since ν/n→0\nu/n\to0. This establishes (21).

Finally, when 0<d<1/20<d<1/2, Proposition [?] gives

B=τ2(1−2d)ν2π∫0π(2−2cos⁡u)d[1+ν(2−2cos⁡u)d]2 du[1+o(1)],\mathcal{B}=\frac{\tau^2(1-2d)\nu^2}{\pi}\int_0^\pi\frac{(2-2\cos u)^d}{[1+\nu(2-2\cos u)^d]^2}\,du[1+o(1)],

and

V=1π∫0πf(u)[1+ν(2−2cos⁡u)d]2 du[1+o(1)].\mathcal{V}=\frac{1}{\pi}\int_0^\pi\frac{f(u)}{[1+\nu(2-2\cos u)^d]^2}\,du[1+o(1)].

Since ν≍1\nu\asymp1, both leading terms are finite and nonnegative. This proves (22). ■

We now provide a few remarks below.

Remark 3.5 When 0<d<1/20<d<1/2, the nontrivial regime under Assumption 2.1 is ν≍1\nu\asymp1. In this regime, both terms in (22) remain non-negligible, so the MSE does not vanish. Thus minimizing (3) cannot yield a consistent trend estimator in this range of dd.

Remark 3.6 In Proposition [?] and Theorem [31], we analyze the standard MSE, defined as the Euclidean distance between the true random trend μ\mu and its estimate μ^\hat{\mu}. It is also of interest to consider the weighted MSE, i.e., the quadratic form [6] (μ^−μ)′Rε−1(μ^−μ)(\hat{\mu}-\mu)^{\prime}R_{\varepsilon}^{-1}(\hat{\mu}-\mu), where RεR_{\varepsilon} is as in (25). If the spectral density of {εt}\{\varepsilon_t\} is bounded away from zero and infinity — which holds, for example, for the usual ARMA model — then the theory of Toeplitz matrices [17] implies that all eigenvalues of RεR_{\varepsilon} are bounded between two positive constants. In this case, there exist constants k1>0k_1>0 and k2>0k_2>0 such that

k1∥μ^−μ∥2≤(μ^−μ)′Rε−1(μ^−μ)≤k2∥μ^−μ∥2.k_1\lVert\hat{\mu}-\mu\rVert^2\leq(\hat{\mu}-\mu)^{\prime}R_{\varepsilon}^{-1}(\hat{\mu}-\mu)\leq k_2\lVert\hat{\mu}-\mu\rVert^2.

Consequently, all asymptotic convergence rate results established for the standard MSE also hold when the distance is measured by the weighted MSE. We will pursue this direction in the future work.

Practical selection of the parameters

In this section, we develop a practical, data-driven rule for selecting the penalty parameter ν\nu and the differencing order dd when d≥1/2d\geq1/2. For d<1/2d<1/2, the signal component is asymptotically negligible relative to the observation noise, so consistent estimation of dd is not feasible in this setting. The procedure is motivated by the Mallows-type criterion in Section 3 of [5].

Recall the estimator in (8). For each (ν,d)∈(0,∞)×[1/2,∞)(\nu,d)\in(0,\infty)\times[1/2,\infty), let

Hν,d=(I+ν(S(r−1))⊤S(r−1))−1,μ^(ν,d)=Hν,dY,H_{\nu,d}=\left(I+\nu\left(\mathcal{S}^{(r-1)}\right)^{\top}\mathcal{S}^{(r-1)}\right)^{-1},\qquad\hat{\mu}(\nu,d)=H_{\nu,d}Y,

where r=⌈d⌉+1r=\lceil d\rceil+1. Let Gn⊂(0,∞)×[1/2,∞)G_n\subset(0,\infty)\times[1/2,\infty) be a finite candidate grid. In implementation, the candidate orders dd are chosen on a user-specified grid in [1/2,dmax⁡][1/2,d_{\max}], and, for each candidate dd, the candidate penalty parameters ν\nu are chosen on a logarithmic grid over a range determined by the sample size nn.

Let f^(u)\hat{f}(u) be an estimator of the spectral density of {εt}\{\varepsilon_t\}, obtained, for example, from a preliminary trend estimate and a fitted short-range-dependent error model. Such a residual spectral-density estimator is standard; see, for example, [5], Section 3 and [3]. Following [5], our primary criterion is

ϕn(ν,d)=1n∥Y−μ^(ν,d)∥22+2n∑j=1nf^(xj)1+νs(xj)d,(27)\phi_n(\nu,d)=\frac{1}{n}\lVert Y-\hat{\mu}(\nu,d)\rVert_2^2+\frac{2}{n}\sum_{j=1}^{n}\frac{\hat{f}(x_j)}{1+\nu s(x_j)^d}, \tag*{(27)}

where we recall (16) and (17). We then select

(ν^,d^)∈arg⁡min⁡(ν,d)∈Gn ϕn(ν,d).(28)(\hat{\nu},\hat{d}) \in\underset{(\nu,d)\in\mathcal{G}_n}{\arg\min}\ \phi_n(\nu,d). \tag*{(28)}

In our implementation, f^\hat{f} is obtained by fitting autoregressive models to preliminary residuals and selecting the order by BIC.

For large ν\nu, the weights in the correction term concentrate near frequency zero. Thus, when f^(u)\hat{f}(u) is sufficiently smooth near zero, the full-spectrum criterion in (27) has the simpler approximation

ϕn(0)(ν,d)=1n∥Y−μ^(ν,d)∥22+2f^(0)n∑j=1n(1+νs(xj)d)−1.(29)\phi_n^{(0)}(\nu,d) = \frac{1}{n}\lVert\mathbf{Y}-\hat{\boldsymbol{\mu}}(\nu,d)\rVert_2^2 + \frac{2\hat{f}(0)}{n}\sum_{j=1}^{n}\left(1+\nu s(x_j)^d\right)^{-1}. \tag*{(29)}

The full-spectrum criterion is preferred in implementation; (29) is useful as a computational simplification and for asymptotic interpretation.

Before concluding this section, we summarize the implementation of the selection procedure. Let k=⌈n1/2⌉k=\lceil n^{1/2}\rceil. Here, kk is the half-width of the local fitting window, so that the interior window contains 2k+12k+1 observations.

  1. For each t=1,…,nt=1,\ldots,n, define

It={max⁡(1,t−k),…,min⁡(n,t+k)}.\mathcal{I}_t=\{\max(1,t-k),\ldots,\min(n,t+k)\}.

Fit a local linear regression to the observations {Yj:j∈It}\{Y_j:j\in\mathcal{I}_t\} by minimizing ∑j∈It{Yj−a−b(j−t)}2\sum_{j\in\mathcal{I}_t}\{Y_j-a-b(j-t)\}^2, over aa and bb, and let μ^t+\hat{\mu}_t^+ be the resulting intercept estimate a^\hat{a}. This automatically provides the appropriate boundary modification.

  1. Form the preliminary residuals

ϵ~t=Yt−μ^t+,t=1,…,n.\widetilde{\epsilon}_t=Y_t-\hat{\mu}_t^+,\qquad t=1,\ldots,n.
  1. Fit AR⁡(p)\operatorname{AR}(p) models to {ϵ~t}t=1n\{\widetilde{\epsilon}_t\}_{t=1}^{n} over a prespecified collection of orders. Select the order p^\hat{p} using BIC, and let f^\hat{f} denote the spectral density associated with the fitted AR⁡(p^)\operatorname{AR}(\hat{p}) model.

  1. Evaluate the criterion in (27) at every (ν,d)∈Gn(\nu,d)\in\mathcal{G}_n, and let

(ν^,d^)∈arg⁡min⁡(ν,d)∈Gn ϕn(ν,d).(\hat{\nu},\hat{d}) \in\underset{(\nu,d)\in\mathcal{G}_n}{\arg\min}\ \phi_n(\nu,d).

Finally, report μ^(ν^,d^)\hat{\mu}(\hat{\nu},\hat{d}) in (8).

Numerical studies

In this section, we conduct Monte Carlo experiments using the exact finite-sequence fractional operators in (4) and the estimator in (8). Throughout, we set

τ2=σϵ2=1.\tau^2=\sigma_{\epsilon}^2=1.

The innovation sequences {γt}\{\gamma_t\} in (2) and {ϵt}\{\epsilon_t\} in (1) are Gaussian sequences, and σγ2\sigma_{\gamma}^2 is specified according to (12). The penalty parameter is set to its model value ν=σϵ2/σγ2\nu=\sigma_{\epsilon}^2/\sigma_{\gamma}^2, as in (3).

In what follows, we consider different combinations of dd and nn, where d∈{0.25,0.5,0.75,1,1.5}d \in\{0.25, 0.5, 0.75, 1, 1.5\} and n∈{400,600,800,1000,1200,1400,1600,1800,2000,2200}n \in\{400, 600, 800, 1000, 1200, 1400, 1600, 1800, 2000, 2200\}, and generate 500 independent data sets for each combination. In each replication, we construct μ=Sdγ\mu= S_d\gamma and compute μ^\widehat{\mu} from (8). The replication-specific per-coordinate squared error is

MSE⁡n,m=1n∥μ^(m)−μ(m)∥22,m=1,…,500.\operatorname{MSE}_{n,m} = \frac{1}{n}\left\lVert\widehat{\mu}^{(m)}-\mu^{(m)}\right\rVert_2^2,\qquad m=1,\ldots,500.

The empirical MSE reported below is the average of these 500 values.

First, we compare the empirical MSE with the theoretical approximation derived in Theorem 3.1. In this experiment, the differencing order, penalty parameter, and error spectrum are known. Thus, the comparison directly assesses the finite-sample accuracy of the asymptotic risk approximation. The results are reported in Figure 5.1.

Figure 5.1. Empirical MSE (points) and theoretical approximation (dashes).

The theoretical approximations agree closely with the empirical MSEs, and their accuracy generally improves as nn increases, supporting the asymptotic results in Theorem 3.1. Moreover, the figure illustrates the phase transition predicted by the theory. When d>1/2d > 1/2, the risk decreases polynomially with nn. At the boundary d=1/2d = 1/2, the decrease is substantially slower. When d<1/2d < 1/2, the risk approaches a positive limiting value. Across the displayed sample sizes, the theoretical approximations capture these distinct finite-sample behaviors.

Second, we examine the practical selection procedure described in Section 4. We consider data-generating trend orders d∈{0.5,0.75,1}d \in\{0.5, 0.75, 1\} under three error specifications: white noise; an AR(1) process satisfying εt=ϕεt−1+wt\varepsilon_t = \phi\varepsilon_{t-1} + w_t; and an ARMA(1, 1) process satisfying εt=ϕεt−1+θwt−1+wt\varepsilon_t = \phi\varepsilon_{t-1} + \theta w_{t-1} + w_t, where {wt}\{w_t\} is a Gaussian white noise sequence. As outlined at the end of Section 4, for each simulated data set, we first obtain a preliminary local-linear trend estimate. We then fit autoregressive models of orders 0,…,50,\ldots,5 to the resulting residuals, select the order using BIC, and use the spectral density of the selected autoregressive model as f^\widehat{f} in (27). The candidate penalty values are ν∈{0.5,1,1.5,2,2.5}log⁡n\nu\in\{0.5, 1, 1.5, 2, 2.5\}\log n when d=0.5d = 0.5, and ν∈{0.5,1,1.5,2,2.5}n2d−1\nu\in\{0.5, 1, 1.5, 2, 2.5\}n^{2d-1} when d>0.5d > 0.5. For each replication, let (ν^,d^)(\widehat{\nu},\widehat{d}) denote the resulting fitted pair. We compute the empirical MSE of μ^(ν^,d^)\widehat{\mu}(\widehat{\nu},\widehat{d}) in (8) and compare it with the theoretical risk approximation evaluated at the same fitted pair. The latter uses (18) when d^=0.5\widehat{d}=0.5 and the right-hand side of (19) when d^>0.5\widehat{d} > 0.5. The known simulation spectrum is used only to evaluate this theoretical benchmark; the parameter-selection procedure itself uses the residual-based BIC–AR estimate f^\widehat{f}.

Figure 5.2 shows that, after parameter selection, the plug-in theoretical approximation continues to track the empirical risk reasonably well for different values of d≥0.5d \ge0.5. This comparison provides a numerical check that the practical residual-based implementation selects penalties in a range for which the risk approximation remains informative.

Figure 5.2. Empirical MSE (points) and theoretical approximation evaluated at the fitted pair (dashes). Error bars are Monte Carlo 95% confidence intervals.

Proof of Proposition 3.4

Some preliminary results

In this section, we establish auxiliary results used in the proof of Proposition 3.4. The following theorem provides a precise approximation for the bias and variance components in (3.14) whose proof will be given in Appendix B.2. Recall xjx_j and s(⋅)s(\cdot) from (3.4) and (3.5). Recall the spectral density function f(⋅)f(\cdot) in (3.1) and the bias-variance decomposition in (3.14).

Theorem A.1 Suppose the assumptions of Proposition 3.4 hold. Then we have that for all ν>0\nu> 0

E∥μ^−μ‾∥2=∑j=1n(1+νs(xj)d)−2f(xj)+O(1),(30)\mathbb{E}\lVert\widehat{\mu}-\overline{\mu}\rVert^2 = \sum_{j=1}^{n}(1+\nu s(x_j)^d)^{-2}f(x_j)+O(1), \tag*{(30)}

where xjx_j and s(⋅)s(\cdot) are defined in (3.4) and (3.5), respectively. Moreover, we have that

E∥μ‾−μ∥2=σγ2ν∑j=1n(1+νs(xj)d)−2νs(xj)d+O(σγ2ν).(31)\mathbb{E}\lVert\overline{\mu}-\mu\rVert^2 = \sigma_{\gamma}^{2}\nu\sum_{j=1}^{n}(1+\nu s(x_j)^d)^{-2}\nu s(x_j)^d+O(\sigma_{\gamma}^{2}\nu). \tag*{(31)}

The following lemma establishes the asymptotic behavior of the Riemann-type sums in (A.1) and (A.2).

Lemma A.2 Let

hV(z)=1(1+z)2,hB(z)=z(1+z)2,h_{\mathrm{V}}(z)=\frac{1}{(1+z)^2}, \qquad h_{\mathrm{B}}(z)=\frac{z}{(1+z)^2},

and let gg be continuously differentiable on [0,π][0,\pi]. Define

an=πν1/(2d)n+1,bn=πnν1/(2d)n+1,ψ(x)=(sin⁡(x/2)x/2)2d.(32)a_n=\frac{\pi\nu^{1/(2d)}}{n+1},\qquad b_n=\frac{\pi n\nu^{1/(2d)}}{n+1},\qquad\psi(x)=\left(\frac{\sin(x/2)}{x/2}\right)^{2d}. \tag*{(32)}

Then, for every d>0d>0, ν>0\nu>0 and h∈{hV,hB}h\in\{h_{\mathrm{V}},h_{\mathrm{B}}\}

∑j=1nh(νs(xj)d)g(xj)=(n+1)ν−1/(2d)π∫anbnh(w2dψ(ν−1/(2d)w))g(ν−1/(2d)w) dw+O(1).\sum_{j=1}^{n}h(\nu s(x_j)^d)g(x_j)=\frac{(n+1)\nu^{-1/(2d)}}{\pi}\int_{a_n}^{b_n}h(w^{2d}\psi(\nu^{-1/(2d)}w))g(\nu^{-1/(2d)}w)\,\mathrm{d}w+O(1).

Proof. For u∈[0,π]u\in[0,\pi], define

Fν(u)=h(νs(u)d)g(u).F_\nu(u)=h(\nu s(u)^d)g(u).

Since

s(u)=2(1−cos⁡u)=4sin⁡2(u/2),s(u)=2(1-\cos u)=4\sin^2(u/2),

we have

s(u)d=(4sin⁡2(u/2))d=u2d(sin⁡(u/2)u/2)2d.s(u)^d=(4\sin^2(u/2))^d=u^{2d}\left(\frac{\sin(u/2)}{u/2}\right)^{2d}.

Recall (A.3). This yields that s(u)d=u2dψ(u)s(u)^d=u^{2d}\psi(u). Consequently, we see that

n+1π∫π/(n+1)πn/(n+1)h(νs(u)d)g(u) du=(n+1)ν−1/(2d)π∫anbnh(w2dψ(ν−1/(2d)w))g(ν−1/(2d)w) dw.(33)\begin{aligned} \frac{n+1}{\pi}\int_{\pi/(n+1)}^{\pi n/(n+1)}h(\nu s(u)^d)g(u)\,\mathrm{d}u &=\frac{(n+1)\nu^{-1/(2d)}}{\pi}\int_{a_n}^{b_n}h(w^{2d}\psi(\nu^{-1/(2d)}w))g(\nu^{-1/(2d)}w)\,\mathrm{d}w. \tag*{(33)} \end{aligned}

where the equality follows from the change of variables w=ν1/(2d)uw=\nu^{1/(2d)}u.

It remains to justify the replacement of the sum by the integral. The mesh points xj=π(n+1−j)/(n+1)x_j=\pi(n+1-j)/(n+1) have spacing π/(n+1)\pi/(n+1). We use the following elementary Riemann-sum error bound for functions of bounded variation (see Chapter 7 of [1]). If FF has bounded variation on [0,π][0,\pi], then

∣πn+1∑j=1nF(xj)−∫π/(n+1)πn/(n+1)F(u) du∣≤πn+1TV⁡[0,π](F).\left|\frac{\pi}{n+1}\sum_{j=1}^{n}F(x_j)-\int_{\pi/(n+1)}^{\pi n/(n+1)}F(u)\,\mathrm{d}u\right|\leq\frac{\pi}{n+1}\operatorname{TV}_{[0,\pi]}(F).

Consequently,

∣∑j=1nF(xj)−n+1π∫π/(n+1)πn/(n+1)F(u) du∣≤TV⁡[0,π](F).\left|\sum_{j=1}^{n}F(x_j)-\frac{n+1}{\pi}\int_{\pi/(n+1)}^{\pi n/(n+1)}F(u)\,\mathrm{d}u\right|\leq\operatorname{TV}_{[0,\pi]}(F).

We now verify that TV⁡(Fν)\operatorname{TV}(F_\nu) is uniformly bounded. The function ss is increasing on [0,π][0,\pi]. Moreover, hVh_{\mathrm{V}} is decreasing, while hBh_{\mathrm{B}} increases on [0,1][0,1] and decreases on [1,∞)[1,\infty). Hence, for either choice of hh, there exists some constant Ch>0C_h>0,

TV⁡(h(νs(⋅)d))≤Ch,\operatorname{TV}(h(\nu s(\cdot)^d))\leq C_h,

uniformly in ν>0\nu>0. Since gg is continuously differentiable on [0,π][0,\pi], it has bounded variation. The product rule for total variation (see Chapter 6 of [1]) therefore implies

TV⁡(Fν)≤C,\operatorname{TV}(F_\nu)\leq C,

where CC does not depend on nn or ν\nu. Consequently, together with (A.4), we can conclude the proof. ■

Technical proof of the Proposition 3.4

Recall (26). Theorem A.1 gives

ν=1n∑j=1nf(xj)(1+νs(xj)d)2+O(n−1)\nu= \frac{1}{n}\sum_{j=1}^{n}\frac{f(x_j)}{(1+\nu s(x_j)^d)^2}+O(n^{-1})

and

B=σγ2νn∑j=1nνs(xj)d(1+νs(xj)d)2+O(σγ2νn).\mathcal{B} = \frac{\sigma_\gamma^2\nu}{n}\sum_{j=1}^{n}\frac{\nu s(x_j)^d}{(1+\nu s(x_j)^d)^2}+O\left(\frac{\sigma_\gamma^2\nu}{n}\right).

For d>1/2d>1/2, for the variance term, we apply Lemma A.2 with

h(z)=1(1+z)2,g(u)=f(u).(34)h(z)=\frac{1}{(1+z)^2},\qquad g(u)=f(u). \tag*{(34)}

Since an↓0a_n\downarrow0, bn↑∞b_n\uparrow\infty, ψ(ν−1/(2d)w)→1\psi(\nu^{-1/(2d)}w)\to1, and f(ν−1/(2d)w)→f(0)f(\nu^{-1/(2d)}w)\to f(0) for every fixed ww, dominated convergence gives

ν=ν−1/(2d)π∫anbnf(ν−1/(2d)w)(1+w2dψ(ν−1/(2d)w))2 dw+O(n−1)=f(0)ν−1/(2d)π∫0∞1(1+w2d)2 dw(1+o(1))+O(n−1)=c1ν−1/(2d)(1+o(1))+O(n−1).\begin{aligned} \nu=\frac{\nu^{-1/(2d)}}{\pi}\int_{a_n}^{b_n}\frac{f(\nu^{-1/(2d)}w)}{(1+w^{2d}\psi(\nu^{-1/(2d)}w))^2}\,dw+O(n^{-1}) \\ &=\frac{f(0)\nu^{-1/(2d)}}{\pi}\int_{0}^{\infty}\frac{1}{(1+w^{2d})^2}\,dw(1+o(1))+O(n^{-1}) \\ &=c_1\nu^{-1/(2d)}(1+o(1))+O(n^{-1}). \end{aligned}

Similarly, for the bias term, we apply Lemma A.2 with

h(z)=z(1+z)2,g(u)=1.(35)h(z)=\frac{z}{(1+z)^2},\qquad g(u)=1. \tag*{(35)}

Using (12) that

σγ2=τ2(2d−1)n−(2d−1)(1+o(1)),\sigma_\gamma^2=\tau^2(2d-1)n^{-(2d-1)}(1+o(1)),

we obtain

B=σγ2ν1−1/(2d)π∫0∞w2d(1+w2d)2 dw(1+o(1))+O(n−1)=c2(ν1/(2d)n)2d−1(1+o(1))+O(n−1).\begin{aligned} \mathcal{B}=\frac{\sigma_\gamma^2\nu^{1-1/(2d)}}{\pi}\int_{0}^{\infty}\frac{w^{2d}}{(1+w^{2d})^2}\,dw(1+o(1))+O(n^{-1}) \\ &=c_2\left(\frac{\nu^{1/(2d)}}{n}\right)^{2d-1}(1+o(1))+O(n^{-1}). \end{aligned}

Because ν1/(2d)/n→0\nu^{1/(2d)}/n\to0, the remainder O(n−1)O(n^{-1}) is negligible relative to ν−1/(2d)\nu^{-1/(2d)}, and the remainder O(σγ2ν/n)O(\sigma_\gamma^2\nu/n) is negligible relative to σγ2ν1−1/(2d)\sigma_\gamma^2\nu^{1-1/(2d)}. Thus, both additive remainders are absorbed into the respective o(1)o(1) terms. This concludes the proof for d>1/2d>1/2.

For d=1/2d=1/2, the identity σγ2=τ2/log⁡n\sigma_\gamma^2=\tau^2/\log n gives

B=τ2ν2nlog⁡n∑j=1ns(xj)1/2(1+νs(xj)1/2)2+O(νnlog⁡n).\mathcal{B}=\frac{\tau^2\nu^2}{n\log n}\sum_{j=1}^{n}\frac{s(x_j)^{1/2}}{(1+\nu s(x_j)^{1/2})^2}+O\left(\frac{\nu}{n\log n}\right).

Likewise, we have

V=1n∑j=1nf(xj)(1+νs(xj)1/2)2+O(n−1).\mathcal{V} = \frac{1}{n}\sum_{j=1}^{n}\frac{f(x_j)}{(1+\nu s(x_j)^{1/2})^2}+O(n^{-1}).

These are precisely the bias and variance expressions in the proposition. The further asymptotic simplifications of these sums are recorded separately in Lemma A.3.

Finally, let 0<d<1/20<d<1/2, and define

hν(u)=s(u)d(1+νs(u)d)2,vν(u)=f(u)(1+νs(u)d)2.h_\nu(u)=\frac{s(u)^d}{(1+\nu s(u)^d)^2},\qquad v_\nu(u)=\frac{f(u)}{(1+\nu s(u)^d)^2}.

Since ν≍1\nu\asymp1, the functions hνh_\nu and vνv_\nu are uniformly continuously differentiable on [0,π][0,\pi]. By the first-order Euler–Maclaurin formula (see Chapter 7 of [1]), we have that

1n∑j=1nhν(xj)=1π∫0πhν(u) du+o(1),\frac{1}{n}\sum_{j=1}^{n}h_\nu(x_j)=\frac{1}{\pi}\int_{0}^{\pi}h_\nu(u)\,\mathrm{d}u+o(1),

and

1n∑j=1nvν(xj)=1π∫0πvν(u) du+o(1).\frac{1}{n}\sum_{j=1}^{n}v_\nu(x_j)=\frac{1}{\pi}\int_{0}^{\pi}v_\nu(u)\,\mathrm{d}u+o(1).

Since σγ2=τ2(1−2d)(1+o(1))\sigma_\gamma^2=\tau^2(1-2d)(1+o(1)) according to (2.9), we obtain from (A.2) that

B=τ2(1−2d)ν2π∫0πs(u)d(1+νs(u)d)2 du(1+o(1)),\mathcal{B}=\frac{\tau^2(1-2d)\nu^2}{\pi}\int_{0}^{\pi}\frac{s(u)^d}{(1+\nu s(u)^d)^2}\,\mathrm{d}u(1+o(1)),

where we used the fact that the remainder σγ2ν/n\sigma_\gamma^2\nu/n is o(1)o(1). Similarly,

V=1π∫0πf(u)(1+νs(u)d)2 du(1+o(1)).\mathcal{V}=\frac{1}{\pi}\int_{0}^{\pi}\frac{f(u)}{(1+\nu s(u)^d)^2}\,\mathrm{d}u(1+o(1)).

This proves the 0<d<1/20<d<1/2 part and completes the proof of Proposition 3.4.

Boundary asymptotics at d=1/2d=1/2

The case d=1/2d=1/2 is the boundary between the consistent and inconsistent regimes and requires a separate asymptotic argument. We record the required scaled-sum approximations below which is used in the proof of Theorem 3.1.

Lemma A.3 Suppose Assumption 2.1 holds and d=1/2d=1/2, ν→∞\nu\to\infty, and ν/n→0\nu/n\to0. Then

1n∑j=1nf(xj)(1+νs(xj)1/2)2=f(0)πν(1+o(1)),\frac{1}{n}\sum_{j=1}^{n}\frac{f(x_j)}{(1+\nu s(x_j)^{1/2})^2}=\frac{f(0)}{\pi\nu}(1+o(1)),

and

ν2n∑j=1ns(xj)1/2(1+νs(xj)1/2)2=log⁡νπ(1+o(1)).\frac{\nu^2}{n}\sum_{j=1}^{n}\frac{s(x_j)^{1/2}}{(1+\nu s(x_j)^{1/2})^2}=\frac{\log\nu}{\pi}(1+o(1)).

Proof. Recall (32). For d=1/2d = 1/2, we have

ψ(u)=sin⁡(u/2)u/2,an=πνn+1,bn=πnνn+1,\psi(u) = \frac{\sin(u/2)}{u/2}, \qquad a_n = \frac{\pi\nu}{n+1}, \qquad b_n = \frac{\pi n\nu}{n+1},

and s(u)1/2=uψ(u)s(u)^{1/2} = u\psi(u).

We first consider the variance sum. Apply Lemma A.2 with (34), we have that

1n∑j=1nf(xj)(1+νs(xj)1/2)2=n+1nπν∫anbnf(ν−1w)(1+wψ(ν−1w))2 dw+O(n−1).\frac{1}{n}\sum_{j=1}^{n}\frac{f(x_j)}{(1+\nu s(x_j)^{1/2})^2} = \frac{n+1}{n\pi\nu}\int_{a_n}^{b_n}\frac{f(\nu^{-1}w)}{(1+w\psi(\nu^{-1}w))^2}\,\mathrm{d}w + O(n^{-1}).

Since ν/n→0\nu/n \to0 and ν→∞\nu\to\infty, we have that an⟶0a_n \longrightarrow0, bn⟶∞b_n \longrightarrow\infty. Moreover, for every fixed w>0w > 0,

ψ(ν−1w)⟶1,f(ν−1w)⟶f(0).\psi(\nu^{-1}w) \longrightarrow1, \qquad f(\nu^{-1}w) \longrightarrow f(0).

Because ψ\psi is continuous and strictly positive on [0,π][0,\pi], there exists a constant c>0c > 0 such that

ψ(u)≥c,0≤u≤π.\psi(u) \ge c, \qquad0 \le u \le\pi.

Therefore,

0≤f(ν−1w)(1+wψ(ν−1w))2≤∥f∥∞(1+cw)2.0 \le\frac{f(\nu^{-1}w)}{(1+w\psi(\nu^{-1}w))^2} \le\frac{\lVert f\rVert_\infty}{(1+cw)^2}.

The dominating function is integrable on [0,∞)[0,\infty). Hence, by dominated convergence, we have

∫anbnf(ν−1w)(1+wψ(ν−1w))2 dw=f(0)∫0∞1(1+w)2 dw(1+o(1))=f(0)(1+o(1)).(36)\begin{aligned} \int_{a_n}^{b_n}\frac{f(\nu^{-1}w)}{(1+w\psi(\nu^{-1}w))^2}\,\mathrm{d}w &= f(0)\int_{0}^{\infty}\frac{1}{(1+w)^2}\,\mathrm{d}w(1+o(1)) \\ &= f(0)(1+o(1)). \tag*{(36)} \end{aligned}

It follows that

1n∑j=1nf(xj)(1+νs(xj)1/2)2=f(0)πν(1+o(1)).\frac{1}{n}\sum_{j=1}^{n}\frac{f(x_j)}{(1+\nu s(x_j)^{1/2})^2} = \frac{f(0)}{\pi\nu}(1+o(1)).

For the bias part, applying Lemma A.2 with (35) and multiplying the resulting identity by ν/n\nu/n, we obtain

ν2n∑j=1ns(xj)1/2(1+νs(xj)1/2)2=n+1nπ∫anbnwψ(ν−1w)(1+wψ(ν−1w))2 dw+O(νn).\frac{\nu^2}{n}\sum_{j=1}^{n}\frac{s(x_j)^{1/2}}{(1+\nu s(x_j)^{1/2})^2} = \frac{n+1}{n\pi}\int_{a_n}^{b_n}\frac{w\psi(\nu^{-1}w)}{(1+w\psi(\nu^{-1}w))^2}\,\mathrm{d}w + O\left(\frac{\nu}{n}\right).

Since ν/n→0\nu/n \to0, it remains to evaluate the integral. By an argument similar to (36), we have

∫anbnwψ(ν−1w)(1+wψ(ν−1w))2 dw=∫anbnw(1+w)2 dw+O(1).\int_{a_n}^{b_n}\frac{w\psi(\nu^{-1}w)}{(1+w\psi(\nu^{-1}w))^2}\,\mathrm{d}w = \int_{a_n}^{b_n}\frac{w}{(1+w)^2}\,\mathrm{d}w + O(1).

Finally, by direct calculation, we have

∫anbnw(1+w)2 dw=log⁡(1+bn)+11+bn−log⁡(1+an)−11+an=log⁡ν+O(1).\int_{a_n}^{b_n} \frac{w}{(1+w)^2}\,\mathrm{d}w = \log(1+b_n) + \frac{1}{1+b_n} - \log(1+a_n) - \frac{1}{1+a_n} = \log\nu+ O(1).

Consequently, as ν↑∞\nu\uparrow\infty, we have

ν2n∑j=1ns(xj)1/2(1+νs(xj)1/2)2=log⁡νπ(1+o(1)).\frac{\nu^2}{n}\sum_{j=1}^{n}\frac{s(x_j)^{1/2}}{(1+\nu s(x_j)^{1/2})^2} = \frac{\log\nu}{\pi}(1+o(1)).

This completes the proof. ■

Proof of Theorem A.1

Technical preparation

In this section, we collect and prove auxiliary results that will be used in the proof of Theorem A.1. We first collect some results from linear algebra. For a n×nn \times n symmetric matrix AA, we denote its eigenvalues by λ1(A)≥⋯≥λn(A)\lambda_1(A) \ge\cdots\ge\lambda_n(A), and its operator norm by ∥A∥\lVert A\rVert.

Lemma B.1 Let AA and BB be two real symmetric matrices so that A−BA-B has rank at most rr. Then for r+1≤j≤nr+1 \le j \le n the following inequalities hold: (a) λj(A)≤λj−r(B)\lambda_j(A) \le\lambda_{j-r}(B), and (b) λj(B)≤λj−r(A)\lambda_j(B) \le\lambda_{j-r}(A).

Proof. See [23], Theorem 4.3.6. ■

Lemma B.2 Let AA and BB be two n×nn \times n nonnegative definite matrices with the property rank⁡(A−B)≤r\operatorname{rank}(A-B) \le r. Assume that all the eigenvalues of AA and BB are bounded above by a constant c>0c>0, and the values of the integer rr and the constant cc do not depend on nn. Then we have

∑1≤j≤nλj(A)−∑1≤j≤nλj(B)=O(1).\sum_{1\le j\le n}\lambda_j(A)-\sum_{1\le j\le n}\lambda_j(B)=O(1).

Proof. The result follows directly from Lemma B.1. ■

We next collect and prove auxiliary results concerning the matrix SS introduced below (6). The analysis relies on its connection with Toeplitz and Hankel matrices. Recall that a matrix T=(bjk)T=(b_{jk}) is Toeplitz if bjk=bj−kb_{jk}=b_{j-k}. If bj−kb_{j-k} is given by ∫−ππexp⁡{i(j−k)u}f(u) du/(2π)\int_{-\pi}^{\pi}\exp\{i(j-k)u\}f(u)\,\mathrm{d}u/(2\pi), then ff is called the symbol of TT. We write T(f)T(f) for the corresponding infinite-dimensional Toeplitz matrix and Tn(f)T_n(f) for its leading n×nn \times n principal submatrix.

The finite-sample boundary correction arising in the analysis of STSS^{\mathsf T}S has Hankel structure. Accordingly, a matrix H=(bjk)H=(b_{jk}) is called a Hankel matrix if bjk=bj+kb_{jk}=b_{j+k}. If bj+kb_{j+k} is given by ∫−ππexp⁡{i(j+k)u}f(u) du/(2π)\int_{-\pi}^{\pi}\exp\{i(j+k)u\}f(u)\,\mathrm{d}u/(2\pi), then ff is called the symbol of HH. We write H(f)H(f) for the corresponding infinite-dimensional Hankel matrix and Hn(f)H_n(f) for its leading n×nn \times n principal submatrix.

With the above preparation, now we connect SS with the Toeplitz matrix. For d>0d>0, recall that the (t,j)(t,j)th entry of the matrix SS is given by

S(t,j)={(−1)t−j(dt−j),1≤j≤t,0,j>t.S(t,j)= \begin{cases} (-1)^{t-j}\binom{d}{t-j}, & 1\le j\le t,\\ 0, & j>t. \end{cases}

Equivalently, according to the generalized binomial expansion, we can write

S(t,j)=12π∫−ππexp⁡(i(t−j)u)(1−exp⁡(−iu))d du.S(t,j)=\frac{1}{2\pi}\int_{-\pi}^{\pi}\exp(\mathrm{i}(t-j)u)(1-\exp(-\mathrm{i}u))^d\,du.

More precisely, let T(s0d)T(s_0^d) denote the infinite lower triangular Toeplitz operator with symbol s0(u)ds_0(u)^d, where s0(u)=1−exp⁡(−iu)s_0(u)=1-\exp(-\mathrm{i}u). Then SS is the leading n×nn\times n principal submatrix of T(s0d)T(s_0^d). Recall from (17) that s(u)=∣s0(u)∣2s(u)=|s_0(u)|^2. Although the corresponding infinite product has symbol s(u)ds(u)^d, the finite-dimensional matrix S⊤SS^\top S is not exactly Tn(sd)T_n(s^d), because truncation at the sample boundary produces a Hankel-type correction. We formally establish the approximation results as follows.

To state the result, let Wn\mathcal{W}_n denote the flip operator defined by

Wn(z1,…,zn)⊤=(zn,…,z1)⊤,\mathcal{W}_n(z_1,\ldots,z_n)^\top=(z_n,\ldots,z_1)^\top,

for every z=(z1,…,zn)⊤∈Rnz=(z_1,\ldots,z_n)^\top\in\mathbb{R}^n.

Lemma B.3 Let Tn(g)T_n(g) be the Toeplitz matrix with symbol g(u)=∑j∈Zgjexp⁡(−iju)g(u)=\sum_{j\in\mathbb{Z}}g_j\exp(-\mathrm{i}ju), where gj=g−jg_j=g_{-j} for every j∈Zj\in\mathbb{Z} and ∑j∈Z∣jgj∣<∞\sum_{j\in\mathbb{Z}}|jg_j|<\infty. Define

gn(u)=∑∣j∣≤ngjexp⁡(−iju),g_n(u)=\sum_{|j|\leq n}g_j\exp(-\mathrm{i}ju),

and let Rn(g)=(Rn(j,k;g))j,k=1nR_n(g)=(R_n(j,k;g))_{j,k=1}^n, where we recall (16) and denote

Rn(j,k;g)=2n+1∑t=1ng(xt)sin⁡(jxt)sin⁡(kxt),1≤j,k≤n.R_n(j,k;g)=\frac{2}{n+1}\sum_{t=1}^{n}g(x_t)\sin(jx_t)\sin(kx_t),\qquad1\leq j,k\leq n.

Then we have that

∥Tn(g)−Rn(g)−Hn(g)−WnHn(g)Wn∥=O(∥gn−g∥∞).\lVert T_n(g)-R_n(g)-H_n(g)-\mathcal{W}_nH_n(g)\mathcal{W}_n\rVert=O(\lVert g_n-g\rVert_\infty).

Proof. First, suppose that j+k≤nj+k\leq n. By the definitions of the Toeplitz and Hankel entries and the identity cos⁡((j−k)u)−cos⁡((j+k)u)=2sin⁡(ju)sin⁡(ku)\cos((j-k)u)-\cos((j+k)u)=2\sin(ju)\sin(ku), we have

Tn(j−k;gn)−Hn(j+k;gn)=2π∫0πsin⁡(ju)sin⁡(ku)gn(u) du=2n+1∑t=1nsin⁡(jxt)sin⁡(kxt)gn(xt)=Rn(j,k;gn).\begin{aligned} T_n(j-k;g_n)-H_n(j+k;g_n) &=\frac{2}{\pi}\int_{0}^{\pi}\sin(ju)\sin(ku)g_n(u)\,du\\ &=\frac{2}{n+1}\sum_{t=1}^{n}\sin(jx_t)\sin(kx_t)g_n(x_t)\\ &=R_n(j,k;g_n). \end{aligned}

Indeed, the second equality is the discrete sine-transform quadrature formula. Since gng_n is an even trigonometric polynomial of degree at most nn, this formula follows by expanding gng_n in its Fourier series and applying discrete cosine orthogonality on the grid {xt}t=1n\{x_t\}_{t=1}^n. The condition j+k≤nj+k\leq n ensures that no aliasing terms arise in this calculation. Next, suppose that n+1≤j+k≤2nn+1\le j+k\le2n. Applying the same discrete sine-transform identity to the indices n+1−jn+1-j and n+1−kn+1-k gives

Tn(j−k;gn)−Hn(2n+2−j−k;gn)=2π∫0πsin⁡((n+1−j)u)sin⁡((n+1−k)u)gn(u) du=2n+1∑t=1nsin⁡((n+1−j)xt)sin⁡((n+1−k)xt)gn(xt)=2n+1∑t=1nsin⁡(jxt)sin⁡(kxt)gn(xt)=Rn(j,k;gn).\begin{aligned} T_n(j-k;g_n)-H_n(2n+2-j-k;g_n) &=\frac{2}{\pi}\int_0^\pi\sin((n+1-j)u)\sin((n+1-k)u)g_n(u)\,\mathrm{d}u\\ &=\frac{2}{n+1}\sum_{t=1}^n\sin((n+1-j)x_t)\sin((n+1-k)x_t)g_n(x_t)\\ &=\frac{2}{n+1}\sum_{t=1}^n\sin(jx_t)\sin(kx_t)g_n(x_t)\\ &=R_n(j,k;g_n). \end{aligned}

For the penultimate equality, note that the two sine factors acquire the same sign under the replacement j↦n+1−jj\mapsto n+1-j and k↦n+1−kk\mapsto n+1-k, so their product is unchanged. Because gng_n has Fourier coefficients equal to zero outside {−n,…,n}\{-n,\ldots,n\}, Hn(j+k;gn)=0H_n(j+k;g_n)=0 whenever j+k>nj+k>n. Moreover, the (j,k)(j,k)th entry of WnHn(gn)WnW_nH_n(g_n)W_n is Hn(2n+2−j−k;gn)H_n(2n+2-j-k;g_n), which is zero whenever j+k≤n+1j+k\le n+1. Thus, the first identity applies when j+k≤nj+k\le n, both Hankel terms vanish when j+k=n+1j+k=n+1, and the second identity applies when j+k≥n+2j+k\ge n+2. Combining these cases entrywise yields

Tn(gn)−Rn(gn)=Hn(gn)+WnHn(gn)Wn.T_n(g_n)-R_n(g_n)=H_n(g_n)+W_nH_n(g_n)W_n.

It remains to replace gng_n by gg. Put h=g−gnh=g-g_n. For the Hankel term, Hn(h)H_n(h) is a finite compression of the Hankel operator associated with the symbol hh. This operator is obtained by composing multiplication by hh, which has operator norm ∥h∥∞\lVert h\rVert_\infty, with an orthogonal projection and a reflection, both of which have operator norm one. Therefore,

∥Hn(g)−Hn(gn)∥=∥Hn(h)∥≤∥h∥∞=∥g−gn∥∞.\lVert H_n(g)-H_n(g_n)\rVert=\lVert H_n(h)\rVert\le\lVert h\rVert_\infty=\lVert g-g_n\rVert_\infty.

For the sine-basis term, let UnU_n be the orthogonal discrete sine-transform matrix with (j,t)(j,t)th entry (2/(n+1))1/2sin⁡(jxt)(2/(n+1))^{1/2}\sin(jx_t). By the definition of Rn(⋅)R_n(\cdot), Rn(h)R_n(h) is obtained by conjugating the diagonal matrix with diagonal entries h(x1),…,h(xn)h(x_1),\ldots,h(x_n) by UnU_n. Orthogonal conjugation preserves the operator norm. Hence,

∥Rn(g)−Rn(gn)∥=∥Rn(h)∥≤max⁡1≤t≤n∣h(xt)∣≤∥g−gn∥∞.\lVert R_n(g)-R_n(g_n)\rVert=\lVert R_n(h)\rVert\le\max_{1\le t\le n}|h(x_t)|\le\lVert g-g_n\rVert_\infty.

Moreover, WnW_n is an orthogonal and symmetric matrix; that is, WnT=Wn=Wn−1W_n^{\mathsf{T}}=W_n=W_n^{-1}. Therefore, conjugation by WnW_n preserves the operator norm, and hence

∥Wn(Hn(g)−Hn(gn))Wn∥=∥Hn(g)−Hn(gn)∥.\lVert W_n(H_n(g)-H_n(g_n))W_n\rVert=\lVert H_n(g)-H_n(g_n)\rVert.

Furthermore, the (j,k)(j,k)th entry of Tn(g)T_n(g) is gj−kg_{j-k}, where 1≤j,k≤n1\le j,k\le n. Thus, only Fourier coefficients gℓg_\ell with ∣ℓ∣≤n−1|\ell|\le n-1 enter Tn(g)T_n(g). Since gng_n retains all Fourier coefficients of gg with ∣ℓ∣≤n|\ell|\le n, the matrices Tn(g)T_n(g) and Tn(gn)T_n(g_n) have identical entries. Therefore,

Tn(g)=Tn(gn).T_n(g)=T_n(g_n).

Combining this equality with the identity established above for gng_n, the difference in the statement of the lemma is the sum of the differences between the corresponding RnR_n and HnH_n terms. The triangle inequality, together with the preceding bounds, shows that its operator norm is bounded by 3∥g−gn∥∞3\lVert g-g_n\rVert_\infty. This proves the result. ■

The following results provide the key asymptotic expressions used in the proof of Theorem A.1.

Lemma B.4 Let d>0d>0 and ν>0\nu>0. Suppose that g:[0,π]→Rg:[0,\pi]\to\mathbb{R} is of bounded variation, and let Rn(g)R_n(g) be defined as in Lemma B.3. Then, as n→∞n\to\infty,

tr⁡((I+νSTS)−2Rn(g))=∑j=1ng(xj)(1+νs(xj)d)2+O(1),\operatorname{tr}\left((I+\nu S^{\mathsf{T}}S)^{-2}R_n(g)\right)=\sum_{j=1}^{n}\frac{g(x_j)}{(1+\nu s(x_j)^d)^2}+O(1),

and

tr⁡(νSTS(I+νSTS)−2)=∑j=1nνs(xj)d(1+νs(xj)d)2+O(1).\operatorname{tr}\left(\nu S^{\mathsf{T}}S(I+\nu S^{\mathsf{T}}S)^{-2}\right)=\sum_{j=1}^{n}\frac{\nu s(x_j)^d}{(1+\nu s(x_j)^d)^2}+O(1).

Proof. It suffices to prove the first assertion when gg is bounded, nonnegative, and decreasing. The result for a general function of bounded variation then follows by applying the Jordan decomposition and using linearity. Indeed, let

Vg(u)=TV⁡[0,u](g),0≤u≤π,V_g(u)=\operatorname{TV}_{[0,u]}(g),\qquad0\leq u\leq\pi,

and define

g2(u)=Vg(π)−Vg(u)+∥g∥∞,g1(u)=g(u)+g2(u).g_2(u)=V_g(\pi)-V_g(u)+\lVert g\rVert_\infty,\qquad g_1(u)=g(u)+g_2(u).

Then g1g_1 and g2g_2 are bounded, nonnegative, and decreasing on [0,π][0,\pi], and g=g1−g2g=g_1-g_2. Since both Rn(⋅)R_n(\cdot) and the right-hand side of the asserted approximation are linear in gg, the general result follows by applying the decreasing-case result separately to g1g_1 and g2g_2.

Write A=STSA=S^{\mathsf{T}}S, and let d=ℓ+βd=\ell+\beta, where ℓ\ell is a nonnegative integer and 0<β≤10<\beta\leq1. We invoke the finite-rank comparison for fractional-difference operators established in [4]. In the notation used here, this result implies that, for every p∈(0,π)p\in(0,\pi), AA can be compared with

(1−β)s(p)βI+βs(p)β−1Tn(s)ℓ+1,(1-\beta)s(p)^\beta I+\beta s(p)^{\beta-1}T_n(s)^{\ell+1},

up to a symmetric perturbation of rank at most 2ℓ2\ell, where we recall (17).

We first prove the approximation involving Rn(g)R_n(g). The matrix Tn(s)T_n(s) is diagonalized by the discrete sine basis {et}t=1n\{e_t\}_{t=1}^{n} whose entries are {2/(n+1)sin⁡(jxt)}\{\sqrt{2/(n+1)}\sin(jx_t)\}, with eigenvalues s(xt)s(x_t) [17]. Since Rn(g)R_n(g) is diagonalized by the same basis, its eigenvalues are g(xt)g(x_t). For a fixed p∈(0,π)p\in(0,\pi), the comparison from [4], Theorem 6(a) and Theorem 7(b), (d), followed by the interlacing theorem in Theorem B.1, gives the lower bound

λj+2ℓ((I+νA)−1Rn(g)(I+νA)−1)≥[1+ν((1−β)s(p)β+βs(p)β−1s(xn+1−j)ℓ+1)]−2g(xn+1−j),\lambda_{j+2\ell}\left((I+\nu A)^{-1}R_n(g)(I+\nu A)^{-1}\right)\geq\left[1+\nu\left((1-\beta)s(p)^\beta+\beta s(p)^{\beta-1}s(x_{n+1-j})^{\ell+1}\right)\right]^{-2}g(x_{n+1-j}),

for j=1,…,n−2ℓj=1,\ldots,n-2\ell.

To identify the right-hand side, choose p=xn+1−jp=x_{n+1-j}. Since d=ℓ+βd=\ell+\beta, we have

(1−β)s(p)β+βs(p)β−1s(p)ℓ+1=s(p)d.(1-\beta)s(p)^\beta+\beta s(p)^{\beta-1}s(p)^{\ell+1}=s(p)^d.

Moreover, {xt}t=1n\{x_t\}_{t=1}^{n} is decreasing in tt, s(xt)s(x_t) is decreasing in tt, and g(xt)g(x_t) is increasing in tt, because gg is decreasing on [0,π][0,\pi]. Thus, the relevant eigenvalues are ordered as required for the interlacing argument. Summing the preceding inequality over j=1,…,n−2ℓj=1,\ldots,n-2\ell gives

tr⁡((I+νA)−2Rn(g))≥∑j=1n−2ℓg(xn+1−j)(1+νs(xn+1−j)d)2=∑j=1ng(xj)(1+νs(xj)d)2+O(1).\operatorname{tr}\left((I+\nu A)^{-2}R_n(g)\right) \ge\sum_{j=1}^{n-2\ell}\frac{g(x_{n+1-j})}{\left(1+\nu s(x_{n+1-j})^d\right)^2} = \sum_{j=1}^{n}\frac{g(x_j)}{\left(1+\nu s(x_j)^d\right)^2}+O(1).

The omitted 2ℓ2\ell terms are uniformly bounded because gg is bounded. The companion comparison for SSTSS^{\mathsf T} in [4], together with the same interlacing argument, yields the reverse bound

tr⁡((I+νA)−2Rn(g))≤∑j=1ng(xj)(1+νs(xj)d)2+O(1).\operatorname{tr}\left((I+\nu A)^{-2}R_n(g)\right) \le\sum_{j=1}^{n}\frac{g(x_j)}{\left(1+\nu s(x_j)^d\right)^2}+O(1).

This proves the first assertion.

It remains to prove the bias-trace approximation. Since Rn(1)=IR_n(1)=I, the same argument, with g≡1g\equiv1, applies to the two resolvent powers q=1q=1 and q=2q=2, and gives

tr⁡((I+νA)−q)=∑j=1n(1+νs(xj)d)−q+O(1),q=1,2.\operatorname{tr}\left((I+\nu A)^{-q}\right) = \sum_{j=1}^{n}\left(1+\nu s(x_j)^d\right)^{-q}+O(1),\qquad q=1,2.

Finally, the elementary identity

νA(I+νA)−2=(I+νA)−1−(I+νA)−2,\nu A(I+\nu A)^{-2}=(I+\nu A)^{-1}-(I+\nu A)^{-2},

implies that

tr⁡(νA(I+νA)−2)=∑j=1n[11+νs(xj)d−1(1+νs(xj)d)2]+O(1)=∑j=1nνs(xj)d(1+νs(xj)d)2+O(1).\begin{aligned} \operatorname{tr}\left(\nu A(I+\nu A)^{-2}\right) = \sum_{j=1}^{n}\left[\frac{1}{1+\nu s(x_j)^d}-\frac{1}{\left(1+\nu s(x_j)^d\right)^2}\right]+O(1) \\ &= \sum_{j=1}^{n}\frac{\nu s(x_j)^d}{\left(1+\nu s(x_j)^d\right)^2}+O(1). \end{aligned}

This proves the second assertion. ■

Proof of Theorem A.1

For notational simplicity, let S‾=S(r−1)\overline{S}=S^{(r-1)} be the matrix obtained from SS by deleting its first r−1r-1 rows, and define

U‾d=S‾TS‾,Ud=STS.\overline{U}_d=\overline{S}^{\mathsf T}\overline{S},\qquad U_d=S^{\mathsf T}S.

Both U‾d\overline{U}_d and UdU_d are n×nn\times n symmetric matrices. Throughout this proof, ff denotes the spectral density in (13). Recall (8). We can write

Hν,d=(I+νU‾d)−1,μ^=Hν,dY,μ‾=Hν,dμ.H_{\nu,d}=(I+\nu\overline{U}_d)^{-1},\qquad\widehat{\mu}=H_{\nu,d}Y,\qquad\overline{\mu}=H_{\nu,d}\mu.

Since μ^−μ‾=Hν,dε\widehat{\mu}-\overline{\mu}=H_{\nu,d}\varepsilon and E(εεT)=Tn(f)\mathbb{E}(\varepsilon\varepsilon^{\mathsf{T}})=T_n(f), the variance component is

E∥μ^−μ‾∥22=tr⁡(Hν,d2Tn(f)).\mathbb{E}\left\|\widehat{\mu}-\overline{\mu}\right\|_2^2=\operatorname{tr}\left(H_{\nu,d}^2T_n(f)\right).

The matrices U‾d\overline{U}_d and UdU_d differ by a matrix of rank at most r−1r-1. Indeed, the resolvent identity gives

Hν,d−(I+νUd)−1=νHν,d(Ud−U‾d)(I+νUd)−1.H_{\nu,d}-(I+\nu U_d)^{-1}=\nu H_{\nu,d}(U_d-\overline{U}_d)(I+\nu U_d)^{-1}.

Hence, this difference has rank at most r−1r-1. Consequently, Hν,d2−(I+νUd)−2H_{\nu,d}^2-(I+\nu U_d)^{-2} has rank at most 2(r−1)2(r-1). Both resolvents are contractions because U‾d\overline{U}_d and UdU_d are positive semidefinite, while ∥Tn(f)∥≤∥f∥∞\left\|T_n(f)\right\|\leq\left\|f\right\|_\infty because ff is bounded [17]. Thus, the trace of the resulting bounded-rank perturbation is O(1)O(1), and hence

tr⁡(Hν,d2Tn(f))=tr⁡((I+νUd)−2Tn(f))+O(1).\operatorname{tr}\left(H_{\nu,d}^2T_n(f)\right)=\operatorname{tr}\left((I+\nu U_d)^{-2}T_n(f)\right)+O(1).

Let Bν=(I+νUd)−2B_\nu=(I+\nu U_d)^{-2}. By Lemma B.3, we may write

Tn(f)−Rn(f)=Hn(f)+WnHn(f)Wn+En,T_n(f)-R_n(f)=H_n(f)+\mathcal{W}_nH_n(f)\mathcal{W}_n+E_n,

where

∥En∥=O(∥f−fn∥∞).\left\|E_n\right\|=O\left(\left\|f-f_n\right\|_\infty\right).

Assumption 2.1(1) implies that ∑m∈Z∣mfm∣<∞\sum_{m\in\mathbb{Z}}\left|mf_m\right|<\infty. To bound the Hankel term, write

Hn(f)=∑m=22nfmJm.H_n(f)=\sum_{m=2}^{2n}f_mJ_m.

where JmJ_m has ones on its mmth anti-diagonal and zeros elsewhere. Since ∥Jm∥1≤m−1\left\|J_m\right\|_1\leq m-1, we have

∥Hn(f)∥1≤∑m=22n(m−1)∣fm∣≤∑m∈Z∣mfm∣=O(1).\left\|H_n(f)\right\|_1\leq\sum_{m=2}^{2n}(m-1)\left|f_m\right|\leq\sum_{m\in\mathbb{Z}}\left|mf_m\right|=O(1).

Because Wn\mathcal{W}_n is orthogonal,

∥WnHn(f)Wn∥1=∥Hn(f)∥1=O(1).\left\|\mathcal{W}_nH_n(f)\mathcal{W}_n\right\|_1=\left\|H_n(f)\right\|_1=O(1).

Moreover,

n∥f−fn∥∞≤n∑∣m∣>n∣fm∣≤∑∣m∣>n∣mfm∣=o(1).n\left\|f-f_n\right\|_\infty\leq n\sum_{\lvert m\rvert>n}\left|f_m\right|\leq\sum_{\lvert m\rvert>n}\left|mf_m\right|=o(1).

Since ∥Bν∥≤1\left\|B_\nu\right\|\leq1, the trace inequality ∣tr⁡(BνC)∣≤∥Bν∥∥C∥1\left|\operatorname{tr}(B_\nu C)\right|\leq\left\|B_\nu\right\|\left\|C\right\|_1 shows that each Hankel term contributes O(1)O(1). Also,

∣tr⁡(BνEn)∣≤n∥Bν∥∥En∥=o(1).\left|\operatorname{tr}(B_\nu E_n)\right|\leq n\left\|B_\nu\right\|\left\|E_n\right\|=o(1).

Consequently,

tr⁡(BνTn(f))=tr⁡(BνRn(f))+O(1).\operatorname{tr}(B_\nu T_n(f))=\operatorname{tr}(B_\nu R_n(f))+O(1).

Furthermore, (1) of Assumption 2.1 implies ∑h∈Z∣hρ(h)∣<∞\sum_{h\in\mathbb{Z}}\left|h\rho(h)\right|<\infty. Hence the spectral density ff is continuously differentiable on [0,π][0,\pi], and therefore has bounded variation. Thus, the first assertion of Lemma B.4 applies with g=fg=f and gives

E∥μ^−μ‾∥22=∑j=1nf(xj)(1+νs(xj)d)2+O(1),\mathbb{E}\left\|\widehat{\mu}-\overline{\mu}\right\|_2^2=\sum_{j=1}^{n}\frac{f(x_j)}{\left(1+\nu s(x_j)^d\right)^2}+O(1),

which proves (A.1).

We next consider the bias component. Recall (2.6) and S‾μ1=0\overline{S}\mu_1=0 and

S‾Sd=[0In−r+1]\overline{S}S_d = \begin{bmatrix} 0 & I_{n-r+1} \end{bmatrix}

where the zero block has r−1r-1 columns. Consequently, S‾SdSdTS‾T=In−r+1\overline{S}S_dS_d^{\mathsf{T}}\overline{S}^{\mathsf{T}}=I_{n-r+1}. It follows that

μ‾−μ=−νHν,dS‾TS‾Sdγ.\overline{\mu}-\mu=-\nu H_{\nu,d}\overline{S}^{\mathsf{T}}\overline{S}S_d\gamma.

Using E(γγT)=σγ2I\mathbb{E}(\gamma\gamma^{\mathsf{T}})=\sigma_\gamma^2I, we obtain

E∥μ‾−μ∥22=σγ2ν2tr⁡(Hν,dS‾TS‾SdSdTS‾TS‾Hν,d)=σγ2ν2tr⁡(Hν,d2U‾d)=σγ2νtr⁡(Hν,d2νU‾d).\mathbb{E}\left\|\overline{\mu}-\mu\right\|_2^2 =\sigma_\gamma^2\nu^2\operatorname{tr}\left(H_{\nu,d}\overline{S}^{\mathsf{T}}\overline{S}S_dS_d^{\mathsf{T}}\overline{S}^{\mathsf{T}}\overline{S}H_{\nu,d}\right) =\sigma_\gamma^2\nu^2\operatorname{tr}\left(H_{\nu,d}^2\overline{U}_d\right) =\sigma_\gamma^2\nu\operatorname{tr}\left(H_{\nu,d}^2\nu\overline{U}_d\right).

Let qν(x)=νx(1+νx)−2q_\nu(x)=\nu x(1+\nu x)^{-2}. Since Hν,d2νU‾d=qν(U‾d)H_{\nu,d}^2\nu\overline{U}_d=q_\nu(\overline{U}_d), and U‾d−Ud\overline{U}_d-U_d has rank at most r−1r-1, Lemma eq:B.2 and the uniform bound ∥qν∥∞≤1/4\left\|q_\nu\right\|_\infty\leq1/4 imply that

tr⁡(Hν,d2νU‾d)=tr⁡(qν(Ud))+O(1).\operatorname{tr}\left(H_{\nu,d}^2\nu\overline{U}_d\right)=\operatorname{tr}\left(q_\nu(U_d)\right)+O(1).

By Lemma B.4, we have

tr⁡(qν(Ud))=∑j=1nνs(xj)d(1+νs(xj)d)2+O(1).\operatorname{tr}\left(q_\nu(U_d)\right)=\sum_{j=1}^{n}\frac{\nu s(x_j)^d}{(1+\nu s(x_j)^d)^2}+O(1).

Consequently,

E∥μ‾−μ∥22=σγ2ν∑j=1nνs(xj)d(1+νs(xj)d)2+O(σγ2ν),\mathbb{E}\left\|\overline{\mu}-\mu\right\|_2^2 =\sigma_\gamma^2\nu\sum_{j=1}^{n}\frac{\nu s(x_j)^d}{(1+\nu s(x_j)^d)^2}+O(\sigma_\gamma^2\nu),

which proves (A.2).

References

  1. [1]T. M. Apostol. Mathematical Analysis: A Modern Approach to Advanced Calculus. Addison–Wesley, Reading, MA, 2 edition, 1974.
  2. [2]G. E. P. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung. Time Series Analysis: Forecasting and Control. Wiley, Hoboken, NJ, 5th edition, 2015.
  3. [3]P. J. Brockwell and R. A. Davis. Introduction to Time Series and Forecasting. Springer Texts in Statistics. Springer, New York, 3rd edition, 2016.
  4. [4]P. Burman. Approximate singular values of the fractional difference and summation operators. Linear Algebra and its Applications, 416(2):677–687, 2006.DOI
  5. [5]P. Burman and R. H. Shumway. Estimation of trend in state-space models: Asymptotic mean square error and rate of convergence. The Annals of Statistics, 37(6B):3715–3742, 2009.
  6. [6]R. J. Carroll and D. Ruppert. Transformation and Weighting in Regression. Monographs on Statistics and Applied Probability. Chapman and Hall, New York, NY, 1988.
  7. [7]N. H. Chan and W. Palma. State space modeling of long-memory processes. The Annals of Statistics, 26(2):719–740, 1998.
  8. [8]N. H. Chan and W. Palma. Estimation of long-memory time series models: A survey of different likelihood-based methods. In Econometric Analysis of Financial and Economic Time Series, pages 89–121. Emerald Group Publishing Limited, 2006.
  9. [9]J. J. Commandeur and S. J. Koopman. An introduction to state space time series analysis. Oxford university press, 2007.
  10. [10]X. Ding and Z. Zhou. Autoregressive approximations to nonstationary time series with inference and applications. The Annals of Statistics, 51(3):1207–1231, 2023.DOI
  11. [11]X. Ding and Z. Zhou. On the partial autocorrelation function for locally stationary time series: characterization, estimation and inference. Biometrika, 112(2):asaf016, 2025.
  12. [12]J. Durbin and S. J. Koopman. Time Series Analysis by State Space Methods. Oxford Statistical Science Series. Oxford University Press, Oxford, 2nd edition, 2012.
  13. [13]K. B. Ensor, J. C. Schedler, T. Sun, R. Schneider, A. Mulenga, J. Wu, L. B. Stadler, and L. Hopkins. Online trend estimation and detection of trend deviations in sub-sewershed time series of sars-cov-2 rna measured in wastewater. Scientific Reports, 14(1):5575, 2024.
  14. [14]J. Fan and Q. Yao. Nonlinear Time Series: Nonparametric and Parametric Methods. Springer, 2003.
  15. [15]S. Grassi and P. S. de Magistris. When long memory meets the kalman filter: A comparative study. Computational Statistics & Data Analysis, 76:301–319, 2014.
  16. [16]H. L. Gray and N. F. Zhang. On a new definition of the fractional difference. Mathematics of Computation, 50(182):513–529, 1988.
  17. [17]U. Grenander and G. Szegö. Toeplitz Forms and Their Applications, volume 2 of California Monographs in Mathematical Sciences. University of California Press, Berkeley, 1958.
  18. [18]N. Haldrup and M. Ø. Nielsen. Estimation of fractional integration in the presence of data noise. Computational Statistics & Data Analysis, 51(6):3100–3114, 2007.
  19. [19]T. Hartl. Robust trend estimation for strongly persistent data with unobserved memory. Journal of Business & Economic Statistics, 2025.
  20. [20]T. Hartl and R. Jucknewitz. Approximate state space modelling of unobserved fractional components. Econometric Reviews, 41(1):75–98, 2022.
  21. [21]U. Hassler. Time series analysis with long memory in view. John Wiley & Sons, 2018.
  22. [22]M. Hatanaka. Time–Series–Based Econometrics: Unit Roots and Co–integrations. Advanced Texts in Econometrics. Oxford University Press, Oxford, 1996.
  23. [23]R. A. Horn and C. R. Johnson. Matrix Analysis. Cambridge University Press, Cambridge, 1985.
  24. [24]M. Mudelsee. Trend analysis of climate time series: A review of methods. Earth-Science Reviews, 190:310–322, 2019.DOI
  25. [25]M. Ø. Nielsen. Asymptotics for the conditional-sum-of-squares estimator in multivariate fractional time-series models. Journal of Time Series Analysis, 36(2):154–188, 2015.
  26. [26]W. Palma. Long-memory time series: theory and methods. John Wiley & Sons, 2007.
  27. [27]A. Patharkar, F. Cai, F. Al-Hindawi, and T. Wu. Predictive modeling of biomedical temporal data in healthcare applications: review and future directions. Frontiers in Physiology, 15:1386760, 2024.
  28. [28]R. H. Shumway and D. S. Stoffer. Time Series Analysis and Its Applications: With R Examples. Springer Texts in Statistics. Springer, 5th edition, 2025.
  29. [29]C. Stone. Optimal global rates of convergence for nonparametric regression. Annals of Statistics, 10:1040–1053, 1982.
  30. [30]R. S. Tsay. Analysis of Financial Time Series. Wiley Series in Probability and Statistics. Wiley, Hoboken, NJ, 3rd edition, 2010.
  31. [31]G. Wahba. Improper priors, spline smoothing and the problem of guarding against model errors in regression. Journal of the Royal Statistical Society Series B: Statistical Methodology, 40(3):364–372, 1978.DOI

Paper details

vlnc:W_mxjmas9u
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents