Introduction
The accuracy of functional mean estimation depends critically on the number of measurements available within each curve. With sparse temporal observations, estimation is governed by nonparametric smoothing, whereas sufficiently dense observations allow curve-level variation to become the dominant source of uncertainty. For spatially indexed functional data, this transition is further affected by dependence between curves observed at different locations. In particular, increasing the spatial sampling density may substantially increase the total number of measurements without providing a comparable increase in independent information about the underlying functional process. This raises a fundamental question: How do spatial dependence and spatial sampling density affect the sparse-to-dense transition in functional mean estimation? The problem arises naturally in environmental monitoring, where temperature, pollution, and water-quality profiles are observed at spatially dependent locations and often at irregular time points.
For independent functional data, the effect of within-curve sampling is well understood. Sparse longitudinal methods were developed by Yao et al. [41 ] , and Hall et al. [18 ] studied the role of the sampling plan in functional principal component analysis. For functional mean estimation, Cai and Yuan [9 ] established minimax L 2 L^2 L 2 -rates and identified the transition to the root-sample-size rate available from complete trajectories. Zhang and Wang [43 ] provided a unified local-smoothing theory across sparse, semi-dense, and ultra-dense designs. Related developments include simultaneous inference for dense functional means ([10 ] ), optimal sup-norm rates under synchronous designs ([2 ] ), non-asymptotic high-dimensional phase transitions ([17 ] ), and simultaneous inference across sampling regimes ([8 ] ). Together these results show how the number and placement of measurements within a curve determine whether the estimate of mean behaves as a nonparametric or parametric object.
Dependence across curves introduces a second source of information loss. Functional time-series methods describe linear and weakly dependent processes ([3 , 19 ] ), mean inference ([21 , 22 , 11 , 29 ] ), frequency-domain structure ([34 ] ), high-dimensional panels ([44 ] ), and locally stationary processes ([40 , 15 ] ). Sparse-to-dense transitions have also been studied for structural-break detection and relevant-hypothesis testing in discretely observed functional time series ([6 , 7 ] ). Spatial functional data, however, are indexed by pairing with multidimensional geometry rather than an ordered temporal lag. Functional kriging, spatial functional statistics, and spatially dependent principal components have been developed by Delicado et al. [13 ] , Giraldo et al. [16 ] , Martínez-Hernández and Genton [30 ] , and Zhang and Li [42 ] ; see also the broader second-generation framework of Koner and Staicu [25 ] .
While the sparse-to-dense transition is well understood for independent functional observations, much less is known about how it changes under spatial dependence and simultaneous infill and increasing-domain sampling. A central difficulty is that spatial replication generates two different information scales. Additional nearby sites contribute new temporal observations, but their underlying functional trajectories may contain largely redundant information. Consequently, the total number of measurements and the effective number of spatial curves can grow at substantially different rates. As we demonstrate below, this distinction changes the temporal sampling density at which the leading estimation error switches from an observation-limited to a curve-limited regime. Understanding and quantifying this shift is the principal objective of the present paper.
At sites S i = ( S i , 1 , … , S i , d ) ⊤ ∈ R d S_i = (S_{i,1}, \ldots, S_{i,d})^\top\in\mathbb{R}^d S i = ( S i , 1 , … , S i , d ) ⊤ ∈ R d , 1 ≤ i ≤ m n 1 \le i \le m_n 1 ≤ i ≤ m n , and times T i j ∈ [ 0 , 1 ] T_{ij} \in[0,1] T ij ∈ [ 0 , 1 ] , we observe
Y i j = η ( S i , T i j ) + σ ( S i , T i j ) ϵ i j , 1 ≤ i ≤ m n , 1 ≤ j ≤ N i , (1) Y_{ij} = \eta(S_i,T_{ij}) + \sigma(S_i,T_{ij})\epsilon_{ij}, \qquad1 \le i \le m_n,\qquad1 \le j \le N_i, \tag*{(1)} Y ij = η ( S i , T ij ) + σ ( S i , T ij ) ϵ ij , 1 ≤ i ≤ m n , 1 ≤ j ≤ N i , (1)
where η ( s , t ) = m ( s , t ) + X ( s , t ) \eta(s,t) = m(s,t) + X(s,t) η ( s , t ) = m ( s , t ) + X ( s , t ) and E { X ( s , t ) } = 0 \mathbb{E}\{X(s,t)\} = 0 E { X ( s , t )} = 0 . The random field X X X represents spatially dependent curve-to-curve variation and ϵ i j \epsilon_{ij} ϵ ij is an independent measurement error. We consider both a stationary mean m ( t ) = m ( s , t ) m(t) = m(s,t) m ( t ) = m ( s , t ) , shared by all sites, and a locally stationary mean surface M ( u , t ) M(u,t) M ( u , t ) , where M ( u , t ) M(u,t) M ( u , t ) approximates the spatially varying mean m ( s , t ) m(s,t) m ( s , t ) after appropriately rescaling the expanding spatial domain to the cube [ 0 , 1 ] d [0,1]^d [ 0 , 1 ] d ; see Kurisu [26 ] . This decomposition separates spatially dependent curve-level variation from observation-level variation generated by random sampling times and measurement errors, and this distinction is reflected explicitly in the convergence rates.
Our main contribution is to characterize how spatial dependence and the spatial sampling density affect the sparse-to-dense transition in functional mean estimation. We establish dependence-adjusted uniform convergence rates that reveal two distinct information scales. The observation-level component is governed by the nominal number of sampled measurements, whereas the curve-level component is governed by a dependence-adjusted effective sample size m e f f , n = m n / Θ n m_{\mathrm{eff},n} = m_n/\Theta_n m eff , n = m n / Θ n , where Θ n \Theta_n Θ n quantifies the inflation caused by spatial dependence. The balance between these two components determines the temporal sampling density at which the dominant source of estimation error changes. Under increasing-domain sparse-space sampling, the effective sample size is of the same order as the nominal number of curves, recovering the familiar sparse-to-dense transition. Under spatial infill with persistent local dependence, however, the effective sample size is governed by the spatial domain volume and can therefore be substantially smaller than the nominal number of curves. Consequently, additional spatial observations may continue to improve the observation-level component while contributing comparatively little independent curve-level information. This shifts the temporal phase-transition boundary and reveals an interaction between spatial and temporal sampling that cannot be captured by simply replacing the nominal sample size with an effective sample size. For spatially varying means, spatial localization additionally affects the transition through the kernel-window volume or the spatial-spline resolution.
To develop this theory for different mean structures, we consider three spline-based estimators. A pooled least-squares estimator recovers the stationary mean, a local-kernel estimator recovers the mean curve at an interior spatial point, and a tensor-product spatial spline estimates the full mean surface. Their linear representations reduce the stochastic analysis to empirical Gram matrices and spatially dependent site-level scores and allow the same distinction between observation-level and curve-level information to be traced across the three estimation problems.
We complement the theoretical results by a simulation study that varies sample size, spatial geometry and temporal sampling density, together with additional distributional robustness experiments. Three environmental applications illustrate the estimators under regular-grid, irregular-network, and highly time-sparse sampling designs.
Statistical framework and methodology
We consider the model ( 1.1 ) (1.1) ( 1.1 ) and assume that the spatial locations S i S_i S i are fixed and the within-site sample sizes N i N_i N i are treated as fixed whereas the observation times T i j T_{ij} T ij and the measurement errors ε i j \varepsilon_{ij} ε ij are random. For each fixed site, ( ω , t ) ↦ X ( s , t ; ω ) (\omega,t) \mapsto X(s,t;\omega) ( ω , t ) ↦ X ( s , t ; ω ) is jointly measurable with respect to the σ \sigma σ -field generated by that curve and the Borel sets of [ 0 , 1 ] [0,1] [ 0 , 1 ] . More specifically, we make the following assumptions regarding the measurement locations and errors in model ( 1.1 ) (1.1) ( 1.1 ) .
Assumption 2.1 (Time design). The N i N_i N i define the number of measurements of each location. The variables { T i j : 1 ≤ i ≤ m n , 1 ≤ j ≤ N i } \{T_{ij}: 1 \le i \le m_n, 1 \le j \le N_i\} { T ij : 1 ≤ i ≤ m n , 1 ≤ j ≤ N i } are i.i.d. with density f T f_T f T on [ 0 , 1 ] [0,1] [ 0 , 1 ] satisfying 0 < c f ≤ f T ( t ) ≤ C f < ∞ 0 < c_f \le f_T(t) \le C_f < \infty 0 < c f ≤ f T ( t ) ≤ C f < ∞ for all t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , and the entire collection of time points is independent of the latent random field X X X . The errors ε i j \varepsilon_{ij} ε ij are i.i.d., centered, variance one, and independent of the latent field X X X and of the { T i j } \{T_{ij}\} { T ij } .
Assumption 2.2 (Infill-increasing spatial design). For each n n n , let L 1 , n , … , L d , n > 0 L_{1,n},\ldots,L_{d,n}>0 L 1 , n , … , L d , n > 0 and S n = ∏ r = 1 d [ 0 , L r , n ] S_n = \prod_{r=1}^{d}[0,L_{r,n}] S n = ∏ r = 1 d [ 0 , L r , n ] , the spatial domain with volume, V n = ∣ S n ∣ = ∏ r = 1 d L r , n V_n = |S_n| = \prod_{r=1}^{d}L_{r,n} V n = ∣ S n ∣ = ∏ r = 1 d L r , n , and L n = V n 1 / d L_n = V_n^{1/d} L n = V n 1/ d . For δ n > 0 \delta_n>0 δ n > 0 , partition R d \mathbb{R}^d R d into cubes Δ n , ℓ = ∏ r = 1 d [ ℓ r δ n , ( ℓ r + 1 ) δ n ) \Delta_{n,\ell} = \prod_{r=1}^{d}[\ell_r\delta_n,(\ell_r+1)\delta_n) Δ n , ℓ = ∏ r = 1 d [ ℓ r δ n , ( ℓ r + 1 ) δ n ) , ℓ = ( ℓ 1 , … , ℓ d ) ∈ Z d \ell=(\ell_1,\ldots,\ell_d)\in\mathbb{Z}^d ℓ = ( ℓ 1 , … , ℓ d ) ∈ Z d , and set I n = { ℓ ∈ Z d : Δ n , ℓ ∩ S n ≠ ∅ } \mathcal{I}_n=\{\ell\in\mathbb{Z}^d:\Delta_{n,\ell}\cap S_n\ne\varnothing\} I n = { ℓ ∈ Z d : Δ n , ℓ ∩ S n = ∅ } . Each ℓ ∈ I n \ell\in\mathcal{I}_n ℓ ∈ I n contains exactly one spatial location S ℓ ∈ Δ n , ℓ ∩ S n S_\ell\in\Delta_{n,\ell}\cap S_n S ℓ ∈ Δ n , ℓ ∩ S n . Thus { S i } i = 1 m n = { S ℓ : ℓ ∈ I n } \{S_i\}_{i=1}^{m_n}=\{S_\ell:\ell\in\mathcal{I}_n\} { S i } i = 1 m n = { S ℓ : ℓ ∈ I n } and m n = ∣ I n ∣ ≍ V n δ n − d m_n=|\mathcal{I}_n|\asymp V_n\delta_n^{-d} m n = ∣ I n ∣ ≍ V n δ n − d .
Note that V n V_n V n measures the size of the observation domain, whereas δ n \delta_n δ n controls the spatial sampling resolution. Throughout the asymptotic theory, we consider an increasing-domain, that is V n → ∞ V_n\to\infty V n → ∞ , and distinguish two spatial regimes: (i) if δ n ≍ 1 \delta_n\asymp1 δ n ≍ 1 we refer to this as an sparse space regime; (ii) if δ n → 0 \delta_n\to0 δ n → 0 we refer to this as a dense space regime.
Assumption 2.3 (Full-dimensional domain growth). There exist constants 0 < c L < C L < ∞ 0<c_L<C_L<\infty 0 < c L < C L < ∞ such that c L ≤ L r , n / L n ≤ C L c_L\le L_{r,n}/L_n\le C_L c L ≤ L r , n / L n ≤ C L for every 1 ≤ r ≤ d 1\le r\le d 1 ≤ r ≤ d and all sufficiently large n n n .
Let M 1 , n = ∑ i = 1 m n N i M_{1,n}=\sum_{i=1}^{m_n}N_i M 1 , n = ∑ i = 1 m n N i denote the number of all observations and M 2 , n = ∑ i = 1 m n N i ( N i − 1 ) M_{2,n}=\sum_{i=1}^{m_n}N_i(N_i-1) M 2 , n = ∑ i = 1 m n N i ( N i − 1 ) , and set N ‾ n = M 1 , n / m n \overline{N}_n=M_{1,n}/m_n N n = M 1 , n / m n , and N ‾ 2 , n = M 2 , n / m n \overline{N}_{2,n}=M_{2,n}/m_n N 2 , n = M 2 , n / m n . Thus N ‾ n \overline{N}_n N n is the average number of observations per curve; its growth relative to the dependence-adjusted effective size determines the time-sampling regime. For the estimation of the mean function m : R d × [ 0 , 1 ] → R m:\mathbb{R}^d\times[0,1]\to\mathbb{R} m : R d × [ 0 , 1 ] → R , defined by m ( s , t ) = E [ η ( s , t ) ] m(s,t)=\mathbb{E}[\eta(s,t)] m ( s , t ) = E [ η ( s , t )] , we will distinguish the cases of (spatial) stationarity , that is m ( s , t ) = m ( t ) m(s,t)=m(t) m ( s , t ) = m ( t ) (for all s s s ) and a locally stationary mean that can be approximated, after rescaling the spatial domain, by a smooth surface M ( u , t ) M(u,t) M ( u , t ) , see [26 ] .
Our approach is based on a B-spline estimator. More specifically, we fix an integer p ≥ 2 p \ge2 p ≥ 2 , and let B J ( t ) = ( B 1 , J ( t ) , … , B J + p , J ( t ) ) T \mathbf{B}_{J}(t) = (B_{1,J}(t), \ldots, B_{J+p,J}(t))^{\mathsf{T}} B J ( t ) = ( B 1 , J ( t ) , … , B J + p , J ( t ) ) T denote the normalized B-spline basis of order p p p on a quasi-uniform partition of the interval [ 0 , 1 ] [0,1] [ 0 , 1 ] with J J J interior knots. Throughout this paper, J m = J m n J_m = J_{m_n} J m = J m n , depends on m n m_n m n , denotes the time-direction spline resolution, and we use the notation B m ( t ) = B J m ( t ) \mathbf{B}_m(t) = \mathbf{B}_{J_m}(t) B m ( t ) = B J m ( t ) where the dimension of this vector is K m = J m + p K_m = J_m + p K m = J m + p . For further details on B-spline theory and its use in regression, see, for example, de Boor [12 ] .
Stationary mean
In this section, we consider the model (1) with a stationary mean, that is m ( s , t ) = m ( t ) m(s,t) = m(t) m ( s , t ) = m ( t ) , for all s ∈ R d s \in\mathbb{R}^{d} s ∈ R d . The spatial locations do not enter the target, so observations from all sites can be pooled to estimate a single temporal curve.
Accordingly, pooling all observations gives the least-squares projection onto the time-spline space, that is
m ^ J m ( t ) = B m ( t ) T β ^ m , J m , (2) \widehat{m}_{J_m}(t) = \mathbf{B}_m(t)^{\mathsf{T}}\widehat{\boldsymbol{\beta}}_{m,J_m}, \tag*{(2)} m J m ( t ) = B m ( t ) T β m , J m , (2)
where
β ^ m , J m : = arg min β ∈ R K m ∑ i = 1 m n ∑ j = 1 N i { Y i j − B m ( T i j ) T β } 2 . \widehat{\boldsymbol{\beta}}_{m,J_m} := \underset{\boldsymbol{\beta}\in\mathbb{R}^{K_m}}{\operatorname{arg\,min}} \sum_{i=1}^{m_n}\sum_{j=1}^{N_i} \left\{Y_{ij}-\mathbf{B}_m(T_{ij})^{\mathsf{T}}\boldsymbol{\beta}\right\}^{2}. β m , J m := β ∈ R K m arg min i = 1 ∑ m n j = 1 ∑ N i { Y ij − B m ( T ij ) T β } 2 .
Equivalently, β ^ m , J m = Q ^ m , J m − 1 r ^ m , J m \widehat{\boldsymbol{\beta}}_{m,J_m} = \widehat{Q}_{m,J_m}^{-1}\widehat{\boldsymbol{r}}_{m,J_m} β m , J m = Q m , J m − 1 r m , J m , where the empirical Gram matrix Q ^ m , J m \widehat{Q}_{m,J_m} Q m , J m and the vector r ^ m , J m \widehat{\boldsymbol{r}}_{m,J_m} r m , J m are given by
Q ^ m , J m : = 1 M 1 , n ∑ i = 1 m n ∑ j = 1 N i B m ( T i j ) B m ( T i j ) T , r ^ m , J m : = 1 M 1 , n ∑ i = 1 m n ∑ j = 1 N i B m ( T i j ) Y i j . \widehat{Q}_{m,J_m} := \frac{1}{M_{1,n}}\sum_{i=1}^{m_n}\sum_{j=1}^{N_i}\mathbf{B}_m(T_{ij})\mathbf{B}_m(T_{ij})^{\mathsf{T}}, \qquad\widehat{\boldsymbol{r}}_{m,J_m} := \frac{1}{M_{1,n}}\sum_{i=1}^{m_n}\sum_{j=1}^{N_i}\mathbf{B}_m(T_{ij})Y_{ij}. Q m , J m := M 1 , n 1 i = 1 ∑ m n j = 1 ∑ N i B m ( T ij ) B m ( T ij ) T , r m , J m := M 1 , n 1 i = 1 ∑ m n j = 1 ∑ N i B m ( T ij ) Y ij .
The corresponding population projection is given by m J m ( t ) : = B m ( t ) T β m , J m m_{J_m}(t) := \mathbf{B}_m(t)^{\mathsf{T}}\boldsymbol{\beta}_{m,J_m} m J m ( t ) := B m ( t ) T β m , J m , where
β m , J m : = Q m , J m − 1 r m , J m , Q m , J m : = E { B m ( T ) B m ( T ) T } , r m , J m : = E { B m ( T ) m ( T ) } , \boldsymbol{\beta}_{m,J_m} := Q_{m,J_m}^{-1}\boldsymbol{r}_{m,J_m}, \qquad Q_{m,J_m} := \mathbb{E}\left\{\mathbf{B}_m(T)\mathbf{B}_m(T)^{\mathsf{T}}\right\}, \qquad\boldsymbol{r}_{m,J_m} := \mathbb{E}\left\{\mathbf{B}_m(T)m(T)\right\}, β m , J m := Q m , J m − 1 r m , J m , Q m , J m := E { B m ( T ) B m ( T ) T } , r m , J m := E { B m ( T ) m ( T ) } ,
and T T T has the density f T f_T f T specified in Assumption 2.1. This projection satisfies the orthogonality condition
E [ B m ( T ) { m ( T ) − m J m ( T ) } ] = 0. (3) \mathbb{E}\left[\mathbf{B}_m(T)\{m(T)-m_{J_m}(T)\}\right] = 0. \tag*{(3)} E [ B m ( T ) { m ( T ) − m J m ( T )} ] = 0. (3)
Note that the normal equations give an exact centering for the stochastic expansion. More precisely, with the notation
ξ i , J m : = ∑ j = 1 N i B m ( T i j ) { Y i j − m J m ( T i j ) } ∈ R K m \xi_{i,J_m} := \sum_{j=1}^{N_i}\mathbf{B}_m(T_{ij})\{Y_{ij}-m_{J_m}(T_{ij})\} \in\mathbb{R}^{K_m} ξ i , J m := j = 1 ∑ N i B m ( T ij ) { Y ij − m J m ( T ij )} ∈ R K m
we have, by Assumption 2.1 and (3) , E ξ i , J m = 0 \mathbb{E}\xi_{i,J_m}=0 E ξ i , J m = 0 , and
r ^ m , J m − Q ^ m , J m β m , J m = M 1 , n − 1 ∑ i = 1 m n ξ i , J m . (4) \widehat{\boldsymbol{r}}_{m,J_m}-\widehat{Q}_{m,J_m}\boldsymbol{\beta}_{m,J_m}=M_{1,n}^{-1}\sum_{i=1}^{m_n}\xi_{i,J_m}. \tag*{(4)} r m , J m − Q m , J m β m , J m = M 1 , n − 1 i = 1 ∑ m n ξ i , J m . (4)
The estimator (2) pools data from all sites and smooths the mean over time t t t . The number J m J_m J m (of interior knots) controls bias-variance trade-off: a larger value of J m J_m J m value of reduces the bias but increases the variance. More observations within a curve reduce observation-level variation. Once the curves are sampled sufficiently densely in time, however, the dominant uncertainty arises from the latent curve-level variation, for which spatial dependence reduces the effective amount of replication. § 3 makes this distinction more precise.
Locally stationary mean
In this section, we consider the problem of estimating the spatially varying mean surface M ( u , t ) M(u,t) M ( u , t ) without assuming spatial stationarity of the mean. When the mean changes gradually over space, pooling all sites would obscure spatial heterogeneity. We instead represent the mean on a rescaled domain by a smooth function M : [ 0 , 1 ] d × [ 0 , 1 ] → R M:[0,1]^d \times[0,1] \to\mathbb{R} M : [ 0 , 1 ] d × [ 0 , 1 ] → R , that is
sup s ∈ S n , t ∈ [ 0 , 1 ] ∣ m ( s , t ) − M ( U ( s ) , t ) ∣ ≤ C L n − 1 , (5) \sup_{s \in\mathcal{S}_n,t \in[0,1]} \left|m(s,t)-M\left(U(s),t\right)\right| \le CL_n^{-1}, \tag*{(5)} s ∈ S n , t ∈ [ 0 , 1 ] sup ∣ m ( s , t ) − M ( U ( s ) , t ) ∣ ≤ C L n − 1 , (5)
where U ( s ) = ( s 1 / L 1 , n , … , s d / L d , n ) ⊤ U(s)=(s_1/L_{1,n},\ldots,s_d/L_{d,n})^\top U ( s ) = ( s 1 / L 1 , n , … , s d / L d , n ) ⊤ . This approximation is in the spirit of local-stationarity conditions for spatial regression and random fields [26 , 28 ] . It allows the mean function to depend on the spatial domain and hence implicitly on n n n . In particular, it describes a sequence of spatially varying mean surfaces that becomes asymptotically representable by a fixed function M M M on the rescaled domain. For a detailed formulation of the assumption (which also describes the condition on the temporal smoothness precisely) we refer to Assumption 3.9 in the following section.
The stationary and local settings share the difficulties caused by irregular observations over time and dependence across sites. Estimation of a spatially varying mean additionally requires smoothing in the spatial direction. We consider two approaches for this purpose. The (spatial) local-kernel estimator introduced in § 2.2.1 uses observations from sites in a neighborhood of a specified spatial location, with a bandwidth a n a_n a n controlling the degree of localization. The spatial-spline estimator discussed in § 2.2.2 instead represents the entire mean surface by a tensor-product spline, with the spatial resolution J s J_s J s controlling its flexibility. Accordingly, the convergence rates derived in the following § 3 reflect the different spatial resolutions of the two procedures: localization through the bandwidth a n a_n a n for the kernel estimator and through J s J_s J s for the spatial-spline estimator.
Local-kernel estimator
We write U i = U ( S i ) = ( S i 1 / L 1 , n , … , S i d / L d , n ) ⊤ ∈ [ 0 , 1 ] d U_i=U(S_i)=(S_{i1}/L_{1,n},\ldots,S_{id}/L_{d,n})^\top\in[0,1]^d U i = U ( S i ) = ( S i 1 / L 1 , n , … , S i d / L d , n ) ⊤ ∈ [ 0 , 1 ] d and fix an interior point u 0 ∈ ( 0 , 1 ) d u_0 \in(0,1)^d u 0 ∈ ( 0 , 1 ) d . Let K : R d → [ 0 , ∞ ) K:\mathbb{R}^d \to[0,\infty) K : R d → [ 0 , ∞ ) be a bounded Lipschitz kernel supported on [ − 1 , 1 ] d [-1,1]^d [ − 1 , 1 ] d , let a n ↓ 0 a_n \downarrow0 a n ↓ 0 denote a bandwidth, and define the weights W i ( u 0 ) : = K { ( U i − u 0 ) / a n } W_i(u_0):=K\{(U_i-u_0)/a_n\} W i ( u 0 ) := K {( U i − u 0 ) / a n } . Weighting sites within an a n a_n a n -neighborhood of u 0 u_0 u 0 gives the kernel estimator
M ^ J m ( u 0 , t ) = B m ( t ) ⊤ β ^ m , J m ( u 0 ) , (6) \widehat{M}_{J_m}(u_0,t)=\mathbf{B}_m(t)^\top\widehat{\boldsymbol{\beta}}_{m,J_m}(u_0), \tag*{(6)} M J m ( u 0 , t ) = B m ( t ) ⊤ β m , J m ( u 0 ) , (6)
where
β ^ m , J m ( u 0 ) = arg min β ∈ R K m ∑ i = 1 m n W i ( u 0 ) ∑ j = 1 N i { Y i j − B m ( T i j ) T β } 2 . (7) \widehat{\beta}_{m,J_m}(u_0)=\underset{\beta\in\mathbb{R}^{K_m}}{\arg\min}\sum_{i=1}^{m_n}W_i(u_0)\sum_{j=1}^{N_i}\left\{Y_{ij}-B_m(T_{ij})^{\mathsf{T}}\beta\right\}^2. \tag*{(7)} β m , J m ( u 0 ) = β ∈ R K m arg min i = 1 ∑ m n W i ( u 0 ) j = 1 ∑ N i { Y ij − B m ( T ij ) T β } 2 . (7)
Let M 1 , n ( u 0 ) : = ∑ i = 1 m n W i ( u 0 ) N i M_{1,n}(u_0):=\sum_{i=1}^{m_n}W_i(u_0)N_i M 1 , n ( u 0 ) := ∑ i = 1 m n W i ( u 0 ) N i . Then β ^ m , J m ( u 0 ) = Q ^ m , J m ( u 0 ) − 1 r ^ m , J m ( u 0 ) \widehat{\beta}_{m,J_m}(u_0)=\widehat{Q}_{m,J_m}(u_0)^{-1}\widehat{r}_{m,J_m}(u_0) β m , J m ( u 0 ) = Q m , J m ( u 0 ) − 1 r m , J m ( u 0 ) , where
Q ^ m , J m ( u 0 ) : = 1 M 1 , n ( u 0 ) ∑ i = 1 m n W i ( u 0 ) ∑ j = 1 N i B m ( T i j ) B m ( T i j ) T , \widehat{Q}_{m,J_m}(u_0):=\frac{1}{M_{1,n}(u_0)}\sum_{i=1}^{m_n}W_i(u_0)\sum_{j=1}^{N_i}B_m(T_{ij})B_m(T_{ij})^{\mathsf{T}}, Q m , J m ( u 0 ) := M 1 , n ( u 0 ) 1 i = 1 ∑ m n W i ( u 0 ) j = 1 ∑ N i B m ( T ij ) B m ( T ij ) T ,
r ^ m , J m ( u 0 ) : = 1 M 1 , n ( u 0 ) ∑ i = 1 m n W i ( u 0 ) ∑ j = 1 N i B m ( T i j ) Y i j . \widehat{r}_{m,J_m}(u_0):=\frac{1}{M_{1,n}(u_0)}\sum_{i=1}^{m_n}W_i(u_0)\sum_{j=1}^{N_i}B_m(T_{ij})Y_{ij}. r m , J m ( u 0 ) := M 1 , n ( u 0 ) 1 i = 1 ∑ m n W i ( u 0 ) j = 1 ∑ N i B m ( T ij ) Y ij .
Thus, the local estimator has the same temporal spline structure as the stationary estimator (2) , but only observations from sites receiving positive kernel weight contribute to the fit. The bandwidth a n a_n a n therefore determines both the spatial smoothing bias and the amount of spatial information available for estimating M ( u 0 , ⋅ ) M(u_0,\cdot) M ( u 0 , ⋅ ) .
Tensor-product spatial-spline estimator
When the full spatial surface is of interest, a global tensor-product spline borrows information smoothly across all sites and avoids fitting a separate local model at each location. Let A s ( u ) A_s(u) A s ( u ) be a normalized tensor-product B-spline basis constructed from quasi-uniform partitions in the d d d spatial coordinates, with J s = J s n J_s=J_{sn} J s = J s n subintervals in each coordinate and fixed order p s p_s p s . Its dimension is K s ≍ J s d K_s\asymp J_s^d K s ≍ J s d . With Z J s , J m ( u , t ) : = A s ( u ) ⊗ B m ( t ) ∈ R K s K m Z_{J_s,J_m}(u,t):=A_s(u)\otimes B_m(t)\in\mathbb{R}^{K_sK_m} Z J s , J m ( u , t ) := A s ( u ) ⊗ B m ( t ) ∈ R K s K m , define
M ^ J s , J m s p ( u , t ) = Z J s , J m ( u , t ) T θ ^ M , J s , J m , (8) \widehat{M}^{\mathrm{sp}}_{J_s,J_m}(u,t)=Z_{J_s,J_m}(u,t)^{\mathsf{T}}\widehat{\theta}_{M,J_s,J_m}, \tag*{(8)} M J s , J m sp ( u , t ) = Z J s , J m ( u , t ) T θ M , J s , J m , (8)
where
θ ^ M , J s , J m : = arg min θ ∈ R K s K m ∑ i = 1 m n ∑ j = 1 N i { Y i j − Z J s , J m ( U i , T i j ) T θ } 2 . \widehat{\theta}_{M,J_s,J_m}:=\underset{\theta\in\mathbb{R}^{K_sK_m}}{\arg\min}\sum_{i=1}^{m_n}\sum_{j=1}^{N_i}\left\{Y_{ij}-Z_{J_s,J_m}(U_i,T_{ij})^{\mathsf{T}}\theta\right\}^2. θ M , J s , J m := θ ∈ R K s K m arg min i = 1 ∑ m n j = 1 ∑ N i { Y ij − Z J s , J m ( U i , T ij ) T θ } 2 .
The empirical Gram matrix and empirical moment vector are given by
Q ^ M , J s , J m : = 1 M 1 , n ∑ i = 1 m n ∑ j = 1 N i Z J s , J m ( U i , T i j ) Z J s , J m ( U i , T i j ) T , \widehat{Q}_{M,J_s,J_m}:=\frac{1}{M_{1,n}}\sum_{i=1}^{m_n}\sum_{j=1}^{N_i}Z_{J_s,J_m}(U_i,T_{ij})Z_{J_s,J_m}(U_i,T_{ij})^{\mathsf{T}}, Q M , J s , J m := M 1 , n 1 i = 1 ∑ m n j = 1 ∑ N i Z J s , J m ( U i , T ij ) Z J s , J m ( U i , T ij ) T ,
r ^ M , J s , J m : = 1 M 1 , n ∑ i = 1 m n ∑ j = 1 N i Z J s , J m ( U i , T i j ) Y i j , \widehat{r}_{M,J_s,J_m}:=\frac{1}{M_{1,n}}\sum_{i=1}^{m_n}\sum_{j=1}^{N_i}Z_{J_s,J_m}(U_i,T_{ij})Y_{ij}, r M , J s , J m := M 1 , n 1 i = 1 ∑ m n j = 1 ∑ N i Z J s , J m ( U i , T ij ) Y ij ,
respectively, and the corresponding population quantities are given by
Q M , J s , J m : = 1 M 1 , n ∑ i = 1 m n N i ∫ 0 1 Z J s , J m ( U i , x ) Z J s , J m ( U i , x ) T f T ( x ) d x , Q_{M,J_s,J_m}:=\frac{1}{M_{1,n}}\sum_{i=1}^{m_n}N_i\int_0^1 Z_{J_s,J_m}(U_i,x)Z_{J_s,J_m}(U_i,x)^{\mathsf{T}}f_T(x)\,\mathrm{d}x, Q M , J s , J m := M 1 , n 1 i = 1 ∑ m n N i ∫ 0 1 Z J s , J m ( U i , x ) Z J s , J m ( U i , x ) T f T ( x ) d x ,
r M , J s , J m : = 1 M 1 , n ∑ i = 1 m n N i ∫ 0 1 Z J s , J m ( U i , x ) m ( S i , x ) f T ( x ) d x , r_{M,J_s,J_m}:=\frac{1}{M_{1,n}}\sum_{i=1}^{m_n}N_i\int_0^1 Z_{J_s,J_m}(U_i,x)m(S_i,x)f_T(x)\,\mathrm{d}x, r M , J s , J m := M 1 , n 1 i = 1 ∑ m n N i ∫ 0 1 Z J s , J m ( U i , x ) m ( S i , x ) f T ( x ) d x ,
with
θ M , J s , J m ∘ : = Q M , J s , J m − 1 r M , J s , J m , M J s , J m ∘ ( u , t ) : = Z J s , J m ( u , t ) T θ M , J s , J m ∘ . \theta^\circ_{M,J_s,J_m}:=Q_{M,J_s,J_m}^{-1}r_{M,J_s,J_m},\qquad M^\circ_{J_s,J_m}(u,t):=Z_{J_s,J_m}(u,t)^{\mathsf{T}}\theta^\circ_{M,J_s,J_m}. θ M , J s , J m ∘ := Q M , J s , J m − 1 r M , J s , J m , M J s , J m ∘ ( u , t ) := Z J s , J m ( u , t ) T θ M , J s , J m ∘ .
Note that the population projection M J s , J m ∘ M^\circ_{J_s,J_m} M J s , J m ∘ is defined in terms of the actual mean values m ( S i , ⋅ ) m(S_i,\cdot) m ( S i , ⋅ ) at the sampled sites, rather than the approximating surface M ( U i , ⋅ ) M(U_i,\cdot) M ( U i , ⋅ ) . This choice yields an exact centering of the sample normal equations. The difference between m ( S i , t ) m(S_i,t) m ( S i , t ) and M ( U i , t ) M(U_i,t) M ( U i , t ) is treated separately as the deterministic local-stationarity approximation error.
The defining population normal equation is given by
∑ i = 1 m n N i ∫ 0 1 Z J s , J m ( U i , x ) { m ( S i , x ) − M J s , J m ∘ ( U i , x ) } f T ( x ) d x = 0. \sum_{i=1}^{m_n} N_i \int_0^1 Z_{J_s,J_m}(U_i,x)\left\{m(S_i,x)-M^\circ_{J_s,J_m}(U_i,x)\right\}f_T(x)\,dx=0. i = 1 ∑ m n N i ∫ 0 1 Z J s , J m ( U i , x ) { m ( S i , x ) − M J s , J m ∘ ( U i , x ) } f T ( x ) d x = 0.
and the centered site score for the joint fit is defined as
ω i , J s , J m : = ∑ j = 1 N i Z J s , J m ( U i , T i j ) { Y i j − M J s , J m ∘ ( U i , T i j ) } − E i , \omega_{i,J_s,J_m}:=\sum_{j=1}^{N_i} Z_{J_s,J_m}(U_i,T_{ij})\left\{Y_{ij}-M^\circ_{J_s,J_m}(U_i,T_{ij})\right\}-\mathbb{E}_i, ω i , J s , J m := j = 1 ∑ N i Z J s , J m ( U i , T ij ) { Y ij − M J s , J m ∘ ( U i , T ij ) } − E i ,
where E i = E [ ∑ j = 1 N i Z J s , J m ( U i , T i j ) { Y i j − M J s , J m ∘ ( U i , T i j ) } ] \mathbb{E}_i=\mathbb{E}\left[\sum_{j=1}^{N_i}Z_{J_s,J_m}(U_i,T_{ij})\left\{Y_{ij}-M^\circ_{J_s,J_m}(U_i,T_{ij})\right\}\right] E i = E [ ∑ j = 1 N i Z J s , J m ( U i , T ij ) { Y ij − M J s , J m ∘ ( U i , T ij ) } ] . The population normal equation and Assumption 2.1 give ∑ i E i = 0 \sum_i\mathbb{E}_i=0 ∑ i E i = 0 , matching the aggregate centered score to the sample normal equations.
Theoretical guarantees
In this section we demonstrate that the methodology developed in § 2 provides consistent estimators in the different scenarios. We are particularly interested in uniform convergence rates and the effects of sparse and dense sampling in the spatial and temporal domain. For this purpose, we first introduce several conditions on the mean functions and the sampling framework that are required to derive uniform convergence rates for all three estimators introduced in § 2 . Assumptions specific to the estimation of a stationary mean or to a locally stationary setting are stated in the corresponding subsections below.
The rates presented in this section have a common structure. In all three estimation problems, the stochastic error separates into an observation-level component, governed by the nominal number of measurements, and a curve-level component, governed by the dependence-adjusted effective sample size m e f f , n m_{\mathrm{eff},n} m eff , n defined in (10) below. The assumptions introduced next provide the dependence, tail, and smoothness conditions needed to establish this decomposition.
For an integer q ≥ 0 q\ge0 q ≥ 0 and ν ∈ ( 0 , 1 ] \nu\in(0,1] ν ∈ ( 0 , 1 ] , let C q , ν [ 0 , 1 ] C^{q,\nu}[0,1] C q , ν [ 0 , 1 ] denote the Hölder class with smoothness index q + ν q+\nu q + ν and norm
∥ f ∥ C q , ν = ∑ ℓ = 0 q ∥ f ( ℓ ) ∥ ∞ + sup x ≠ y ∣ f ( q ) ( x ) − f ( q ) ( y ) ∣ ∣ x − y ∣ ν . \lVert f\rVert_{C^{q,\nu}}=\sum_{\ell=0}^{q}\lVert f^{(\ell)}\rVert_{\infty}+\sup_{x\ne y}\frac{\left|f^{(q)}(x)-f^{(q)}(y)\right|}{\left|x-y\right|^\nu}. ∥ f ∥ C q , ν = ℓ = 0 ∑ q ∥ f ( ℓ ) ∥ ∞ + x = y sup ∣ x − y ∣ ν f ( q ) ( x ) − f ( q ) ( y ) .
For our theoretical analysis, we combine a deterministic infill-increasing design with polynomial beta-mixing. The design is a fixed-site analogue of the increasing-domain and infill regimes in spatial statistics considered, for example, in Lahiri [27 ] or Jenish and Prucha [24 ] ; the mixing condition supports the covariance bounds [35 , 5 ] used in the proofs of our main results.
For a Borel set A ⊂ R d A \subset\mathbb{R}^{d} A ⊂ R d , let F X ( A ) \mathcal{F}_{X}(A) F X ( A ) denote the σ \sigma σ -field generated by { X ( s , ⋅ ) : s ∈ A } \{X(s,\cdot):s \in A\} { X ( s , ⋅ ) : s ∈ A } . The local model may depend on n n n , with this index suppressed and all constants uniform in n n n . For r ≥ 0 r \ge0 r ≥ 0 , we define the pairwise spatial beta-mixing coefficients by
β X ( r ) : = sup s ≠ s ′ { β { F X ( { s } ) , F X ( { s ′ } ) } : ∥ s − s ′ ∥ 2 ≥ r } . \beta_{X}(r) := \sup_{s \ne s'} \left\{\beta\{\mathcal{F}_{X}(\{s\}),\mathcal{F}_{X}(\{s'\})\}: \lVert s-s'\rVert_{2} \ge r\right\}. β X ( r ) := s = s ′ sup { β { F X ({ s }) , F X ({ s ′ })} : ∥ s − s ′ ∥ 2 ≥ r } .
where, for sub-σ \sigma σ -fields A , B \mathcal{A},\mathcal{B} A , B , the coefficient β ( A , B ) \beta(\mathcal{A},\mathcal{B}) β ( A , B ) is defined by
β ( A , B ) : = 1 2 sup { A u } , { B v } ∑ u , v ∣ Pr ( A u ∩ B v ) − Pr ( A u ) Pr ( B v ) ∣ . \beta(\mathcal{A},\mathcal{B}) := \frac{1}{2}\sup_{\{A_{u}\},\{B_{v}\}}\sum_{u,v}\left\lvert\Pr(A_{u}\cap B_{v})-\Pr(A_{u})\Pr(B_{v})\right\rvert. β ( A , B ) := 2 1 { A u } , { B v } sup u , v ∑ ∣ Pr ( A u ∩ B v ) − Pr ( A u ) Pr ( B v ) ∣ .
and the supremum is taken over all finite partitions { A u } \{A_{u}\} { A u } and { B v } \{B_{v}\} { B v } of the underlying sample space with A u ∈ A A_{u} \in\mathcal{A} A u ∈ A and B v ∈ B B_{v} \in\mathcal{B} B v ∈ B . Our next assumption refers to the decay rate of the spatial beta-mixing coefficients.
Assumption 3.1 (Polynomial beta-mixing). There exist constants C β > 0 C_{\beta} > 0 C β > 0 and α β > 2 d \alpha_{\beta} > 2d α β > 2 d such that β X ( r ) ≤ C β ( 1 + r ) − α β \beta_{X}(r) \le C_{\beta}(1+r)^{-\alpha_{\beta}} β X ( r ) ≤ C β ( 1 + r ) − α β for r ≥ 0 r \ge0 r ≥ 0 .
To measure the effect of spatial dependence we introduce the coefficient
Θ n : = 1 + sup 1 ≤ i ≤ m n ∑ k ≠ i β X ( ∥ S i − S k ∥ 2 ) 1 / 2 , (9) \Theta_{n} := 1+\sup_{1 \le i \le m_{n}}\sum_{k \ne i}\beta_{X}(\lVert S_{i}-S_{k}\rVert_{2})^{1/2}, \tag*{(9)} Θ n := 1 + 1 ≤ i ≤ m n sup k = i ∑ β X (∥ S i − S k ∥ 2 ) 1/2 , (9)
and call
m e f f , n : = m n / Θ n (10) m_{\mathrm{eff},n} := m_{n}/\Theta_{n} \tag*{(10)} m eff , n := m n / Θ n (10)
the (dependence-adjusted) effective sample size . Spatial dependence reduces the amount of independent information contained in the m n m_{n} m n sampled curves. The factor Θ n \Theta_{n} Θ n quantifies this redundancy, so that m e f f , n = m n / Θ n m_{\mathrm{eff},n}=m_{n}/\Theta_{n} m eff , n = m n / Θ n provides a dependence-adjusted measure of the effective number of curves for the purpose of our convergence-rate analysis; see Lemma S.2.4 in the supplementary material for more details. Thus, adding nearby sites may contribute much less information than adding the same number of well-separated sites. By contrast, independently sampled observation times and measurement errors still average over the actual observations, so this part of the estimation error benefits from the nominal sampling scale. This distinction leads to the two information scales appearing in the convergence rates presented below. Under Assumptions 2.2 and 3.1, Lemma S.2.2 in the supplementary material gives m e f f , n ≳ V n m_{\mathrm{eff},n} \gtrsim V_{n} m eff , n ≳ V n . In the case of a sparse space regime, we have m n ≍ m e f f , n ≍ V n m_{n} \asymp m_{\mathrm{eff},n} \asymp V_{n} m n ≍ m eff , n ≍ V n , whereas in the dense-space infill regime, we have m e f f , n ≍ V n m_{\mathrm{eff},n} \asymp V_{n} m eff , n ≍ V n under Assumption 2.3 and a short-range nonindependence condition (see Proposition S.2.3 in the supplementary material for details).
The next assumptions describe irregular time sampling, tails, and smoothness, paralleling the independent-curve framework of Cai and Yuan [9 ] and Zhang and Wang [43 ] .
Assumption 3.2 (Uniform tails). There exists a constant K 0 > 0 K_{0} > 0 K 0 > 0 such that
sup s ∈ R d , t ∈ [ 0 , 1 ] ∥ X ( s , t ) ∥ ψ 2 ≤ K 0 , sup i , j ∥ ε i j ∥ ψ 2 ≤ K 0 , sup s , t ∣ σ ( s , t ) ∣ ≤ K 0 . \sup_{s\in\mathbb{R}^{d},t\in[0,1]} \lVert X(s,t)\rVert_{\psi_{2}} \le K_{0}, \qquad\sup_{i,j}\lVert\varepsilon_{ij}\rVert_{\psi_{2}} \le K_{0}, \qquad\sup_{s,t}|\sigma(s,t)| \le K_{0}. s ∈ R d , t ∈ [ 0 , 1 ] sup ∥ X ( s , t ) ∥ ψ 2 ≤ K 0 , i , j sup ∥ ε ij ∥ ψ 2 ≤ K 0 , s , t sup ∣ σ ( s , t ) ∣ ≤ K 0 . Moreover, we assume that the finite-dimensional distributions of the centered field are uniformly sub-Gaussian relative to their variance: for every r ≥ 1 r \ge1 r ≥ 1 , every deterministic collection ( s ℓ , t ℓ ) ℓ = 1 r (s_{\ell},t_{\ell})_{\ell=1}^{r} ( s ℓ , t ℓ ) ℓ = 1 r , and every vector ( a 1 , … , a r ) T ∈ R r (a_{1},\ldots,a_{r})^{\mathsf{T}} \in\mathbb{R}^{r} ( a 1 , … , a r ) T ∈ R r ,
∥ ∑ ℓ = 1 r a ℓ X ( s ℓ , t ℓ ) ∥ ψ 2 ≤ K 0 [ Var { ∑ ℓ = 1 r a ℓ X ( s ℓ , t ℓ ) } ] 1 / 2 . (11) \left\lVert\sum_{\ell=1}^{r}a_{\ell}X(s_{\ell},t_{\ell}) \right\rVert_{\psi_{2}} \le K_{0}\left[\operatorname{Var}\left\{\sum_{\ell=1}^{r}a_{\ell}X(s_{\ell},t_{\ell})\right\}\right]^{1/2}. \tag*{(11)} ℓ = 1 ∑ r a ℓ X ( s ℓ , t ℓ ) ψ 2 ≤ K 0 [ Var { ℓ = 1 ∑ r a ℓ X ( s ℓ , t ℓ ) } ] 1/2 . (11)
Assumption 3.3 (Balanced within-site counts). There exist constants 0 < c N < C N < ∞ 0 < c_{N} < C_{N} < \infty 0 < c N < C N < ∞ such that c N N ‾ n ≤ N i ≤ C N N ‾ n c_{N}\overline{N}_{n} \le N_{i} \le C_{N}\overline{N}_{n} c N N n ≤ N i ≤ C N N n for every 1 ≤ i ≤ m n 1 \le i \le m_{n} 1 ≤ i ≤ m n , where N ‾ n = M 1 , n / m n \overline{N}_{n} = M_{1,n}/m_{n} N n = M 1 , n / m n .
Assumption 3.4 (Temporal mean smoothness). In the stationary setting, the mean function m m m is an element of C q m , ν m [ 0 , 1 ] C^{q_{m},\nu_{m}}[0,1] C q m , ν m [ 0 , 1 ] ; in the locally stationary setting, the function t ↦ M ( u , t ) t \mapsto M(u,t) t ↦ M ( u , t ) is an element of C q m , ν m [ 0 , 1 ] C^{q_{m},\nu_{m}}[0,1] C q m , ν m [ 0 , 1 ] for all u ∈ [ 0 , 1 ] d u \in[0,1]^{d} u ∈ [ 0 , 1 ] d . The spline order satisfies p ≥ q m ∗ + 1 p \ge q_{m}^{*} + 1 p ≥ q m ∗ + 1 , where q m ∗ = q m + ν m > 0 q_{m}^{*} = q_{m} + \nu_{m} > 0 q m ∗ = q m + ν m > 0 .
Assumptions 2.1 and 3.2 to 3.4 collect familiar regularity conditions for irregularly sampled functional data. Random exogenous observation times with a density bounded above and away from zero, comparable within-curve sample sizes, and smooth mean functions have close counterparts in sparse-to-dense functional mean estimation [41 , 9 , 43 ] . The covariance-relative finite-dimensional condition in (11) is stronger than the preceding pointwise ψ 2 \psi_{2} ψ 2 bound because it controls every deterministic linear combination. It holds automatically for Gaussian fields and parallels covariance-scaled sub-Gaussian conditions for Hilbert-space-valued functional observations [17 ] . The balance condition is used to express the phase-transition rates through N ‾ n \overline{N}_{n} N n and to simplify the variance terms for the local-kernel and spatial-spline estimators.
Stationary mean
We first establish uniform convergence for the pooled stationary-mean estimator (2) in the situation, when the mean function does not depend on the spatial locations. For this purpose we require a further assumption.
Assumption 3.5 (Mean leverage condition). The site counts satisfy max 1 ≤ i ≤ m n N i = O ( M 1 , n / m n ) \max_{1\le i\le m_{n}} N_{i} = O(M_{1,n}/m_{n}) max 1 ≤ i ≤ m n N i = O ( M 1 , n / m n ) .
Assumption 3.5 is equivalent to max i = 1 m n N i / N ‾ n = O ( 1 ) \max_{i=1}^{m_{n}}N_{i}/\overline{N}_{n} = O(1) max i = 1 m n N i / N n = O ( 1 ) and prevents individual sites from contributing a disproportionately large fraction of the total observations. Unlike Assumption 3.3, it does not impose a uniform lower bound on the within-site sample sizes. The condition is not essential for obtaining a convergence rate, but it simplifies the representation of the rates. For a family of p n p_{n} p n score coordinates, write b n = ( e + p n + m e f f , n ) 2 / α β b_{n} = (e+p_{n}+m_{\mathrm{eff},n})^{2/\alpha_{\beta}} b n = ( e + p n + m eff , n ) 2/ α β .
Theorem 3.6. Suppose that m ( s , t ) ≡ m ( t ) m(s,t) \equiv m(t) m ( s , t ) ≡ m ( t ) and that Assumptions 2.1 , 2.2 , 3.1 , 3.2 , 3.4 and 3.5 are satisfied. Take p n = K m p_n = K_m p n = K m in the definition of b n b_n b n . Let J m → ∞ J_m \to\infty J m → ∞ and suppose J m { [ log K m / ( M 1 , n J m ) ] 1 / 2 + log K m / M 1 , n } → 0 J_m\{\left[\log K_m/(M_{1,n}J_m)\right]^{1/2} + \log K_m/M_{1,n}\} \to0 J m { [ log K m / ( M 1 , n J m ) ] 1/2 + log K m / M 1 , n } → 0 . Then estimator (2) satisfies that
sup t ∈ [ 0 , 1 ] ∣ m ^ J m ( t ) − m ( t ) ∣ = O p ( J m − q m ∗ + log K m [ J m M 1 , n + M 2 , n + ( Θ n − 1 ) ( M 1 , n + M 2 , n ) M 1 , n 2 ] + J m δ n − d b n d log K m m n ) . (12) \sup_{t \in[0,1]} \left|\widehat{m}_{J_m}(t) - m(t)\right| = O_p\left(J_m^{-q_m^*} + \sqrt{\log K_m\left[\frac{J_m}{M_{1,n}} + \frac{M_{2,n} + (\Theta_n - 1)(M_{1,n} + M_{2,n})}{M_{1,n}^2}\right]} + \frac{J_m\delta_n^{-d}b_n^{d}\log K_m}{m_n}\right). \tag*{(12)} t ∈ [ 0 , 1 ] sup ∣ m J m ( t ) − m ( t ) ∣ = O p J m − q m ∗ + log K m [ M 1 , n J m + M 1 , n 2 M 2 , n + ( Θ n − 1 ) ( M 1 , n + M 2 , n ) ] + m n J m δ n − d b n d log K m . (12)
The first term on the right-hand side of (12) is the spline-approximation error, while the second term corresponds to the stochastic part. Under the balanced-count condition this term can be further simplified.
Corollary 3.7. Suppose that the conditions of Theorem 3.6 and Assumption 3.3 are satisfied. If, additionally, δ n − d b n d J m { log K m / [ m n ( Θ n + J m / N ‾ n ) ] } 1 / 2 → 0 \delta_n^{-d}b_n^dJ_m\{\log K_m/[m_n(\Theta_n + J_m/\overline{N}_n)]\}^{1/2} \to0 δ n − d b n d J m { log K m / [ m n ( Θ n + J m / N n )] } 1/2 → 0 , then (12) turns into
sup t ∈ [ 0 , 1 ] ∣ m ^ J m ( t ) − m ( t ) ∣ = O p ( J m − q m ∗ + log K m ( J m m n N ‾ n + 1 m e f f , n ) ) . (13) \sup_{t \in[0,1]} \left|\widehat{m}_{J_m}(t) - m(t)\right| = O_p\left(J_m^{-q_m^*} + \sqrt{\log K_m\left(\frac{J_m}{m_n\overline{N}_n} + \frac{1}{m_{\mathrm{eff},n}}\right)}\right). \tag*{(13)} t ∈ [ 0 , 1 ] sup ∣ m J m ( t ) − m ( t ) ∣ = O p ( J m − q m ∗ + log K m ( m n N n J m + m eff , n 1 ) ) . (13)
The stochastic part in (13) consists of two components with different information scales. The term J m / ( m n N ‾ n ) J_m/(m_n\overline{N}_n) J m / ( m n N n ) corresponds to observation-level variation, whereas m e f f , n − 1 m_{\mathrm{eff},n}^{-1} m eff , n − 1 represents the curve-level component. Independently sampled observation times and measurement errors contribute to the former, while variation of the latent random field and spatial dependence contribute to the latter. Increasing the number of observations per curve reduces the observation-level component until the curve-level component becomes dominant. Ignoring logarithmic factors, the balance between these two components yields the temporal phase-transition boundary
N ‾ n ≍ m e f f , n 1 / ( 2 q m ∗ ) Θ n . (14) \overline{N}_n \asymp\frac{m_{\mathrm{eff},n}^{1/(2q_m^*)}}{\Theta_n}. \tag*{(14)} N n ≍ Θ n m eff , n 1/ ( 2 q m ∗ ) . (14)
Below this boundary, balancing the temporal spline bias with the observation-level component gives J m ≍ ( m n N ‾ n ) 1 / ( 2 q m ∗ + 1 ) J_m \asymp(m_n\overline{N}_n)^{1/(2q_m^*+1)} J m ≍ ( m n N n ) 1/ ( 2 q m ∗ + 1 ) and the rate
sup t ∈ [ 0 , 1 ] ∣ m ^ J m ( t ) − m ( t ) ∣ = O p { ( m n N ‾ n ) − q m ∗ / ( 2 q m ∗ + 1 ) } . \sup_{t \in[0,1]} \left|\widehat{m}_{J_m}(t) - m(t)\right| = O_p\left\{(m_n\overline{N}_n)^{-q_m^*/(2q_m^*+1)}\right\}. t ∈ [ 0 , 1 ] sup ∣ m J m ( t ) − m ( t ) ∣ = O p { ( m n N n ) − q m ∗ / ( 2 q m ∗ + 1 ) } .
up to logarithmic factors. At or above the boundary, choosing J m ≍ m e f f , n 1 / ( 2 q m ∗ ) J_m \asymp m_{\mathrm{eff},n}^{1/(2q_m^*)} J m ≍ m eff , n 1/ ( 2 q m ∗ ) yields the rate m e f f , n − 1 / 2 m_{\mathrm{eff},n}^{-1/2} m eff , n − 1/2 , again up to logarithmic factors. We refer to temporal sampling below the boundary as sparse and to sampling at or above it as dense.
An important consequence of (14) is that the temporal phase-transition boundary depends not only on the effective number of spatial curves, but also on the spatial dependence inflation factor Θ n \Theta_n Θ n . In the sparse-space regime, Θ n ≍ 1 \Theta_n \asymp1 Θ n ≍ 1 and m e f f , n ≍ m n m_{\mathrm{eff},n} \asymp m_n m eff , n ≍ m n , so that (14) reduces to the familiar boundary N ‾ n ≍ m n 1 / ( 2 q m ∗ ) \overline{N}_n \asymp m_n^{1/(2q_m^*)} N n ≍ m n 1/ ( 2 q m ∗ ) . By contrast, under full-dimensional spatial infill with persistent local dependence, we have m e f f , n ≍ V n m_{\mathrm{eff},n} \asymp V_n m eff , n ≍ V n and Θ n ≍ δ n − d \Theta_n \asymp\delta_n^{-d} Θ n ≍ δ n − d , and the boundary becomes N ‾ n ≍ V n 1 / ( 2 q m ∗ ) δ n d \overline{N}_n \asymp V_n^{1/(2q_m^*)}\delta_n^d N n ≍ V n 1/ ( 2 q m ∗ ) δ n d . Thus, increasing spatial sampling density shifts the temporal phase transition towards fewer observations per curve. Depending on the relative rates of domain expansion and spatial infill, this boundary may even converge to zero. Since N ‾ n ≥ 1 \overline{N}_n \ge1 N n ≥ 1 , all feasible temporal sampling densities then lie asymptotically in the dense-time regime of the leading convergence-rate bound. In this situation, additional spatial observations may still reduce the observation-level variation, but the leading curve-level error remains of order V n − 1 / 2 V_n^{-1/2} V n − 1/2 . This illustrates how spatial infill can alter the temporal phase transition rather than merely reducing the effective number of independent curves.
Remark 3.8. If the data are independent, we have Θ n = 1 \Theta_n = 1 Θ n = 1 and Corollary 3.7 recovers the sparse-to-dense structure in the work of Cai and Yuan [9 ] and Zhang and Wang [43 ] . For q m ∗ = 2 q_m^* = 2 q m ∗ = 2 , the phase transition boundary is N ‾ n ≍ m n 1 / 4 \overline{N}_n \asymp m_n^{1/4} N n ≍ m n 1/4 after logarithmic factors are suppressed. The same algebra applies to a stationary functional time series observed on the fixed-spacing lattice. Indeed, Assumption 3.1 with d = 1 d = 1 d = 1 and α β > 2 \alpha_\beta> 2 α β > 2 implies ∑ h ≥ 1 β X ( Δ h ) 1 / 2 < ∞ \sum_{h \ge1} \beta_X(\Delta_h)^{1/2} < \infty ∑ h ≥ 1 β X ( Δ h ) 1/2 < ∞ , hence Θ n = O ( 1 ) \Theta_n = O(1) Θ n = O ( 1 ) and m e f f , n ≍ m n m_{\mathrm{eff},n} \asymp m_n m eff , n ≍ m n . The resulting rate is of order ( m n N ‾ n ) − q m ∗ / ( 2 q m ∗ + 1 ) (m_n\overline{N}_n)^{-q_m^*/(2q_m^*+1)} ( m n N n ) − q m ∗ / ( 2 q m ∗ + 1 ) below the phase transition boundary and m n − 1 / 2 m_n^{-1/2} m n − 1/2 above it. The latter agrees with the root-m n m_n m n mean behavior in Horváth et al. [22 ] , Chen and Song [11 ] , and Li and Yang [29 ] , with serial dependence entering their long-run covariance. This specialization complements sup-norm theory for independent curves [2 ] by making the spatial covariance inflation explicit. The rates established here are upper bounds for the proposed estimators; establishing matching minimax lower bounds and simultaneous inference is left for future work.
Locally stationary mean
We next derive uniform convergence rates for the (spatial) kernel estimator in (6) , with coefficients obtained from (7) , and for the spatial spline estimator (8) . Both methods allow the mean to vary smoothly over space. We first state a condition used to describe this setting.
Assumption 3.9 (Locally stationary mean). There exists a function M : [ 0 , 1 ] d × [ 0 , 1 ] → R M : [0, 1]^d \times[0, 1] \to\mathbb{R} M : [ 0 , 1 ] d × [ 0 , 1 ] → R such that: (i) for some γ u ∈ ( 0 , 1 ] \gamma_u \in(0, 1] γ u ∈ ( 0 , 1 ] , sup t ∈ [ 0 , 1 ] ∣ M ( u , t ) − M ( v , t ) ∣ ≤ C ∥ u − v ∥ 2 γ u \sup_{t \in[0,1]} |M(u,t) - M(v,t)| \le C\lVert u-v\rVert_2^{\gamma_u} sup t ∈ [ 0 , 1 ] ∣ M ( u , t ) − M ( v , t ) ∣ ≤ C ∥ u − v ∥ 2 γ u for all u , v ∈ [ 0 , 1 ] d u,v \in[0,1]^d u , v ∈ [ 0 , 1 ] d ; (ii) the temporal smoothness requirement in Assumption 3.4 holds uniformly in u u u ; (iii) the actual mean satisfies (5) ; and (iv) the family { m ( s , ⋅ ) : s ∈ S n } \{m(s,\cdot) : s \in S_n\} { m ( s , ⋅ ) : s ∈ S n } is uniformly bounded in C q m , ν m [ 0 , 1 ] C^{q_m,\nu_m}[0,1] C q m , ν m [ 0 , 1 ] .
Local-kernel estimator
For the theoretical analysis of the kernel estimator (6) , we make the following assumption regarding the weights in the loss function (7) .
Assumption 3.10. At the fixed point u 0 ∈ ( 0 , 1 ) d u_0 \in(0,1)^d u 0 ∈ ( 0 , 1 ) d , we have ∑ i = 1 m n W i ( u 0 ) ≍ m n a n d \sum_{i=1}^{m_n} W_i(u_0) \asymp m_n a_n^d ∑ i = 1 m n W i ( u 0 ) ≍ m n a n d and ∑ i = 1 m n W i ( u 0 ) 2 ≲ m n a n d \sum_{i=1}^{m_n} W_i(u_0)^2 \lesssim m_n a_n^d ∑ i = 1 m n W i ( u 0 ) 2 ≲ m n a n d .
Recall the notation M 1 , n ( u 0 ) = ∑ i = 1 m n W i ( u 0 ) N i M_{1,n}(u_0) = \sum_{i=1}^{m_n} W_i(u_0)N_i M 1 , n ( u 0 ) = ∑ i = 1 m n W i ( u 0 ) N i and define m u 0 , e f f , n : = { M 1 , n ( u 0 ) } 2 / { Θ n ∑ i = 1 m n W i ( u 0 ) 2 N i 2 } m_{u_0,\mathrm{eff},n} := \{M_{1,n}(u_0)\}^2 / \{\Theta_n \sum_{i=1}^{m_n} W_i(u_0)^2 N_i^2\} m u 0 , eff , n := { M 1 , n ( u 0 ) } 2 / { Θ n ∑ i = 1 m n W i ( u 0 ) 2 N i 2 } . Note that under Assumptions 3.3 and 3.10 we have m u 0 , e f f , n ≍ m_{u_0,\mathrm{eff},n} \asymp m u 0 , eff , n ≍
m e f f , n a n d m_{\mathrm{eff},n}a_n^d m eff , n a n d . Thus, spatial localization reduces both the nominal and the effective curve-level sample sizes by the window-volume factor a n d a_n^d a n d .
The deterministic error of the local kernel estimator (2.5) consists of the temporal spline bias, the spatial smoothing bias a n γ u a_n^{\gamma_u} a n γ u , and the L n − 1 L_n^{-1} L n − 1 approximation error from local stationarity.
Theorem 3.11. Let Assumptions 2.1, 2.2, 3.1-3.3, 3.9, 3.10 be satisfied and let J m → ∞ J_m \to\infty J m → ∞ , J m { [ log K m / ( M 1 , n ( u 0 ) J m ) ] 1 / 2 + log K m / M 1 , n ( u 0 ) } → 0 J_m\{[\log K_m/(M_{1,n}(u_0)J_m)]^{1/2}+\log K_m/M_{1,n}(u_0)\}\to0 J m {[ log K m / ( M 1 , n ( u 0 ) J m ) ] 1/2 + log K m / M 1 , n ( u 0 )} → 0 . Take p n = K m p_n=K_m p n = K m in the definition of b n b_n b n . Then, for every fixed interior point u 0 u_0 u 0 ,
sup t ∈ [ 0 , 1 ] ∣ M ^ J m ( u 0 , t ) − M ( u 0 , t ) ∣ = O p ( J m − q m ∗ + a n γ u + L n − 1 + log K m ( J m m n a n d N ‾ n + 1 m e f f , n a n d ) + J m δ n − d b n d log K m m n a n d ) . (15) \sup_{t\in[0,1]}\left|\widehat{M}_{J_m}(u_0,t)-M(u_0,t)\right|
=O_p\left(J_m^{-q_m^*}+a_n^{\gamma_u}+L_n^{-1}+
\sqrt{\log K_m\left(\frac{J_m}{m_na_n^d\overline{N}_n}+\frac{1}{m_{\mathrm{eff},n}a_n^d}\right)}
+\frac{J_m\delta_n^{-d}b_n^d\log K_m}{m_na_n^d}\right). \tag*{(15)} t ∈ [ 0 , 1 ] sup M J m ( u 0 , t ) − M ( u 0 , t ) = O p ( J m − q m ∗ + a n γ u + L n − 1 + log K m ( m n a n d N n J m + m eff , n a n d 1 ) + m n a n d J m δ n − d b n d log K m ) . (15)
Within the spatial window, the observation-level information is of order m n a n d N ‾ n m_na_n^d\overline{N}_n m n a n d N n , whereas the curve-level effective size is m e f f , n a n d ≍ m u 0 , e f f , n m_{\mathrm{eff},n}a_n^d\asymp m_{u_0,\mathrm{eff},n} m eff , n a n d ≍ m u 0 , eff , n . The balance between these two information scales determines the local time-sampling transition.
Corollary 3.12 (Local phase transition). If the assumptions of Theorem 3.11 and additionally the condition δ n − d b n d J m { log K m / [ a n d m n ( Θ n + J m / N ‾ n ) ] } 1 / 2 → 0 \delta_n^{-d}b_n^dJ_m\{\log K_m/[a_n^dm_n(\Theta_n+J_m/\overline{N}_n)]\}^{1/2}\to0 δ n − d b n d J m { log K m / [ a n d m n ( Θ n + J m / N n )] } 1/2 → 0 are satisfied, the rate in (3.7) simplifies to
sup t ∈ [ 0 , 1 ] ∣ M ^ J m ( u 0 , t ) − M ( u 0 , t ) ∣ = O p ( J m − q m ∗ + a n γ u + L n − 1 + log K m ( J m m n a n d N ‾ n + 1 m e f f , n a n d ) ) . (16) \sup_{t\in[0,1]}\left|\widehat{M}_{J_m}(u_0,t)-M(u_0,t)\right|
=O_p\left(J_m^{-q_m^*}+a_n^{\gamma_u}+L_n^{-1}+
\sqrt{\log K_m\left(\frac{J_m}{m_na_n^d\overline{N}_n}+\frac{1}{m_{\mathrm{eff},n}a_n^d}\right)}\right). \tag*{(16)} t ∈ [ 0 , 1 ] sup M J m ( u 0 , t ) − M ( u 0 , t ) = O p ( J m − q m ∗ + a n γ u + L n − 1 + log K m ( m n a n d N n J m + m eff , n a n d 1 ) ) . (16)
Ignoring logarithmic factors and L n − 1 L_n^{-1} L n − 1 , the time-direction phase transition boundary is
N ‾ n ≳ ( m e f f , n a n d ) 1 / ( 2 q m ∗ ) Θ n . (17) \overline{N}_n\gtrsim\frac{(m_{\mathrm{eff},n}a_n^d)^{1/(2q_m^*)}}{\Theta_n}. \tag*{(17)} N n ≳ Θ n ( m eff , n a n d ) 1/ ( 2 q m ∗ ) . (17)
Below this phase transition boundary, balancing the temporal spline bias with the observation-level component gives J m ≍ ( m n a n d N ‾ n ) 1 / ( 2 q m ∗ + 1 ) J_m\asymp(m_na_n^d\overline{N}_n)^{1/(2q_m^*+1)} J m ≍ ( m n a n d N n ) 1/ ( 2 q m ∗ + 1 ) and the time-smoothing contribution ( m n a n d N ‾ n ) − q m ∗ / ( 2 q m ∗ + 1 ) (m_na_n^d\overline{N}_n)^{-q_m^*/(2q_m^*+1)} ( m n a n d N n ) − q m ∗ / ( 2 q m ∗ + 1 ) in (3.8), up to logarithmic factors. At or above the phase transition boundary, the curve-level component determines the order of the stochastic error. Taking J m ≍ ( m e f f , n a n d ) 1 / ( 2 q m ∗ ) J_m\asymp(m_{\mathrm{eff},n}a_n^d)^{1/(2q_m^*)} J m ≍ ( m eff , n a n d ) 1/ ( 2 q m ∗ ) gives the order ( m e f f , n a n d ) − 1 / 2 (m_{\mathrm{eff},n}a_n^d)^{-1/2} ( m eff , n a n d ) − 1/2 . In both regimes, the complete error bound in (3.8) also contains the spatial terms a n γ u a_n^{\gamma_u} a n γ u and L n − 1 L_n^{-1} L n − 1 .
The preceding results concern estimation at a fixed spatial location. We next strengthen Theorem 3.11 to obtain uniformity over interior spatial points.
Theorem 3.13 (Uniformity over interior spatial points). Let U ∘ = [ c 0 , 1 − c 0 ] d \mathcal{U}_{\circ}=[c_0,1-c_0]^d U ∘ = [ c 0 , 1 − c 0 ] d for some fixed constant c 0 ∈ ( 0 , 1 / 2 ) c_0\in(0,1/2) c 0 ∈ ( 0 , 1/2 ) , and let Assumption 2.1, 2.2, 3.1-3.3 and 3.9 be satisfied. Further,
assume that the kernel K K K is bounded, Lipschitz continuous, and compactly supported and that, for r = 1 , 2 r = 1, 2 r = 1 , 2 ,
c m n a n d ≤ inf u ∈ U 0 ∑ i = 1 m n W i ( u ) r ≤ sup u ∈ U 0 ∑ i = 1 m n W i ( u ) r ≤ C m n a n d . c m_n a_n^d \le\inf_{u\in\mathcal{U}_0} \sum_{i=1}^{m_n} W_i(u)^r \le\sup_{u\in\mathcal{U}_0} \sum_{i=1}^{m_n} W_i(u)^r \le C m_n a_n^d. c m n a n d ≤ u ∈ U 0 inf i = 1 ∑ m n W i ( u ) r ≤ u ∈ U 0 sup i = 1 ∑ m n W i ( u ) r ≤ C m n a n d .
Let U n ⊂ U 0 \mathcal{U}_n \subset\mathcal{U}_0 U n ⊂ U 0 be a regular ( a n 2 d + 2 / J m ) (a_n^{2d+2}/J_m) ( a n 2 d + 2 / J m ) -net, take p n = K m ∣ U n ∣ p_n = K_m|\mathcal{U}_n| p n = K m ∣ U n ∣ in the definition of b n b_n b n , and define ℓ n : = log K m + log ( a n − 1 ) \ell_n := \log K_m + \log(a_n^{-1}) ℓ n := log K m + log ( a n − 1 ) . If J m → ∞ J_m \to\infty J m → ∞ , J m { [ ℓ n / ( m n a n d N ‾ n J m ) ] 1 / 2 + ℓ n / ( m n a n d N ‾ n ) } → 0 J_m\left\{\left[\ell_n/(m_n a_n^d\overline{N}_n J_m)\right]^{1/2}+\ell_n/(m_n a_n^d\overline{N}_n)\right\}\to0 J m { [ ℓ n / ( m n a n d N n J m ) ] 1/2 + ℓ n / ( m n a n d N n ) } → 0 , we have
sup u ∈ U 0 sup t ∈ [ 0 , 1 ] ∣ M ^ J m ( u , t ) − M ( u , t ) ∣ = O p ( J m − q m ∗ + a n q u + L n − 1 + ℓ n ( J m m n a n d N ‾ n + 1 m e f f , n a n d ) + J m δ n − d b n d ℓ n m n a n d ) . (18) \sup_{u\in\mathcal{U}_0}\sup_{t\in[0,1]}|\widehat{M}_{J_m}(u,t)-M(u,t)|
=O_p\left(J_m^{-q_m^*}+a_n^{q_u}+L_n^{-1}+\sqrt{\ell_n\left(\frac{J_m}{m_n a_n^d\overline{N}_n}+\frac{1}{m_{\mathrm{eff},n}a_n^d}\right)}+\frac{J_m^\delta n^{-d}b_n^d\ell_n}{m_n a_n^d}\right). \tag*{(18)} u ∈ U 0 sup t ∈ [ 0 , 1 ] sup ∣ M J m ( u , t ) − M ( u , t ) ∣ = O p ( J m − q m ∗ + a n q u + L n − 1 + ℓ n ( m n a n d N n J m + m eff , n a n d 1 ) + m n a n d J m δ n − d b n d ℓ n ) . (18)
Spatial-spline estimator
The tensor-product estimator (8) recovers the complete surface M M M in one global fit. In place of a local window, it allocates the effective sample across J s d J_s^d J s d spatial degrees of freedom. For its theoretical analysis we require the following assumptions.
Assumption 3.14 (Spatial smoothness of the local mean). Let q u ∈ N 0 q_u\in\mathbb{N}_0 q u ∈ N 0 , ν u ∈ ( 0 , 1 ] \nu_u\in(0,1] ν u ∈ ( 0 , 1 ] , and write q u ∗ = q u + ν u q_u^*=q_u+\nu_u q u ∗ = q u + ν u . For every multi-index α ∈ N 0 d \alpha\in\mathbb{N}_0^d α ∈ N 0 d with ∣ α ∣ ≤ q u |\alpha|\le q_u ∣ α ∣ ≤ q u and every 0 ≤ ℓ ≤ q m 0\le\ell\le q_m 0 ≤ ℓ ≤ q m , the derivative ∂ u α ∂ t ℓ M ( u , t ) \partial_u^\alpha\partial_t^\ell M(u,t) ∂ u α ∂ t ℓ M ( u , t ) exists and is uniformly bounded on [ 0 , 1 ] d × [ 0 , 1 ] [0,1]^d\times[0,1] [ 0 , 1 ] d × [ 0 , 1 ] . Moreover, there exists a constant C > 0 C>0 C > 0 such that
sup t ∈ [ 0 , 1 ] ∣ ∂ u α ∂ t ℓ M ( u , t ) − ∂ u α ∂ t ℓ M ( v , t ) ∣ ≤ C ∥ u − v ∥ 2 ν u \sup_{t\in[0,1]}\left|\partial_u^\alpha\partial_t^\ell M(u,t)-\partial_u^\alpha\partial_t^\ell M(v,t)\right|\le C\|u-v\|_2^{\nu_u} t ∈ [ 0 , 1 ] sup ∂ u α ∂ t ℓ M ( u , t ) − ∂ u α ∂ t ℓ M ( v , t ) ≤ C ∥ u − v ∥ 2 ν u
for all u , v ∈ [ 0 , 1 ] d u,v\in[0,1]^d u , v ∈ [ 0 , 1 ] d , ∣ α ∣ = q u |\alpha|=q_u ∣ α ∣ = q u , and 0 ≤ ℓ ≤ q m 0\le\ell\le q_m 0 ≤ ℓ ≤ q m , and
sup u ∈ [ 0 , 1 ] d ∣ ∂ u α ∂ t q m M ( u , t ) − ∂ u α ∂ t q m M ( u , s ) ∣ ≤ C ∣ t − s ∣ ν m \sup_{u\in[0,1]^d}\left|\partial_u^\alpha\partial_t^{q_m}M(u,t)-\partial_u^\alpha\partial_t^{q_m}M(u,s)\right|\le C|t-s|^{\nu_m} u ∈ [ 0 , 1 ] d sup ∣ ∂ u α ∂ t q m M ( u , t ) − ∂ u α ∂ t q m M ( u , s ) ∣ ≤ C ∣ t − s ∣ ν m
for all s , t ∈ [ 0 , 1 ] s,t\in[0,1] s , t ∈ [ 0 , 1 ] and ∣ α ∣ ≤ q u |\alpha|\le q_u ∣ α ∣ ≤ q u . The spline orders satisfy p s ≥ q u ∗ + 1 p_s\ge q_u^*+1 p s ≥ q u ∗ + 1 , p ≥ q m ∗ + 1 p\ge q_m^*+1 p ≥ q m ∗ + 1 .
Assumption 3.15 (Spatial design stability). The aspect ratios L r , n / L n L_{r,n}/L_n L r , n / L n are bounded above and bounded away from zero. For the rescaled sites U i = U ( S i ) ∈ [ 0 , 1 ] d U_i=U(S_i)\in[0,1]^d U i = U ( S i ) ∈ [ 0 , 1 ] d , define
h U , n : = sup u ∈ [ 0 , 1 ] d min 1 ≤ i ≤ m n ∥ u − U i ∥ 2 , q U , n : = 1 2 min i ≠ k ∥ U i − U k ∥ 2 , h_{U,n}:=\sup_{u\in[0,1]^d}\min_{1\le i\le m_n}\|u-U_i\|_2,\qquad q_{U,n}:=\frac{1}{2}\min_{i\ne k}\|U_i-U_k\|_2, h U , n := u ∈ [ 0 , 1 ] d sup 1 ≤ i ≤ m n min ∥ u − U i ∥ 2 , q U , n := 2 1 i = k min ∥ U i − U k ∥ 2 ,
as the fill distance and separation radius. Assume that h U , n / q U , n ≤ C h_{U,n}/q_{U,n}\le C h U , n / q U , n ≤ C and h U , n → 0 h_{U,n}\to0 h U , n → 0 . Thus the rescaled sites are quasi-uniform and, in particular, h U , n ≍ q U , n ≍ m n − 1 / d h_{U,n}\asymp q_{U,n}\asymp m_n^{-1/d} h U , n ≍ q U , n ≍ m n − 1/ d . The spatial spline resolution satisfies J s h U , n → 0 J_s h_{U,n}\to0 J s h U , n → 0 .
Theorem 3.16 (Spatial-spline local mean: global supnorm rate). *Let Assumptions 2.1, 2.2, 3.1–3.3, 3.9, 3.14, 3.15 be satisfied. Take p n = K s K m p_n=K_sK_m p n = K s K m in the definition of b n b_n b n . If
J s d J m { J s − d log ( K s K m ) / ( M 1 , n J m ) + log ( K s K m ) M 1 , n } → 0 , J_s^dJ_m\left\{\sqrt{J_s^{-d}\log(K_sK_m)/(M_{1,n}J_m)}+\frac{\log(K_sK_m)}{M_{1,n}}\right\}\to0, J s d J m { J s − d log ( K s K m ) / ( M 1 , n J m ) + M 1 , n log ( K s K m ) } → 0 , then
sup u ∈ [ 0 , 1 ] d sup t ∈ [ 0 , 1 ] ∣ M ^ J s , J m s p ( u , t ) − M ( u , t ) ∣ (19) \sup_{u\in[0,1]^d}\sup_{t\in[0,1]}\left|\widehat{M}^{\mathrm{sp}}_{J_s,J_m}(u,t)-M(u,t)\right| \tag*{(19)} u ∈ [ 0 , 1 ] d sup t ∈ [ 0 , 1 ] sup M J s , J m sp ( u , t ) − M ( u , t ) (19) = O p ( J s − q u ∗ + J m − q m ∗ + L n − 1 + J s d log ( K s K m ) ( J m m n N ‾ n + 1 m e f f , n ) + J s d J m δ n − d b n d log ( K s K m ) m n ) . =O_p\left(J_s^{-q_u^*}+J_m^{-q_m^*}+L_n^{-1}+\sqrt{J_s^d\log(K_sK_m)\left(\frac{J_m}{m_n\overline{N}_n}+\frac{1}{m_{\mathrm{eff},n}}\right)}+\frac{J_s^dJ_m\delta_n^{-d}b_n^d\log(K_sK_m)}{m_n}\right). = O p ( J s − q u ∗ + J m − q m ∗ + L n − 1 + J s d log ( K s K m ) ( m n N n J m + m eff , n 1 ) + m n J s d J m δ n − d b n d log ( K s K m ) ) . Note that the approximation errors have the different order J s − q u ∗ J_s^{-q_u^*} J s − q u ∗ and J m − q m ∗ J_m^{-q_m^*} J m − q m ∗ in the space and time direction. The observation-level component allocates m n N ‾ n m_n\overline{N}_n m n N n measurements across J s d J_s^d J s d spatial degrees of freedom, whereas the curve-level component allocates the dependence-adjusted size m e f f , n m_{\mathrm{eff},n} m eff , n across the same spatial resolution.
Corollary 3.17 (Spatial phase transition). *Let the assumptions of Theorem 3.16 be satisfied and assume that δ n − d b n d J m { J s d log ( K s K m ) / [ m n ( Θ n + J m / N ‾ n ) ] } 1 / 2 → 0 \delta_n^{-d}b_n^dJ_m\{J_s^d\log(K_sK_m)/[m_n(\Theta_n+J_m/\overline{N}_n)]\}^{1/2}\to0 δ n − d b n d J m { J s d log ( K s K m ) / [ m n ( Θ n + J m / N n )] } 1/2 → 0 . Then for fixed spatial resolution J s J_s J s and with logarithmic factors suppressed, the time-direction phase transition boundary is
N ‾ n ≍ ( m e f f , n / J s d ) 1 / ( 2 q m ∗ ) Θ n . \overline{N}_n\asymp\frac{(m_{\mathrm{eff},n}/J_s^d)^{1/(2q_m^*)}}{\Theta_n}. N n ≍ Θ n ( m eff , n / J s d ) 1/ ( 2 q m ∗ ) . Below this phase transition boundary, we have J m ≍ ( m n N ‾ n / J s d ) 1 / ( 2 q m ∗ + 1 ) J_m\asymp(m_n\overline{N}_n/J_s^d)^{1/(2q_m^*+1)} J m ≍ ( m n N n / J s d ) 1/ ( 2 q m ∗ + 1 ) and
R m , n : = sup u ∈ [ 0 , 1 ] d sup t ∈ [ 0 , 1 ] ∣ M ^ J s , J m s p ( u , t ) − M ( u , t ) ∣ = O p { J s − q u ∗ + L n − 1 + ( m n N ‾ n / J s d ) − q m ∗ / ( 2 q m ∗ + 1 ) } . R_{m,n}:=\sup_{u\in[0,1]^d}\sup_{t\in[0,1]}\left|\widehat{M}^{\mathrm{sp}}_{J_s,J_m}(u,t)-M(u,t)\right|
=O_p\left\{J_s^{-q_u^*}+L_n^{-1}+(m_n\overline{N}_n/J_s^d)^{-q_m^*/(2q_m^*+1)}\right\}. R m , n := u ∈ [ 0 , 1 ] d sup t ∈ [ 0 , 1 ] sup M J s , J m sp ( u , t ) − M ( u , t ) = O p { J s − q u ∗ + L n − 1 + ( m n N n / J s d ) − q m ∗ / ( 2 q m ∗ + 1 ) } . At or above it, we have J m ≍ ( m e f f , n / J s d ) 1 / ( 2 q m ∗ ) J_m\asymp(m_{\mathrm{eff},n}/J_s^d)^{1/(2q_m^*)} J m ≍ ( m eff , n / J s d ) 1/ ( 2 q m ∗ ) and R m , n = O p { J s − q u ∗ + L n − 1 + ( J s d / m e f f , n ) 1 / 2 } R_{m,n}=O_p\{J_s^{-q_u^*}+L_n^{-1}+(J_s^d/m_{\mathrm{eff},n})^{1/2}\} R m , n = O p { J s − q u ∗ + L n − 1 + ( J s d / m eff , n ) 1/2 } . If L n − 1 L_n^{-1} L n − 1 is of lower order, then the joint minimization in the sparse-time regime gives
R m , n = O p { ( m n N ‾ n ) − q u ∗ q m ∗ / ( 2 q u ∗ q m ∗ + d q m ∗ + q u ∗ ) } , R_{m,n}=O_p\left\{(m_n\overline{N}_n)^{-q_u^*q_m^*/(2q_u^*q_m^*+dq_m^*+q_u^*)}\right\}, R m , n = O p { ( m n N n ) − q u ∗ q m ∗ / ( 2 q u ∗ q m ∗ + d q m ∗ + q u ∗ ) } , whereas in the dense-time spatial regime the rate is given by m e f f , n − q u ∗ / ( 2 q u ∗ + d ) m_{\mathrm{eff},n}^{-q_u^*/(2q_u^*+d)} m eff , n − q u ∗ / ( 2 q u ∗ + d ) .*
The two local-mean procedures have the same leading stochastic structure when their spatial resolutions are matched. Setting a n ≍ J s − 1 a_n\asymp J_s^{-1} a n ≍ J s − 1 makes m e f f , n a n d m_{\mathrm{eff},n}a_n^d m eff , n a n d and m e f f , n / J s d m_{\mathrm{eff},n}/J_s^d m eff , n / J s d equivalent, and hence aligns their time-direction phase-transition boundaries up to logarithmic factors. Their distinction lies in approximation: a local-constant kernel has spatial bias a n γ u a_n^{\gamma_u} a n γ u , whereas a tensor-product spline achieves J s − q u ∗ J_s^{-q_u^*} J s − q u ∗ under the stronger anisotropic smoothness condition.
Simulation study
In this section, we investigate the finite sample properties of the proposed estimators. All reported Monte Carlo averages are based on 200 replications. We use the values m n = 400 m_n=400 m n = 400 , 900, 1600, 2500, 3600, 6400, and 10000 for the number of spatial locations. We first consider the problem of estimating a stationary mean and then retain the same spatial and temporal sampling designs for the problem of estimating a locally stationary mean.
Stationary mean
In dimension d = 2 d = 2 d = 2 , we generated the data from model (1) with m ( s , t ) = 2 sin ( π t ) m(s,t) = 2\sin(\pi t) m ( s , t ) = 2 sin ( π t ) , σ ( s , t ) = 0.3 \sigma(s,t) = 0.3 σ ( s , t ) = 0.3 , T i j ∼ i i d Unif [ 0 , 1 ] T_{ij} \mathrel{\overset{\mathrm{iid}}{\sim}} \operatorname{Unif}[0,1] T ij ∼ iid Unif [ 0 , 1 ] , and independent standard normal distributed errors ϵ i j \epsilon_{ij} ϵ ij . The time points and measurement errors are independent of the latent process. At site S i S_i S i , we define the random field by X ( S i , t ) = ∑ k = 1 4 ξ i k ϕ k ( t ) X(S_i,t) = \sum_{k=1}^{4}\xi_{ik}\phi_k(t) X ( S i , t ) = ∑ k = 1 4 ξ ik ϕ k ( t ) , where
( ϕ 1 ( t ) , ϕ 2 ( t ) , ϕ 3 ( t ) , ϕ 4 ( t ) ) T = 2 ( sin ( 2 π t ) , cos ( 2 π t ) , sin ( 4 π t ) , cos ( 4 π t ) ) T . \left(\phi_1(t),\phi_2(t),\phi_3(t),\phi_4(t)\right)^\mathsf{T}
= \sqrt{2}\left(\sin(2\pi t),\cos(2\pi t),\sin(4\pi t),\cos(4\pi t)\right)^\mathsf{T}. ( ϕ 1 ( t ) , ϕ 2 ( t ) , ϕ 3 ( t ) , ϕ 4 ( t ) ) T = 2 ( sin ( 2 π t ) , cos ( 2 π t ) , sin ( 4 π t ) , cos ( 4 π t ) ) T .
and, for each k k k , ( ξ 1 k , … , ξ m n k ) T \left(\xi_{1k},\ldots,\xi_{m_nk}\right)^\mathsf{T} ( ξ 1 k , … , ξ m n k ) T is a centered Gaussian vector with Cov ( ξ i k , ξ ℓ k ) = 2 − k ρ r , k ( S i − S ℓ ) \operatorname{Cov}(\xi_{ik},\xi_{\ell k}) = 2^{-k}\rho_{r,k}(S_i-S_\ell) Cov ( ξ ik , ξ ℓ k ) = 2 − k ρ r , k ( S i − S ℓ ) for some function ρ i , k \rho_{i,k} ρ i , k . The four score vectors are mutually i i i independent. Writing h = S i − S ℓ h=S_i-S_\ell h = S i − S ℓ , the five choices for the function ρ i , k \rho_{i,k} ρ i , k are
ρ 0 , k ( h ) = 1 ( h = 0 ) , (20) \rho_{0,k}(h)=\mathbf{1}(h=0), \tag*{(20)} ρ 0 , k ( h ) = 1 ( h = 0 ) , (20)
ρ 1 , k ( h ) = exp ( − ∥ h ∥ / α k ) , (21) \rho_{1,k}(h)=\exp\left(-\lVert h\rVert/\alpha_k\right), \tag*{(21)} ρ 1 , k ( h ) = exp ( − ∥ h ∥ / α k ) , (21)
ρ 2 , k ( h ) = exp ( − ∥ h ∥ 2 / α k 2 ) , (22) \rho_{2,k}(h)=\exp\left(-\lVert h\rVert^2/\alpha_k^2\right), \tag*{(22)} ρ 2 , k ( h ) = exp ( − ∥ h ∥ 2 / α k 2 ) , (22)
ρ 3 , k ( h ) = ( 1 + 3 ∥ h ∥ / α k ) exp ( − 3 ∥ h ∥ / α k ) , (23) \rho_{3,k}(h)=\left(1+\sqrt{3}\lVert h\rVert/\alpha_k\right)\exp\left(-\sqrt{3}\lVert h\rVert/\alpha_k\right), \tag*{(23)} ρ 3 , k ( h ) = ( 1 + 3 ∥ h ∥ / α k ) exp ( − 3 ∥ h ∥ / α k ) , (23)
ρ 4 , k ( h ) = exp [ − { ( h 1 / a 1 k ) 2 + ( h 2 / a 2 k ) 2 } 1 / 2 ] . (24) \rho_{4,k}(h)=\exp\left[-\left\{(h_1/a_{1k})^2+(h_2/a_{2k})^2\right\}^{1/2}\right]. \tag*{(24)} ρ 4 , k ( h ) = exp [ − { ( h 1 / a 1 k ) 2 + ( h 2 / a 2 k ) 2 } 1/2 ] . (24)
Here ( α 1 , … , α 4 ) T = ( 0.4 , 0.6 , 0.8 , 1.0 ) T (\alpha_1,\ldots,\alpha_4)^\mathsf{T}=(0.4,0.6,0.8,1.0)^\mathsf{T} ( α 1 , … , α 4 ) T = ( 0.4 , 0.6 , 0.8 , 1.0 ) T , while ( a 1 k , a 2 k ) T (a_{1k},a_{2k})^\mathsf{T} ( a 1 k , a 2 k ) T equals ( 0.3 , 0.8 ) T (0.3,0.8)^\mathsf{T} ( 0.3 , 0.8 ) T , ( 0.4 , 1.0 ) T (0.4,1.0)^\mathsf{T} ( 0.4 , 1.0 ) T , ( 0.5 , 1.2 ) T (0.5,1.2)^\mathsf{T} ( 0.5 , 1.2 ) T , and ( 0.6 , 1.4 ) T (0.6,1.4)^\mathsf{T} ( 0.6 , 1.4 ) T for k = 1 , … , 4 k=1,\ldots,4 k = 1 , … , 4 . Thus the scenario (D0) gives scores with independent components, and (D1)–(D4) use exponential, Gaussian, Matérn-3/2, and anisotropic exponential covariance functions.
For each m n m_n m n , the sites form a m n × m n \sqrt{m_n}\times\sqrt{m_n} m n × m n grid. In the sparse space regime, the nearest-neighbor distance of the spatial locations is one and the spatial area is of order m n m_n m n . In the dense space regime, the spatial locations are given by ( a / m n 1 / 4 , b / m n 1 / 4 ) (a/m_n^{1/4},b/m_n^{1/4}) ( a / m n 1/4 , b / m n 1/4 ) , so the nearest-neighbor distance is m n − 1 / 4 m_n^{-1/4} m n − 1/4 and the volume of the spatial domain is of order m n 1 / 2 m_n^{1/2} m n 1/2 . We therefore have m e f f , n = m n m_{\mathrm{eff},n}=m_n m eff , n = m n under (D0) and under sparse-space scenarios (D1)–(D4), and m e f f , n = m n 1 / 2 m_{\mathrm{eff},n}=m_n^{1/2} m eff , n = m n 1/2 under dense-space scenarios (D1)–(D4).
For the sparse time regime, the numbers of time points at each locations are sampled uniformly from the integers between 2 m n 1 / 5 2m_n^{1/5} 2 m n 1/5 and 4 m n 1 / 5 4m_n^{1/5} 4 m n 1/5 , while in the dense time regime are sampled uniformly from the integers between 1.5 m n 1 / 2 1.5m_n^{1/2} 1.5 m n 1/2 and 2.5 m n 1 / 2 2.5m_n^{1/2} 2.5 m n 1/2 . They are kept fixed for each simulated scenario, and the number of time points is the same for the sparse and the dense space regime. The additional simulation experiment in Section S.1.3 in the supplementary material replaces m n m_n m n by m e f f , n m_{\mathrm{eff},n} m eff , n in these two ranges.
The stationary estimator (2) uses cubic B-splines with dimension proportional to ( m e f f , n N ˉ n ) 1 / 5 (m_{\mathrm{eff},n}\bar{N}_n)^{1/5} ( m eff , n N ˉ n ) 1/5 in the sparse time and to m e f f , n 1 / 4 m_{\mathrm{eff},n}^{1/4} m eff , n 1/4 in the dense time regime. The sup-norm of m ^ J m − m \hat{m}_{J_m}-m m ^ J m − m is calculated as a maximum over 101 equally spaced points.
In Figure 1 we display the simulated sup-norm error for the different cases. The detailed numbers are shown in Table S.1 of supplementary materials. We observe that the errors are decreasing with increasing m n m_n m n for all five dependence choices. In scenario (D0), changing the distance between sites has little effect because the scores are independent. Under
Figure 1. Simulated sup-norm error for the estimator (2) for different values of m n m_n m n under scenario (D0)–(D4).
(D1)–(D4), the errors remain close to the independent case in sparse-space regime, whereas the differences become larger and decrease more slowly in dense-space regime because nearby sites are more strongly correlated and m e f f , n m_{\mathrm{eff},n} m eff , n grows as m n 1 / 2 m_n^{1/2} m n 1/2 . Dense time sampling decreases the error in the sparse space regime and under (D0). In the dense space regime (lower panels in Figure 1 ), the sup-norm errors in models (D1)–(D4) are close and occasionally reverse because additional observations within a curve reduce the measurement error but not the spatially dependent part of the process. These patterns agree with spatial functional results in which mean estimation depends on both spatial dependence and the distances between sites [20 , 42 ] .
Locally stationary mean
In this section, we investigate the performance of the estimator of a locally stationary mean using the same sites, observation times, measurement errors, component and dependence functions of scenario (D0)–(D4), and we set U i = ( ( a − 1 ) / ( m n − 1 ) , ( b − 1 ) / ( m n − 1 ) ) T U_i = ((a - 1)/(\sqrt{m_n} - 1), (b - 1)/(\sqrt{m_n} - 1))^\mathsf{T} U i = (( a − 1 ) / ( m n − 1 ) , ( b − 1 ) / ( m n − 1 ) ) T . The rescaled mean function is given by M ( u , t ) = 2 sin ( π t ) + 0.75 ( u 1 − 1 / 2 ) cos ( 2 π t ) M(u,t) = 2\sin(\pi t) + 0.75(u_1 - 1/2)\cos(2\pi t) M ( u , t ) = 2 sin ( π t ) + 0.75 ( u 1 − 1/2 ) cos ( 2 π t ) , and the variance of marginal scores ξ i k \xi_{ik} ξ ik is changed from 2 − k 2^{-k} 2 − k to λ k ( U i ) = 2 − k { 1 + 0.4 ( U i 1 + U i 2 − 1 ) } \lambda_k(U_i) = 2^{-k}\{1 + 0.4(U_{i1} + U_{i2} - 1)\} λ k ( U i ) = 2 − k { 1 + 0.4 ( U i 1 + U i 2 − 1 )} , which gives { λ k ( U i ) λ k ( U ℓ ) } 1 / 2 ρ r , k ( S i − S ℓ ) \{\lambda_k(U_i)\lambda_k(U_\ell)\}^{1/2}\rho_{r,k}(S_i - S_\ell) { λ k ( U i ) λ k ( U ℓ ) } 1/2 ρ r , k ( S i − S ℓ ) for Cov ( ξ i k , ξ ℓ k ) \operatorname{Cov}(\xi_{ik}, \xi_{\ell k}) Cov ( ξ ik , ξ ℓ k ) .
For the local-kernel estimator (6) , we evaluate the sup-norm of M ^ J m ( u 0 , ⋅ ) − M ( u 0 , ⋅ ) \widehat{M}_{J_m}(u_0,\cdot) - M(u_0,\cdot) M J m ( u 0 , ⋅ ) − M ( u 0 , ⋅ ) at point u 0 = ( 0.5 , 0.5 ) T u_0 = (0.5, 0.5)^\mathsf{T} u 0 = ( 0.5 , 0.5 ) T over the same 101 equally spaced time points. At this location, M ( u 0 , t ) = m ( t ) M(u_0,t)=m(t) M ( u 0 , t ) = m ( t ) , so the stationary and local targets have temporal shape. The bandwidth is proportional to ( m e f f , n N ‾ n ) − 2 / 9 (m_{\mathrm{eff},n}\overline{N}_n)^{-2/9} ( m eff , n N n ) − 2/9 in the sparse time and m e f f , n − 1 / 4 m_{\mathrm{eff},n}^{-1/4} m eff , n − 1/4 in the dense time regime. The time dimension follows the stationary rule after replacing m e f f , n m_{\mathrm{eff},n} m eff , n by m e f f , n a n 2 m_{\mathrm{eff},n}a_n^2 m eff , n a n 2 .
For comparison over a spatial region, we evaluate the local-kernel estimator (6) and spatial-spline estimator (8) of the grid { 0.2 , 0.35 , 0.5 , 0.65 , 0.8 } 2 \{0.2,0.35,0.5,0.65,0.8\}^2 { 0.2 , 0.35 , 0.5 , 0.65 , 0.8 } 2 and 101 equally spaced time points, and take the maximum error over all 25 × 101 25 \times101 25 × 101 points. For the spatial spline, K s K_s K s is max { 4 , ⌈ 0.5 m e f f , n 1 / 4 ⌉ } 2 \max\left\{4,\left\lceil0.5m_{\mathrm{eff},n}^{1/4}\right\rceil\right\}^2 max { 4 , ⌈ 0.5 m eff , n 1/4 ⌉ } 2 , and its time dimension follows the stationary rule with m e f f , n m_{\mathrm{eff},n} m eff , n replaced by m e f f , n / K s m_{\mathrm{eff},n}/K_s m eff , n / K s .
Figure 2 shows the sup-norm error of the estimator (6) and the detailed results are shown in Table S.2 in the supplementary materials. We observe the same effect of spatial dependence as in the stationary experiment. Errors decrease rapidly in the sparse space regime but more slowly in dense space regime under (D1)–(D4). Dense time sampling gives a clearer reduction of the error because the estimator uses only sites near the point u 0 u_0 u 0 .
Figure 2. Simulated sup-norm errors for the local-kernel estimator (2.5) over all seven values of m n m_n m n under scenarios (D0)–(D4).
In Table S.3 in the supplementary material, we display the simulated sup-norm error of the estimators (6) and (8) . We observe that the errors also decrease substantially in the sparse space regime. The maximal error of the spatial spline estimator decreases faster than that of the local-kernel estimator because it fits the complete surface using all sites. In the dense space regime, both methods improve more slowly. In the sparse time regime, the spatial spline estimators becomes more accurate as m n m_n m n increases, whereas under dense time regime the local-kernel estimator remains more accurate over the reported range.
Finally, we study the robustness of the estimators with respect to the score distribution. More specifically, we repeated the experiment for scenario (D1) replacing either the Gaussian scores or the Gaussian measurement errors by standardized uniform, Laplace, or t 5 t_{5} t 5 variables. Note that the Laplace and t 5 t_{5} t 5 distributions do not satisfy the sub-Gaussian assumptions of the theory. For the measurement-error experiment, each distribution was considered with σ ( s , t ) = 0.3 \sigma(s,t)=0.3 σ ( s , t ) = 0.3 and with σ ( s , t ) = 0.3 { 1 + 0.5 sin ( π t ) } \sigma(s,t)=0.3\{1+0.5\sin(\pi t)\} σ ( s , t ) = 0.3 { 1 + 0.5 sin ( π t )} . The results for the four score distributions are displayed in Table S.4 in the supplementary material for the stationary and the locally stationary mean function, where m n = 10000 m_n=10000 m n = 10000 . We do not observe substantial differences for the four score distributions. Corresponding results for alternative distributions of the measurement-error are given in Tables S.5 and S.6 in the supplementary material and show a similar picture. Therefore, the main differences in these experiments are due to spatial sampling and dependence rather than to the particular score or error distribution.
Applications
We illustrate the application of the estimators in three public environmental data sets: North Atlantic sea-surface temperature data (NOAA OISST v2.1) representing spatially and temporally dense data, California ground-level ozone data (EPA AQS ozone) representing spatially sparse and temporally dense data; California stream nitrate data (WQP/USGS nitrate) representing representing spatially and temporally sparse data. The latter example is discussed in Section S.1.2 of the supplementary materials. The spatial and temporal resolutions of the three examples are summarized in Table 1 .
Table 1. Sampling structure for three data sets: Coverage is the percentage of possible calendar-day records after the stated site and quality filters.
North Atlantic sea-surface temperature
The NOAA OISST Version 2.1 Climate Data Record contains North Atlantic sea-surface temperatures and serves as an example of spatially and temporally dense data. It combines bias-adjusted satellite, ship, buoy, and Argo observations on a complete 0.25° global grid, available daily since September 1981 ([23 ] (NOAA National Centers for Environmental Information 2020)). We analyzed 2023 SST over 30.125°–49.625°N and 79.875°–40.375°W, retained every second grid point, and required at least 360 days per ocean cell. All 2565 retained cells had 364 observations, yielding a regular 0.5° grid with 99.7% calendar-day coverage. The day of year was rescaled to the interval [ 0 , 1 ] [0,1] [ 0 , 1 ] . The stationary estimator (2.1) used cubic B-splines with 14 interior knots. The local-kernel estimator (2.5) used weights { 1 − ( r / 225 ) 2 } 2 1 { r < 225 } \{1-(r/225)^2\}^2\mathbf{1}\{r<225\} { 1 − ( r /225 ) 2 } 2 1 { r < 225 } , where r r r is great-circle distance in kilometers. We compared three targets in the warmer Sargasso Sea region (SS1–SS3) and three near the colder Labrador Current (LC1–LC3), retaining the Gulf Stream target (GS) as a further comparison.
In Figure 3 we display the surface of the annual mean temperatures (upper left panel) and the estimators of the mean curves at different spatial locations. We observe a clear annual cycle with substantial regional differences. The Gulf Stream and Sargasso targets (GS and SS1) remain warmer than the Labrador target (LC1) throughout the year. The annual-mean SS1–LC1 contrast is 18.0°C, while the seasonal amplitude is about 15.7°C at LC1 and 9°C at GS and SS1.
Figure 3. North Atlantic SST. Upper left: annual mean SST on the 0.5 ∘ 0.5^\circ 0. 5 ∘ grid and seven target locations. Upper right: pooled mean estimate under stationary assumption and local mean estimate at location GS, SS1, and LC1. Lower panels: local mean estimate at Sargasso and Labrador locations, with separate temperature scales.
Within the Sargasso region, the three curves differ by at most 1.03°C and peak within four days in August (lower left panel). The Labrador curves peak later and differ by up to 3.18°C, with the southern target generally warmer (lower right panel). A common mean therefore describes the selected Sargasso locations more closely than the Labrador locations. Across regions, the differences in level and seasonality favor local rather than basin-wide anomaly baselines.
California ground-level ozone
The U.S. Environmental Protection Agency AQS distributes quality-assured ambient pollution data through AirData [37 ] . To illustrate the estimators for spatially sparse and temporally dense data, we used California ozone records (parameter 44201) from the 2023 daily file. Ozone forms through photochemical reactions involving nitrogen oxides and volatile organic compounds; hot, sunny conditions often raise concentrations, while transport and meteorology produce regional differences [38 ] . We retained “8-HR RUN AVG BEGIN HOUR” records for the 2015 standard, no exceptional-event flag, and at least 75% daily completeness. The response was the daily maximum 8-hour average in ppb. Duplicate site–day pollutant-occurrence codes were reduced to their maximum, and each site was required to have at least 300 valid days. The final sample contains 131 irregular monitors with values N i N_i N i ranging from 304 to 363 and a median of 355. The day of year was rescaled to the interval [ 0 , 1 ] [0,1] [ 0 , 1 ] . The stationary estimator (2.1) used cubic B-splines with 14 interior knots. The local-kernel estimator (2.5) used weights { 1 − ( r / 175 ) 2 } 2 1 { r < 175 } \{1-(r/175)^2\}^2\mathbf{1}\{r<175\} { 1 − ( r /175 ) 2 } 2 1 { r < 175 } , where r r r is great-circle distance in kilometers. To examine variation within the higher- and lower-ozone regions, we compared targets near Fresno, Bakersfield, and Merced in the Central Valley (CV1–CV3), and San Francisco, San Jose, and Santa Rosa in the Bay Area (SF1–SF3). We also included a Los Angeles Basin target (LA).
The statewide mean displayed in Figure 4 rises into summer and declines in autumn, while local means reveal distinct regional patterns. Los Angeles peaks near 68 ppb in summer. The Central Valley curves peak in July at 55.9–65.3 ppb, whereas the Bay Area curves remain lower and peak in April. The Bay Area curves are also closer to one another: their pairwise root-mean-square differences are 0.64–1.16 ppb, compared with 8.29–13.83 ppb between Valley and Bay targets. These patterns favor a common mean for the selected Bay locations but reveal appreciable variation within the Valley. Local means thus distinguish the stronger summer elevation inland from the earlier coastal peak, which the statewide curve obscures.
Figure 4. California ozone. Upper left: monitors colored by annual mean daily maximum 8-hour ozone and seven target locations. Upper right: pooled mean estimate under stationarity assumption and local mean estimates at locations LA, CV1, and SF1. Lower panels: local mean estimate Central Valley and Bay Area locations, with separate concentration scales.
Concluding remarks
This paper develops a unified mean-estimation framework for functional observations whose spatial and temporal sampling densities may vary simultaneously. Across the stationary, local-kernel, and spatial-spline estimators, independent observation-level variation accumulates according to the nominal number of sampled measurements, whereas curve-level spatial variation is governed by m e f f , n m_{\mathrm{eff},n} m eff , n . Localization changes these two information scales to m n a n d N ˉ n m_n a_n^d \bar{N}_n m n a n d N ˉ n and m e f f , n a n d m_{\mathrm{eff},n}a_n^d m eff , n a n d , while a global spatial spline distributes them across J s d J_s^d J s d spatial degrees of freedom. Their balance determines the time-sampling phase transition boundary. The simulations support the resulting information-scale ordering, and the applications illustrate how the estimators recover common temporal structure and regional variation under distinct spatial and temporal sampling designs.
The analysis establishes uniform convergence rate under explicit spatial mixing and design conditions. Matching lower bounds would clarify the optimality of the dependence-adjusted phase boundaries, while simultaneous confidence bands and spatially dependent bootstrap methods would extend the framework from estimation to inference. Other useful directions include data-driven spatial resolution, adaptive smoothing under heterogeneous site density, and joint estimation of the mean and covariance structure.
Declaration of AI use
For this work, generative AI tools were used to edit language, search the literature, develop code, and help check the authors’ mathematical arguments. Before incorporating any AI-assisted material into the manuscript or code, the authors verified it. They take full responsibility for the work.
Data availability statement
The sea-surface temperature data are publicly available from https://doi.org/10.25921/RE9P-PT57. The ozone data are available from the U.S. Environmental Protection Agency’s https://aqs.epa.gov/aqsweb/airdata/download_files.html, and the nitrate data are available from the https://www.waterqualitydata.us/.
Acknowledgements
The work of the authors has been partially supported by the Deutsche Forschungsgemeinschaft (DFG): TRR 391 Spatio-temporal Statistics for the Transition of Energy and Transport (520388526), and Research Unit 5381 Mathematical Statistics in the Information Age (460867398).
References
Supplementary materials
Numerical results
Simulation results in Section 4
The tables in this section give the numerical results discussed in § 4 . Tables S.1 and S.2 report selected values of the errors plotted in Figure 1 and 2 for the stationary and local-kernel estimators, respectively. Table S.3 compares the local-kernel and spatial-spline estimators over the spatial region, and Table S.4 examines different score distributions.
Table S.1. Simulated sup-norm error for the estimator $(2.1). For each m n m_n m n , the five subcolumns correspond to the different scenarios (D0)–(D4).
Table S.2. Simulated sup-norm error for the local-kernel estimator $(2.5) at u 0 = ( 0.5 , 0.5 ) T u_0 = (0.5, 0.5)^\mathsf{T} u 0 = ( 0.5 , 0.5 ) T . For each m n m_n m n , the five subcolumns correspond to the different scenarios (D0)–(D4).
Table S.3. Simulated uniform sup-norm errors for the local-kernel estimator ( x r e f ‘ 2.5 ‘ ) ∗ a n d s p a t i a l − s p l i n e e s t i m a t o r ∗ ({xref}`2.5`) *and spatial-spline estimator* ( x r e f ‘2.5‘ ) ∗ an d s p a t ia l − s pl in ees t ima t or ∗ (2.7) under scenario (D1).
Table S.4. Simulated sup-norm errors for the stationary estimator (2) and pointwise local-kernel estimator (6) for four different score distributions in scenario ( D 1 ) (D1) ( D 1 ) (m n = 10000 m_n = 10000 m n = 10000 ) .
California stream nitrate
The Water Quality Portal (WQP) integrates publicly available discrete water-quality data from the United States Geological Survey (USGS), the Environmental Protection Agency (EPA), and other monitoring organizations [32 ] . We queried the National Water Information System (NWIS) provider only for California stream sites with nitrate results during 2021–2023 representing a spatially sparse and temporally sparse data example. To keep the measurement definition uniform, we retained accepted dissolved-nitrate results reported under USGS parameter 00618 in mg/L as N, and averaged replicate determinations from the same site and day. This yielded 122 sites and 439 distinct site–day observations. Each site had 1 − 28 1-28 1 − 28 observations over three years, with a median of three, giving a temporally sparse design. Nitrate is a mobile form of nitrogen in streams, and excessive nutrient enrichment can contribute to eutrophication [39 ] .
For the right-skewed nitrate concentrations, we fitted log ( 1 + Y i j ) \log(1+Y_{ij}) log ( 1 + Y ij ) using four cubic basis functions with no interior knot. The curves display exp { fitted log-scale curve } − 1 \exp\{\text{fitted log-scale curve}\}-1 exp { fitted log-scale curve } − 1 , a back-transformed summary of the mean on the log scale. This low-dimensional fit captures broad seasonal variation in the sparse observations. We compared the stationary statewide curve with a local curve centered at ( 34.05 ∘ N , 118.25 ∘ W ) (34.05^\circ\mathrm{N}, 118.25^\circ\mathrm{W}) ( 34.0 5 ∘ N , 118.2 5 ∘ W ) in Southern California, using a 500-km biweight bandwidth. The local window contains 25 sites and 83 observations over 30 year–month combinations.
Figure S.1 shows that the observations are concentrated around the Sacramento–San Joaquin Delta. The statewide fit declines from spring to early autumn before rising toward winter; the Southern California fit instead has a marked spring elevation and approaches the statewide level later in the year. The fitted seasonal patterns thus differ between the two spatial scales.
Figure S.1. California WQP/USGS dissolved-nitrate application. The left panel shows the 122 retained stream sites colored by their back-transformed sitewise average of log ( 1 + Y i j ) \log(1 + Y_{ij}) log ( 1 + Y ij ) ; the cross marks the Southern California (SC) target. The right panel shows the back-transformed stationary and SC local log-scale fits.
Additional simulation results
The stationary and locally stationary models, spatial grids, and tuning rules are those in § 4.1 and § 4.2 . All reported errors are averaged over 200 replications. In the tables, Space–time lists the spatial sampling regime first and the temporal sampling regime second. Under scenarios (D1)–(D4) in the dense space regime, the score vectors ( ξ 1 k , … , ξ m n , k ) T (\xi_{1k}, \ldots, \xi_{m_n,k})^\mathsf{T} ( ξ 1 k , … , ξ m n , k ) T were generated on the regular grid by exact block-circulant embedding of the stated covariance functions.
For the measurement-error experiments under (D1), we retain Gaussian score vectors and replace the measurement errors by standardized Gaussian, uniform, Laplace, or t 5 t_5 t 5 variables. We use m n = 10000 m_n = 10000 m n = 10000 and the ranges for N i N_i N i specified in § 4.1 . Tables S.5 and S.6 use σ ( s , t ) = 0.3 \sigma(s,t) = 0.3 σ ( s , t ) = 0.3 and σ ( s , t ) = 0.3 { 1 + 0.5 sin ( π t ) } \sigma(s,t) = 0.3\{1 + 0.5\sin(\pi t)\} σ ( s , t ) = 0.3 { 1 + 0.5 sin ( π t )} , respectively. Stationary errors are computed for estimator (2) over the 101 time points used in the main article. Local errors use estimator (6) at u 0 = ( 0.5 , 0.5 ) T u_0 = (0.5, 0.5)^\mathsf{T} u 0 = ( 0.5 , 0.5 ) T , taking the maximum over the same time points.
Table S.5: Sup-norm errors for the stationary estimator (2) and local-kernel estimator (6) at u 0 = ( 0.5 , 0.5 ) T u_0 = (0.5, 0.5)^\mathsf{T} u 0 = ( 0.5 , 0.5 ) T under (D1) with four measurement-error distributions and σ ( s , t ) = 0.3 \sigma(s,t) = 0.3 σ ( s , t ) = 0.3 , at m n = 10000 m_n = 10000 m n = 10000 , with N i N_i N i determined by m n m_n m n .
Table S.5.
Table S.6: Sup-norm errors for the stationary estimator (2) and local-kernel estimator (6) at u 0 = ( 0.5 , 0.5 ) T u_0 = (0.5, 0.5)^\mathsf{T} u 0 = ( 0.5 , 0.5 ) T under (D1) with four measurement-error distributions and σ ( s , t ) = 0.3 { 1 + 0.5 sin ( π t ) } \sigma(s,t) = 0.3\{1 + 0.5\sin(\pi t)\} σ ( s , t ) = 0.3 { 1 + 0.5 sin ( π t )} , at m n = 10000 m_n = 10000 m n = 10000 , with N i N_i N i determined by m n m_n m n .
Table S.6.
Table Table S.4 gives the corresponding experiment that changes the score innovation distribution while retaining Gaussian errors with σ ( s , t ) = 0.3 \sigma(s,t)=0.3 σ ( s , t ) = 0.3 . The score vectors are obtained by multiplying standardized innovations by the same covariance square root. The three experiments were run separately, so their Gaussian columns provide separate comparisons. Across these experiments, the effects of changing the score innovation or error distribution are small relative to the differences between spatial sampling regimes. The Laplace and t 5 t_5 t 5 settings provide finite-sample evidence beyond the sub-Gaussian conditions used in the theory.
In the main experiment, the two ranges for N i N_i N i depend on m n m_n m n and are therefore the same across spatial designs at a given m n m_n m n . Here we replace m n m_n m n by m e f f , n m_{\mathrm{eff},n} m eff , n in those ranges: N i N_i N i is sampled uniformly from the integers between 2 m e f f , n 1 / 5 2m_{\mathrm{eff},n}^{1/5} 2 m eff , n 1/5 and 4 m e f f , n 1 / 5 4m_{\mathrm{eff},n}^{1/5} 4 m eff , n 1/5 in the sparse time regime, and between 1.5 m e f f , n 1 / 2 1.5m_{\mathrm{eff},n}^{1/2} 1.5 m eff , n 1/2 and 2.5 m e f f , n 1 / 2 2.5m_{\mathrm{eff},n}^{1/2} 2.5 m eff , n 1/2 in the dense time regime, with endpoints rounded down. We use the same values of m e f f , n m_{\mathrm{eff},n} m eff , n as in § 4.1 . This comparison examines whether the effect of spatial dependence persists when the number of observations per site also varies with the spatial design. The models, tuning rules, and error evaluation grids remain unchanged.
Table S.7. Sup-norm errors for the stationary estimator (2) with N i N_i N i determined by m e f f , n m_{\mathrm{eff},n} m eff , n . For each m n m_n m n , the five subcolumns correspond to (D0)–(D4).
Table S.7 shows that the stationary errors decrease as m n m_n m n increases. The decrease remains slower in the dense space regime under (D1)–(D4), even though the numbers of observations per site are now determined by m e f f , n m_{\mathrm{eff},n} m eff , n .
The comparison at u 0 u_0 u 0 uses the locally stationary model in § 4.2 , including the spatially varying mean and score variances. Errors are evaluated at u 0 = ( 0.5 , 0.5 ) ⊤ u_0=(0.5,0.5)^\top u 0 = ( 0.5 , 0.5 ) ⊤ over 101 time points. For the regional comparison, both estimators are evaluated on the same 25 spatial points and 101 time points as in the main article, taking the maximum over all 25 × 101 25\times101 25 × 101 combinations.
Table S.8. Simulated sup-norm errors for the local-kernel estimator (6) at u 0 = ( 0.5 , 0.5 ) ⊤ u_0=(0.5,0.5)^\top u 0 = ( 0.5 , 0.5 ) ⊤ , with N i N_i N i determined by m e f f , n m_{\mathrm{eff},n} m eff , n . For each m n m_n m n , the five subcolumns correspond to (D0)–(D4).
Table S.9. Regional sup-norm errors for the local-kernel estimator (6) and spatial-spline estimator (8) under (D1), with N i N_i N i determined by m e f f , n m_{\mathrm{eff},n} m eff , n .
Table S.8 and Figure S.2 show the same effect of spatial dependence for the local mean at u 0 u_0 u 0 . In the regional comparison in Table S.9, in the dense space regime, the spatial-spline estimator becomes more accurate as m n m_n m n increases under sparse time sampling, whereas the local-kernel estimator remains more accurate under dense time sampling. These patterns agree with the main experiment, although the errors in the dense space regime change when fewer observations are taken within each site.
Figure S.2. Simulated sup-norm errors for the local-kernel estimator (6) at u 0 = ( 0.5 , 0.5 ) ⊤ u_0=(0.5,0.5)^\top u 0 = ( 0.5 , 0.5 ) ⊤ over all seven values of m n m_n m n under (D0)–(D4), with N i N_i N i determined by m e f f , n m_{\mathrm{eff},n} m eff , n .
The distribution experiments under (D1) use m n = 10000 m_n=10000 m n = 10000 and the alternative ranges for N i N_i N i above. We change either the score innovations or the measurement errors, retaining the other part of the model as Gaussian. As in the main experiment, the score comparison uses σ ( s , t ) = 0.3 \sigma(s,t)=0.3 σ ( s , t ) = 0.3 , and the error comparison uses both the constant and time-varying scales. The three experiments are run separately.
Table S.10. Sup-norm errors for the stationary estimator (2) and local-kernel estimator (6) at u 0 = ( 0.5 , 0.5 ) ⊤ u_0=(0.5,0.5)^\top u 0 = ( 0.5 , 0.5 ) ⊤ under (D1) with four score distributions at m n = 10000 m_n=10000 m n = 10000 , with N i N_i N i determined by m e f f , n m_{\mathrm{eff},n} m eff , n .
Table S.11. Sup-norm errors for the stationary estimator (2) and local-kernel estimator (6) at u 0 = ( 0.5 , 0.5 ) ⊤ u_0=(0.5,0.5)^\top u 0 = ( 0.5 , 0.5 ) ⊤ under (D1) with four measurement-error distributions and σ ( s , t ) = 0.3 \sigma(s,t)=0.3 σ ( s , t ) = 0.3 , at m n = 10000 m_n=10000 m n = 10000 , with N i N_i N i determined by m e f f , n m_{\mathrm{eff},n} m eff , n .
Table S.12. Sup-norm errors for the stationary estimator (2) and local-kernel estimator (6) at u 0 = ( 0.5 , 0.5 ) T u_0 = (0.5, 0.5)^\mathsf{T} u 0 = ( 0.5 , 0.5 ) T under (D1) with four measurement-error distributions and σ ( s , t ) = 0.3 { 1 + 0.5 sin ( π t ) } \sigma(s,t) = 0.3\{1 + 0.5\sin(\pi t)\} σ ( s , t ) = 0.3 { 1 + 0.5 sin ( π t )} , at m n = 10000 m_n = 10000 m n = 10000 , with N i N_i N i determined by m e f f , n m_{\mathrm{eff},n} m eff , n .
Tables S.10 to S.12 show that the differences between spatial sampling regimes remain larger than those between the score or measurement-error distributions. Together with the preceding (D0)–(D4) comparisons, these results indicate that the effect of spatial dependence persists under both choices of the ranges for N i N_i N i .
Technical lemmas
Lemma S.2.1. Under Assumption 2.2, there exists a constant C > 0 C > 0 C > 0 such that, for every site S i S_i S i and every integer u ≥ 0 u \ge0 u ≥ 0 ,
# { k : u ≤ ∥ S k − S i ∥ 2 < u + 1 } ≤ C δ n − d ( 1 + u ) d − 1 \#\{k : u \le\lVert S_k-S_i\rVert_2 < u+1\} \le C\delta_n^{-d}(1+u)^{d-1} # { k : u ≤ ∥ S k − S i ∥ 2 < u + 1 } ≤ C δ n − d ( 1 + u ) d − 1 Proof. Fix S i S_i S i . If u ≤ ∥ S k − S i ∥ 2 < u + 1 u \le\lVert S_k-S_i\rVert_2 < u+1 u ≤ ∥ S k − S i ∥ 2 < u + 1 , every point in the design cell containing S k S_k S k lies in the enlarged annulus
A u , n ( S i ) : = { s ∈ R d : ( u − d δ n ) + ≤ ∥ s − S i ∥ 2 < u + 1 + d δ n } . A_{u,n}(S_i) := \left\{s \in\mathbb{R}^d : (u-\sqrt{d}\,\delta_n)_+ \le\lVert s-S_i\rVert_2 < u+1+\sqrt{d}\,\delta_n\right\}. A u , n ( S i ) := { s ∈ R d : ( u − d δ n ) + ≤ ∥ s − S i ∥ 2 < u + 1 + d δ n } . The relevant design cells are disjoint and each has volume δ n d \delta_n^d δ n d , so their number is at most ∣ A u , n ( S i ) ∣ / δ n d |A_{u,n}(S_i)|/\delta_n^d ∣ A u , n ( S i ) ∣/ δ n d . By Assumption 2.2, δ n = O ( 1 ) \delta_n=O(1) δ n = O ( 1 ) , and hence
∣ A u , n ( S i ) ∣ = v d [ ( u + 1 + d δ n ) d − ( u − d δ n ) + d ] ≤ C ( 1 + u ) d − 1 , |A_{u,n}(S_i)|=v_d\left[(u+1+\sqrt{d}\,\delta_n)^d-(u-\sqrt{d}\,\delta_n)_+^d\right]\le C(1+u)^{d-1}, ∣ A u , n ( S i ) ∣ = v d [ ( u + 1 + d δ n ) d − ( u − d δ n ) + d ] ≤ C ( 1 + u ) d − 1 , where the last bound follows from the mean-value theorem for large u u u and by enlarging C C C for bounded u u u . Dividing by δ n d \delta_n^d δ n d proves the result.
Lemma S.2.2. Under Assumptions 2.2 and 3.1, Θ n ≤ C δ n − d \Theta_n \le C\delta_n^{-d} Θ n ≤ C δ n − d and, consequently, m e f f , n ≥ c V n m_{\mathrm{eff},n} \ge cV_n m eff , n ≥ c V n .
Proof. By Lemma S.2.1,
∑ k ≠ i β X ( ∥ S i − S k ∥ 2 ) 1 / 2 ≤ ∑ u = 0 ∞ # { k : u ≤ ∥ S k − S i ∥ 2 < u + 1 } sup r ∈ [ u , u + 1 ) β X ( r ) 1 / 2 . \sum_{k\ne i}\beta_X(\lVert S_i-S_k\rVert_2)^{1/2}
\le\sum_{u=0}^{\infty}\#\{k:u\le\lVert S_k-S_i\rVert_2<u+1\}\sup_{r\in[u,u+1)}\beta_X(r)^{1/2}. k = i ∑ β X (∥ S i − S k ∥ 2 ) 1/2 ≤ u = 0 ∑ ∞ # { k : u ≤ ∥ S k − S i ∥ 2 < u + 1 } r ∈ [ u , u + 1 ) sup β X ( r ) 1/2 . Applying Lemma S.2.1 and Assumption 3.1,
∑ k ≠ i β X ( ∥ S i − S k ∥ 2 ) 1 / 2 ≤ C δ n − d ∑ u = 0 ∞ ( 1 + u ) d − 1 sup r ∈ [ u , u + 1 ) ( 1 + r ) − α β / 2 ≤ C δ n − d ∑ u = 0 ∞ ( 1 + u ) d − 1 − α β / 2 . \sum_{k\ne i}\beta_X(\lVert S_i-S_k\rVert_2)^{1/2}
\le C\delta_n^{-d}\sum_{u=0}^{\infty}(1+u)^{d-1}\sup_{r\in[u,u+1)}(1+r)^{-\alpha\beta/2}
\le C\delta_n^{-d}\sum_{u=0}^{\infty}(1+u)^{d-1-\alpha\beta/2}. k = i ∑ β X (∥ S i − S k ∥ 2 ) 1/2 ≤ C δ n − d u = 0 ∑ ∞ ( 1 + u ) d − 1 r ∈ [ u , u + 1 ) sup ( 1 + r ) − α β /2 ≤ C δ n − d u = 0 ∑ ∞ ( 1 + u ) d − 1 − α β /2 . Because α β > 2 d \alpha\beta>2d α β > 2 d , the exponent d − 1 − α β / 2 d-1-\alpha\beta/2 d − 1 − α β /2 is less than − 1 -1 − 1 , and the series converges. Thus, Θ n ≤ C δ n − d \Theta_n\le C\delta_n^{-d} Θ n ≤ C δ n − d . Since m n ≍ V n δ n − d m_n\asymp V_n\delta_n^{-d} m n ≍ V n δ n − d by Assumption 2.2, it follows that m e f f , n ≥ c V n m_{\mathrm{eff},n}\ge cV_n m eff , n ≥ c V n .
Proposition S.2.3. Assume Assumptions 2.2 and 3.1 and that there exists r 0 > 0 r_0>0 r 0 > 0 such that β X ( r 0 ) > 0 \beta_X(r_0)>0 β X ( r 0 ) > 0 . In the dense-space regime δ n → 0 \delta_n\to0 δ n → 0 , assume additionally Assumption 2.3. Then Θ n ≍ δ n − d \Theta_n\asymp\delta_n^{-d} Θ n ≍ δ n − d and m e f f , n ≍ V n m_{\mathrm{eff},n}\asymp V_n m eff , n ≍ V n .
Proof. The upper bound Θ n ≤ C δ n − d \Theta_n\le C\delta_n^{-d} Θ n ≤ C δ n − d follows from Lemma S.2.2. If δ n ≍ 1 \delta_n\asymp1 δ n ≍ 1 , then Θ n ≥ 1 ≍ δ n − d \Theta_n\ge1\asymp\delta_n^{-d} Θ n ≥ 1 ≍ δ n − d , so it remains to consider the dense-space regime δ n → 0 \delta_n\to0 δ n → 0 .
By Assumptions 2.2 and 2.3, min r L r , n → ∞ \min_r L_{r,n}\to\infty min r L r , n → ∞ . Let x n = ( L 1 , n / 2 , … , L d , n / 2 ) T x_n=(L_{1,n}/2,\ldots,L_{d,n}/2)^\mathsf{T} x n = ( L 1 , n /2 , … , L d , n /2 ) T , and choose the observed site S i S_i S i in the design cell containing x n x_n x n . For all sufficiently large n n n , S i S_i S i is at distance greater than r 0 r_0 r 0 from every face of S n S_n S n . The ball B ( S i , r 0 / 2 ) B(S_i,r_0/2) B ( S i , r 0 /2 ) therefore contains
at least c δ n − d c\delta_n^{-d} c δ n − d complete design cells, each of which contains one observed site. Every such site S k S_k S k satisfies ∥ S i − S k ∥ 2 ≤ r 0 / 2 \lVert S_i-S_k\rVert_2\le r_0/2 ∥ S i − S k ∥ 2 ≤ r 0 /2 . Since r ↦ β X ( r ) r\mapsto\beta_X(r) r ↦ β X ( r ) is nonincreasing,
∑ k ≠ i β X ( ∥ S i − S k ∥ 2 ) 1 / 2 ≥ c δ n − d β X ( r 0 ) 1 / 2 . \sum_{k\ne i}\beta_X\left(\lVert S_i-S_k\rVert_2\right)^{1/2}\ge c\delta_n^{-d}\beta_X(r_0)^{1/2}. k = i ∑ β X ( ∥ S i − S k ∥ 2 ) 1/2 ≥ c δ n − d β X ( r 0 ) 1/2 .
Thus Θ n ≥ c ′ δ n − d \Theta_n\ge c'\delta_n^{-d} Θ n ≥ c ′ δ n − d . Combining the two bounds gives Θ n ≍ δ n − d \Theta_n\asymp\delta_n^{-d} Θ n ≍ δ n − d , and m n ≍ V n δ n − d m_n\asymp V_n\delta_n^{-d} m n ≍ V n δ n − d from Assumption 2.2 then yields m e f f , n ≍ V n m_{\mathrm{eff},n}\asymp V_n m eff , n ≍ V n .
For the sparse two-dimensional simulation lattice, δ n ≍ 1 \delta_n\asymp1 δ n ≍ 1 and V n ≍ m n V_n\asymp m_n V n ≍ m n , so Lemma S.2.2 gives m e f f , n ≍ m n m_{\mathrm{eff},n}\asymp m_n m eff , n ≍ m n . For the dense lattice, δ n ≍ m n − 1 / 4 \delta_n\asymp m_n^{-1/4} δ n ≍ m n − 1/4 , V n ≍ m n 1 / 2 V_n\asymp m_n^{1/2} V n ≍ m n 1/2 . Whenever the simulated score field satisfies Assumption 3.1 and the short-range nonindependence condition of Proposition S.2.3, that proposition gives Θ n ≍ m n 1 / 2 \Theta_n\asymp m_n^{1/2} Θ n ≍ m n 1/2 and m e f f , n ≍ m n 1 / 2 m_{\mathrm{eff},n}\asymp m_n^{1/2} m eff , n ≍ m n 1/2 . For D0, Θ n = 1 \Theta_n=1 Θ n = 1 by independence.
Lemma S.2.4. Under Assumption 2.1, let G i : = σ ( T i j , ϵ i j : 1 ≤ j ≤ N i ) \mathcal{G}_i:=\sigma(T_{ij},\epsilon_{ij}:1\le j\le N_i) G i := σ ( T ij , ϵ ij : 1 ≤ j ≤ N i ) and let { Z i } i = 1 m n \{Z_i\}_{i=1}^{m_n} { Z i } i = 1 m n be real-valued random variables satisfying Z i ∈ F X ( { S i } ) ∨ G i Z_i\in\mathcal{F}_X(\{S_i\})\vee\mathcal{G}_i Z i ∈ F X ({ S i }) ∨ G i . Assume sup i ∥ Z i ∥ 4 < ∞ \sup_i\lVert Z_i\rVert_4<\infty sup i ∥ Z i ∥ 4 < ∞ . Then there exists C > 0 C>0 C > 0 such that, for every i i i ,
∑ k = 1 m n ∣ Cov ( Z i , Z k ) ∣ ≤ C Θ n sup 1 ≤ ℓ ≤ m n ∥ Z ℓ ∥ 4 2 , \sum_{k=1}^{m_n}\left\lvert\operatorname{Cov}(Z_i,Z_k)\right\rvert
\le C\Theta_n\sup_{1\le\ell\le m_n}\lVert Z_\ell\rVert_4^2, k = 1 ∑ m n ∣ Cov ( Z i , Z k ) ∣ ≤ C Θ n 1 ≤ ℓ ≤ m n sup ∥ Z ℓ ∥ 4 2 , and
Var ( ∑ i = 1 m n Z i ) ≤ C Θ n ∑ i = 1 m n ∥ Z i ∥ 4 2 . (25) \operatorname{Var}\left(\sum_{i=1}^{m_n}Z_i\right)
\le C\Theta_n\sum_{i=1}^{m_n}\lVert Z_i\rVert_4^2. \tag*{(25)} Var ( i = 1 ∑ m n Z i ) ≤ C Θ n i = 1 ∑ m n ∥ Z i ∥ 4 2 . (25)
Proof. By Assumption 2.1, the sigma-fields G i \mathcal{G}_i G i are independent across sites and jointly independent of the latent field. Adding these independent sigma-fields does not increase the absolute-regularity coefficient between the two point-site field sigma-fields. Hence, for i ≠ k i\ne k i = k , the strong-mixing coefficient between F X ( { S i } ) ∨ G i \mathcal{F}_X(\{S_i\})\vee\mathcal{G}_i F X ({ S i }) ∨ G i and F X ( { S k } ) ∨ G k \mathcal{F}_X(\{S_k\})\vee\mathcal{G}_k F X ({ S k }) ∨ G k is bounded by β X ( ∥ S i − S k ∥ 2 ) \beta_X(\lVert S_i-S_k\rVert_2) β X (∥ S i − S k ∥ 2 ) . Rio’s covariance inequality (Theorem 1.1 of Rio [35 ] ) therefore gives
∣ Cov ( Z i , Z k ) ∣ ≤ C β X ( ∥ S i − S k ∥ 2 ) 1 / 2 ∥ Z i ∥ 4 ∥ Z k ∥ 4 , i ≠ k . \left\lvert\operatorname{Cov}(Z_i,Z_k)\right\rvert
\le C\beta_X\left(\lVert S_i-S_k\rVert_2\right)^{1/2}\lVert Z_i\rVert_4\lVert Z_k\rVert_4,\qquad i\ne k. ∣ Cov ( Z i , Z k ) ∣ ≤ C β X ( ∥ S i − S k ∥ 2 ) 1/2 ∥ Z i ∥ 4 ∥ Z k ∥ 4 , i = k .
For i = k i=k i = k , Var ( Z i ) ≤ ∥ Z i ∥ 4 2 \operatorname{Var}(Z_i)\le\lVert Z_i\rVert_4^2 Var ( Z i ) ≤ ∥ Z i ∥ 4 2 . Summing over k k k and using equation (9) proves the first assertion. For the second, use ∥ Z i ∥ 4 ∥ Z k ∥ 4 ≤ ( 1 / 2 ) { ∥ Z i ∥ 4 2 + ∥ Z k ∥ 4 2 } \lVert Z_i\rVert_4\lVert Z_k\rVert_4\le(1/2)\{\lVert Z_i\rVert_4^2+\lVert Z_k\rVert_4^2\} ∥ Z i ∥ 4 ∥ Z k ∥ 4 ≤ ( 1/2 ) {∥ Z i ∥ 4 2 + ∥ Z k ∥ 4 2 } in the double covariance sum and again apply equation (9) .
For ( t , t ′ ) ∈ [ 0 , 1 ] 2 (t,t')\in[0,1]^2 ( t , t ′ ) ∈ [ 0 , 1 ] 2 , define the tensor-product basis B J ( 2 ) ( t , t ′ ) : = B J ( t ) ⊗ B J ( t ′ ) ∈ R K J 2 B_J^{(2)}(t,t'):=B_J(t)\otimes B_J(t')\in\mathbb{R}^{K_J^2} B J ( 2 ) ( t , t ′ ) := B J ( t ) ⊗ B J ( t ′ ) ∈ R K J 2 .
Lemma S.2.5. There exists a constant C > 0 C>0 C > 0 , depending only on the spline order and the quasi-uniformity constants of the knot sequence, such that, for every J J J :
(i) sup t ∈ [ 0 , 1 ] ∥ B J ( t ) ∥ 1 = 1 \sup_{t\in[0,1]}\lVert B_J(t)\rVert_1=1 sup t ∈ [ 0 , 1 ] ∥ B J ( t ) ∥ 1 = 1 and sup ( t , t ′ ) ∈ [ 0 , 1 ] 2 ∥ B J ( 2 ) ( t , t ′ ) ∥ 1 = 1 \sup_{(t,t')\in[0,1]^2}\lVert B_J^{(2)}(t,t')\rVert_1=1 sup ( t , t ′ ) ∈ [ 0 , 1 ] 2 ∥ B J ( 2 ) ( t , t ′ ) ∥ 1 = 1 ;
(ii) For each fixed t t t , at most p p p entries of B J ( t ) B_J(t) B J ( t ) are nonzero; for each fixed ( t , t ′ ) (t,t') ( t , t ′ ) , at most p 2 p^2 p 2 entries of B J ( 2 ) ( t , t ′ ) B_J^{(2)}(t,t') B J ( 2 ) ( t , t ′ ) are nonzero;
(iii) For all t , s ∈ [ 0 , 1 ] t,s \in[0,1] t , s ∈ [ 0 , 1 ] , ∥ B J ( t ) − B J ( s ) ∥ 1 ≤ C J ∣ t − s ∣ \lVert\mathbf{B}_J(t)-\mathbf{B}_J(s)\rVert_1 \le C J\lvert t-s\rvert ∥ B J ( t ) − B J ( s ) ∥ 1 ≤ C J ∣ t − s ∣ ;
(iv) For all ( t , t ′ ) , ( s , s ′ ) ∈ [ 0 , 1 ] 2 (t,t'),(s,s') \in[0,1]^2 ( t , t ′ ) , ( s , s ′ ) ∈ [ 0 , 1 ] 2 , ∥ B J ( 2 ) ( t , t ′ ) − B J ( 2 ) ( s , s ′ ) ∥ 1 ≤ C J ( ∣ t − s ∣ + ∣ t ′ − s ′ ∣ ) \lVert\mathbf{B}^{(2)}_J(t,t')-\mathbf{B}^{(2)}_J(s,s')\rVert_1 \le C J(\lvert t-s\rvert+\lvert t'-s'\rvert) ∥ B J ( 2 ) ( t , t ′ ) − B J ( 2 ) ( s , s ′ ) ∥ 1 ≤ C J (∣ t − s ∣ + ∣ t ′ − s ′ ∣) .
Proof. Properties (i) and (ii) follow from the partition-of-unity property and local support of normalized B-splines [12 Chapters IX–XI] . On a quasi-uniform mesh, the derivative of each basis function is bounded by C J CJ C J ; hence ∣ B a , J ( t ) − B a , J ( s ) ∣ ≤ C J ∣ t − s ∣ \lvert B_{a,J}(t)-B_{a,J}(s)\rvert\le CJ\lvert t-s\rvert ∣ B a , J ( t ) − B a , J ( s )∣ ≤ C J ∣ t − s ∣ . Summing over the at most 2 p 2p 2 p basis functions that are nonzero at either t t t or s s s proves (iii). For the tensor-product basis,
B J ( 2 ) ( t , t ′ ) − B J ( 2 ) ( s , s ′ ) = { B J ( t ) − B J ( s ) } ⊗ B J ( t ′ ) + B J ( s ) ⊗ { B J ( t ′ ) − B J ( s ′ ) } . \mathbf{B}^{(2)}_J(t,t')-\mathbf{B}^{(2)}_J(s,s')=\{\mathbf{B}_J(t)-\mathbf{B}_J(s)\}\otimes\mathbf{B}_J(t')+\mathbf{B}_J(s)\otimes\{\mathbf{B}_J(t')-\mathbf{B}_J(s')\}. B J ( 2 ) ( t , t ′ ) − B J ( 2 ) ( s , s ′ ) = { B J ( t ) − B J ( s )} ⊗ B J ( t ′ ) + B J ( s ) ⊗ { B J ( t ′ ) − B J ( s ′ )} .
therefore, (iv) follows from (i) and (iii).
Lemma S.2.6. Let
Q m , J : = E { B J ( T ) B J ( T ) ⊤ } , Q G , J : = E { B J ( 2 ) ( T 1 , T 2 ) B J ( 2 ) ( T 1 , T 2 ) ⊤ } . Q_{m,J}:=\mathbb{E}\left\{\mathbf{B}_J(T)\mathbf{B}_J(T)^\top\right\},\qquad Q_{G,J}:=\mathbb{E}\left\{\mathbf{B}^{(2)}_J(T_1,T_2)\mathbf{B}^{(2)}_J(T_1,T_2)^\top\right\}. Q m , J := E { B J ( T ) B J ( T ) ⊤ } , Q G , J := E { B J ( 2 ) ( T 1 , T 2 ) B J ( 2 ) ( T 1 , T 2 ) ⊤ } . where T , T 1 , T 2 T,T_1,T_2 T , T 1 , T 2 are independent and identically distributed (i.i.d.) with density f T f_T f T from Assumption 2.1 . Then there exist constants 0 < c < C < ∞ 0<c<C<\infty 0 < c < C < ∞ such that, for every J J J ,
c J − 1 ≤ λ min ( Q m , J ) ≤ λ max ( Q m , J ) ≤ C J − 1 , cJ^{-1}\le\lambda_{\min}(Q_{m,J})\le\lambda_{\max}(Q_{m,J})\le CJ^{-1}, c J − 1 ≤ λ m i n ( Q m , J ) ≤ λ m a x ( Q m , J ) ≤ C J − 1 , c J − 2 ≤ λ min ( Q G , J ) ≤ λ max ( Q G , J ) ≤ C J − 2 . cJ^{-2}\le\lambda_{\min}(Q_{G,J})\le\lambda_{\max}(Q_{G,J})\le CJ^{-2}. c J − 2 ≤ λ m i n ( Q G , J ) ≤ λ m a x ( Q G , J ) ≤ C J − 2 . Moreover,
∥ Q m , J − 1 ∥ ∞ ≤ C J , ∥ Q G , J − 1 ∥ ∞ ≤ C J 2 . \lVert Q_{m,J}^{-1}\rVert_\infty\le CJ,\qquad\lVert Q_{G,J}^{-1}\rVert_\infty\le CJ^2. ∥ Q m , J − 1 ∥ ∞ ≤ C J , ∥ Q G , J − 1 ∥ ∞ ≤ C J 2 . Proof. For the one-dimensional basis,
x ⊤ Q m , J x = E ( ∑ a = 1 K J x a B a , J ( T ) ) 2 = ∫ 0 1 ( ∑ a = 1 K J x a B a , J ( t ) ) 2 f T ( t ) d t . x^\top Q_{m,J}x=\mathbb{E}\left(\sum_{a=1}^{K_J}x_aB_{a,J}(T)\right)^2=\int_0^1\left(\sum_{a=1}^{K_J}x_aB_{a,J}(t)\right)^2f_T(t)\,\mathrm{d}t. x ⊤ Q m , J x = E ( a = 1 ∑ K J x a B a , J ( T ) ) 2 = ∫ 0 1 ( a = 1 ∑ K J x a B a , J ( t ) ) 2 f T ( t ) d t . Because f T f_T f T is bounded above and bounded away from zero, this quadratic form is equivalent to the L 2 L^2 L 2 norm of the spline g x ( t ) = ∑ a x a B a , J ( t ) g_x(t)=\sum_a x_aB_{a,J}(t) g x ( t ) = ∑ a x a B a , J ( t ) . For normalized B-splines on a quasi-uniform mesh, coefficient stability [12 Chapter XI] gives
c J − 1 ∥ x ∥ 2 2 ≤ ∫ 0 1 g x ( t ) 2 d t ≤ C J − 1 ∥ x ∥ 2 2 . cJ^{-1}\lVert x\rVert_2^2\le\int_0^1g_x(t)^2\,\mathrm{d}t\le CJ^{-1}\lVert x\rVert_2^2. c J − 1 ∥ x ∥ 2 2 ≤ ∫ 0 1 g x ( t ) 2 d t ≤ C J − 1 ∥ x ∥ 2 2 . This proves the eigenvalue bounds for Q m , J Q_{m,J} Q m , J . When needed, the same argument on [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 gives c J − 2 ∥ x ∥ 2 2 ≤ x ⊤ Q G , J x ≤ C J − 2 ∥ x ∥ 2 2 cJ^{-2}\lVert x\rVert_2^2\le x^\top Q_{G,J}x\le CJ^{-2}\lVert x\rVert_2^2 c J − 2 ∥ x ∥ 2 2 ≤ x ⊤ Q G , J x ≤ C J − 2 ∥ x ∥ 2 2 . The matrix Q m , J Q_{m,J} Q m , J is banded, with bandwidth depending only on p p p , and the inverse of a banded positive-definite matrix has exponentially decaying entries [14 ] . Its eigenvalue scale J − 1 J^{-1} J − 1 therefore gives ∥ Q m , J − 1 ∥ ∞ ≤ C J \lVert Q_{m,J}^{-1}\rVert_\infty\le CJ ∥ Q m , J − 1 ∥ ∞ ≤ C J . Independence and the tensor basis give Q G , J = Q m , J ⊗ Q m , J Q_{G,J}=Q_{m,J}\otimes Q_{m,J} Q G , J = Q m , J ⊗ Q m , J , so Q G , J − 1 = Q m , J − 1 ⊗ Q m , J − 1 Q_{G,J}^{-1}=Q_{m,J}^{-1}\otimes Q_{m,J}^{-1} Q G , J − 1 = Q m , J − 1 ⊗ Q m , J − 1 and ∥ Q G , J − 1 ∥ ∞ = ∥ Q m , J − 1 ∥ ∞ 2 \lVert Q_{G,J}^{-1}\rVert_\infty=\lVert Q_{m,J}^{-1}\rVert_\infty^2 ∥ Q G , J − 1 ∥ ∞ = ∥ Q m , J − 1 ∥ ∞ 2 . Consequently, ∥ Q m , J − 1 ∥ ∞ ≤ C J \lVert Q_{m,J}^{-1}\rVert_\infty\le CJ ∥ Q m , J − 1 ∥ ∞ ≤ C J , ∥ Q G , J − 1 ∥ ∞ ≤ C J 2 \lVert Q_{G,J}^{-1}\rVert_\infty\le CJ^2 ∥ Q G , J − 1 ∥ ∞ ≤ C J 2 .
Lemma S.2.7. Assume Assumptions 3.3 and 3.15, write
π i , n : = N i M 1 , n , g v ( u ) : = A s ( u ) T v , \pi_{i,n} := \frac{N_i}{M_{1,n}}, \qquad g_v(u) := \mathbf{A}_s(u)^\mathsf{T}v, π i , n := M 1 , n N i , g v ( u ) := A s ( u ) T v , and let v ∈ R K s v \in\mathbb{R}^{K_s} v ∈ R K s . There are constants 0 < c < C < ∞ 0 < c < C < \infty 0 < c < C < ∞ , independent of n n n , J s J_s J s , such that
c ∥ g v ∥ L 2 ( [ 0 , 1 ] d ) 2 ≤ ∑ i = 1 m n π i , n g v ( U i ) 2 ≤ C ∥ g v ∥ L 2 ( [ 0 , 1 ] d ) 2 , (26) c\lVert g_v\rVert_{L^2([0,1]^d)}^2 \le\sum_{i=1}^{m_n}\pi_{i,n}g_v(U_i)^2 \le C\lVert g_v\rVert_{L^2([0,1]^d)}^2, \tag*{(26)} c ∥ g v ∥ L 2 ([ 0 , 1 ] d ) 2 ≤ i = 1 ∑ m n π i , n g v ( U i ) 2 ≤ C ∥ g v ∥ L 2 ([ 0 , 1 ] d ) 2 , (26) and hence
c J s − d ∥ v ∥ 2 2 ≤ ∑ i = 1 m n π i , n g v ( U i ) 2 ≤ C J s − d ∥ v ∥ 2 2 . (27) cJ_s^{-d}\lVert v\rVert_2^2 \le\sum_{i=1}^{m_n}\pi_{i,n}g_v(U_i)^2 \le CJ_s^{-d}\lVert v\rVert_2^2. \tag*{(27)} c J s − d ∥ v ∥ 2 2 ≤ i = 1 ∑ m n π i , n g v ( U i ) 2 ≤ C J s − d ∥ v ∥ 2 2 . (27) For every spatial basis coordinate a a a ,
∑ i = 1 m n π i , n A a ( U i ) ≤ C J s − d , ∑ i = 1 m n π i , n A a ( U i ) 2 ≤ C J s − d . (28) \sum_{i=1}^{m_n}\pi_{i,n}A_a(U_i) \le CJ_s^{-d}, \qquad\sum_{i=1}^{m_n}\pi_{i,n}A_a(U_i)^2 \le CJ_s^{-d}. \tag*{(28)} i = 1 ∑ m n π i , n A a ( U i ) ≤ C J s − d , i = 1 ∑ m n π i , n A a ( U i ) 2 ≤ C J s − d . (28)
Proof. Let V i , n V_{i,n} V i , n be the Voronoi cell of U i U_i U i , intersected with [ 0 , 1 ] d [0,1]^d [ 0 , 1 ] d . The fill-distance bound gives diam ( V i , n ) ≤ 2 h U , n \operatorname{diam}(V_{i,n}) \le2h_{U,n} diam ( V i , n ) ≤ 2 h U , n . The separation radius implies that V i , n V_{i,n} V i , n contains the intersection of [ 0 , 1 ] d [0,1]^d [ 0 , 1 ] d with a ball of radius q U , n q_{U,n} q U , n centered at U i U_i U i , whereas the fill-distance bound places it inside a ball of radius h U , n h_{U,n} h U , n . Since h U , n / q U , n ≤ C h_{U,n}/q_{U,n} \le C h U , n / q U , n ≤ C and h U , n ≍ q U , n ≍ m n − 1 / d h_{U,n} \asymp q_{U,n} \asymp m_n^{-1/d} h U , n ≍ q U , n ≍ m n − 1/ d , c m n − 1 ≤ ∣ V i , n ∣ ≤ C m n − 1 cm_n^{-1} \le|V_{i,n}| \le Cm_n^{-1} c m n − 1 ≤ ∣ V i , n ∣ ≤ C m n − 1 uniformly, including boundary cells. Balanced counts give c / m n ≤ π i , n ≤ C / m n c/m_n \le\pi_{i,n} \le C/m_n c / m n ≤ π i , n ≤ C / m n , and hence π i , n ≍ ∣ V i , n ∣ \pi_{i,n} \asymp|V_{i,n}| π i , n ≍ ∣ V i , n ∣ .
Define the piecewise-constant sampler P n g P_ng P n g by P n g ( u ) = g ( U i ) P_ng(u) = g(U_i) P n g ( u ) = g ( U i ) on V i , n V_{i,n} V i , n . Then
∑ i π i , n g ( U i ) 2 ≍ ∥ P n g ∥ L 2 2 . \sum_i \pi_{i,n}g(U_i)^2 \asymp\lVert P_ng\rVert_{L^2}^2. i ∑ π i , n g ( U i ) 2 ≍ ∥ P n g ∥ L 2 2 .
Let Q J s Q_{J_s} Q J s be the quasi-uniform tensor-product spline mesh, whose cell diameters are of order H s ≍ J s − 1 H_s \asymp J_s^{-1} H s ≍ J s − 1 . Since h U , n / H s = J s h U , n → 0 h_{U,n}/H_s = J_sh_{U,n} \to0 h U , n / H s = J s h U , n → 0 , a segment joining u ∈ V i , n u \in V_{i,n} u ∈ V i , n to U i U_i U i crosses only a fixed number of neighboring mesh cells for all sufficiently large n n n . The piecewise mean-value inequality therefore gives
∣ g ( U i ) − g ( u ) ∣ ≤ C h U , n max x ∈ Q + ∥ ∇ g ( x ) ∥ 2 |g(U_i)-g(u)| \le Ch_{U,n}\max_{x\in Q^+}\lVert\nabla g(x)\rVert_2 ∣ g ( U i ) − g ( u ) ∣ ≤ C h U , n x ∈ Q + max ∥ ∇ g ( x ) ∥ 2
whenever u u u lies in a mesh cell Q Q Q , where Q + Q^+ Q + is a fixed enlargement of Q Q Q . Summing over the Voronoi partition and grouping by mesh cells yields
∥ P n g − g ∥ L 2 2 ≤ C h U , n 2 ∑ Q ∈ Q J s ∣ Q ∣ ∥ ∇ g ∥ L ∞ ( Q + ) 2 . \lVert P_ng-g\rVert_{L^2}^2 \le Ch_{U,n}^2\sum_{Q\in Q_{J_s}}|Q|\lVert\nabla g\rVert_{L^\infty(Q^+)}^2. ∥ P n g − g ∥ L 2 2 ≤ C h U , n 2 Q ∈ Q J s ∑ ∣ Q ∣ ∥ ∇ g ∥ L ∞ ( Q + ) 2 .
For fixed-order splines, the local polynomial inverse inequality ([36 Chapters 3 and 12] ) gives ∣ Q ∣ ∥ ∇ g ∥ L ∞ ( Q + ) 2 ≤ C ∥ ∇ g ∥ L 2 ( Q + + ) 2 |Q|\lVert\nabla g\rVert_{L^\infty(Q^+)}^2 \le C\lVert\nabla g\rVert_{L^2(Q^{++})}^2 ∣ Q ∣ ∥ ∇ g ∥ L ∞ ( Q + ) 2 ≤ C ∥ ∇ g ∥ L 2 ( Q ++ ) 2 , with another fixed enlargement Q + + Q^{++} Q ++ . These enlargements have uniformly bounded overlap, and the global spline inverse inequality gives
∥ P n g − g ∥ L 2 2 ≤ C h U , n 2 ∥ ∇ g ∥ L 2 2 ≤ C ( J s h U , n ) 2 ∥ g ∥ L 2 2 . \lVert P_ng-g\rVert_{L^2}^2 \le Ch_{U,n}^2\lVert\nabla g\rVert_{L^2}^2 \le C(J_sh_{U,n})^2\lVert g\rVert_{L^2}^2. ∥ P n g − g ∥ L 2 2 ≤ C h U , n 2 ∥ ∇ g ∥ L 2 2 ≤ C ( J s h U , n ) 2 ∥ g ∥ L 2 2 .
Thus ∥ P n g − g ∥ L 2 = o ( ∥ g ∥ L 2 ) \lVert P_ng-g\rVert_{L^2}=o(\lVert g\rVert_{L^2}) ∥ P n g − g ∥ L 2 = o (∥ g ∥ L 2 ) uniformly over the spline space. The triangle inequality proves ∥ P n g ∥ L 2 ≍ ∥ g ∥ L 2 \lVert P_ng\rVert_{L^2}\asymp\lVert g\rVert_{L^2} ∥ P n g ∥ L 2 ≍ ∥ g ∥ L 2 and hence (S.2.2). Coefficient stability for normalized tensor-product B-splines ([36 Chapter 12] ) gives ∥ g v ∥ L 2 2 ≍ J s − d ∥ v ∥ 2 2 \lVert g_v\rVert_{L^2}^2\asymp J_s^{-d}\lVert v\rVert_2^2 ∥ g v ∥ L 2 2 ≍ J s − d ∥ v ∥ 2 2 , proving (S.2.3).
Finally, each A a A_a A a is bounded by one and supported on a rectangle with volume O ( J s − d ) O(J_s^{-d}) O ( J s − d ) and side lengths of order J s − 1 J_s^{-1} J s − 1 . A packing argument using the separation radius shows that the number of sites in this support is at most
C ( 1 + J s − 1 q U , n ) d ≤ C m n J s − d . C\left(1+\frac{J_s^{-1}}{q_{U,n}}\right)^d\leq Cm_nJ_s^{-d}. C ( 1 + q U , n J s − 1 ) d ≤ C m n J s − d .
where the last inequality uses J s h U , n → 0 J_sh_{U,n}\to0 J s h U , n → 0 and q U , n ≍ m n − 1 / d q_{U,n}\asymp m_n^{-1/d} q U , n ≍ m n − 1/ d . Since π i , n ≤ C / m n \pi_{i,n}\leq C/m_n π i , n ≤ C / m n and 0 ≤ A a ≤ 1 0\leq A_a\leq1 0 ≤ A a ≤ 1 , both bounds in (S.2.4) follow. □ \square □
Lemma S.2.8. Let Z Z Z be a centered random element of a separable real Hilbert space H \mathbb{H} H , with E ∥ Z ∥ 2 < ∞ E\lVert Z\rVert^2<\infty E ∥ Z ∥ 2 < ∞ and covariance operator Σ \Sigma Σ . Suppose that, for every h ∈ H h\in\mathbb{H} h ∈ H ,
E exp { ⟨ h , Z ⟩ } ≤ exp { K 2 ⟨ h , Σ h ⟩ / 2 } . E\exp\{\langle h,Z\rangle\}\leq\exp\{K^2\langle h,\Sigma h\rangle/2\}. E exp {⟨ h , Z ⟩} ≤ exp { K 2 ⟨ h , Σ h ⟩ /2 } . Then, for every x > 0 x>0 x > 0 ,
P ( ∥ Z ∥ 2 > 2 K 2 tr ( Σ ) + 4 K 2 ∥ Σ ∥ o p x ) ≤ e − x . P\left(\lVert Z\rVert^2>2K^2\operatorname{tr}(\Sigma)+4K^2\lVert\Sigma\rVert_{\mathrm{op}}x\right)\leq e^{-x}. P ( ∥ Z ∥ 2 > 2 K 2 tr ( Σ ) + 4 K 2 ∥ Σ ∥ op x ) ≤ e − x .
Proof. The bound follows from Mollenhauer and Schillings ((2023, Proposition 2.2)) with variance proxy K 2 Σ K^2\Sigma K 2 Σ . We include a direct proof. If ∥ Σ ∥ o p = 0 \lVert\Sigma\rVert_{\mathrm{op}}=0 ∥ Σ ∥ op = 0 , then Z = 0 Z=0 Z = 0 almost surely. Otherwise, fix an orthonormal basis and let P r P_r P r be projection onto its first r r r coordinates. Write Σ r = P r Σ P r \Sigma_r=P_r\Sigma P_r Σ r = P r Σ P r on that finite-dimensional space. For an independent g ∼ N ( 0 , I r ) g\sim N(0,I_r) g ∼ N ( 0 , I r ) , Tonelli’s theorem and the assumed linear moment-generating function (MGF) bound give
E exp { t ∥ P r Z ∥ 2 } = E g E Z exp { 2 t ⟨ g , P r Z ⟩ } ≤ E g exp { K 2 t ⟨ g , Σ r g ⟩ } = det ( I r − 2 K 2 t Σ r ) − 1 / 2 . E\exp\{t\lVert P_rZ\rVert^2\}=E_gE_Z\exp\{\sqrt{2t}\langle g,P_rZ\rangle\}\leq E_g\exp\{K^2t\langle g,\Sigma_rg\rangle\}=\det(I_r-2K^2t\Sigma_r)^{-1/2}. E exp { t ∥ P r Z ∥ 2 } = E g E Z exp { 2 t ⟨ g , P r Z ⟩} ≤ E g exp { K 2 t ⟨ g , Σ r g ⟩} = det ( I r − 2 K 2 t Σ r ) − 1/2 .
For 0 ≤ t ≤ ( 4 K 2 ∥ Σ ∥ o p ) − 1 0\leq t\leq(4K^2\lVert\Sigma\rVert_{\mathrm{op}})^{-1} 0 ≤ t ≤ ( 4 K 2 ∥ Σ ∥ op ) − 1 , every eigenvalue u u u of 2 K 2 t Σ r 2K^2t\Sigma_r 2 K 2 t Σ r lies in [ 0 , 1 / 2 ] [0,1/2] [ 0 , 1/2 ] . Using − log ( 1 − u ) ≤ 2 u -\log(1-u)\leq2u − log ( 1 − u ) ≤ 2 u , E exp { t ∥ P r Z ∥ 2 } ≤ exp { 2 K 2 t tr ( Σ ) } E\exp\{t\lVert P_rZ\rVert^2\}\leq\exp\{2K^2t\operatorname{tr}(\Sigma)\} E exp { t ∥ P r Z ∥ 2 } ≤ exp { 2 K 2 t tr ( Σ )} . Since ∥ P r Z ∥ 2 ↑ ∥ Z ∥ 2 \lVert P_rZ\rVert^2\uparrow\lVert Z\rVert^2 ∥ P r Z ∥ 2 ↑ ∥ Z ∥ 2 , monotone convergence yields the same inequality for Z Z Z . Apply Markov’s inequality with t = ( 4 K 2 ∥ Σ ∥ o p ) − 1 t=(4K^2\lVert\Sigma\rVert_{\mathrm{op}})^{-1} t = ( 4 K 2 ∥ Σ ∥ op ) − 1 to obtain the result. □ \square □
Lemma S.2.9. Suppress the array index n n n . Assume Assumptions 2.1 and 3.2 and the stated measurability convention. For fixed N i ≥ 1 N_i\geq1 N i ≥ 1 , let X i ( t ) = X ( S i , t ) X_i(t)=X(S_i,t) X i ( t ) = X ( S i , t ) . Suppose that, for every i , k , t , t ′ i,k,t,t' i , k , t , t ′ , ∣ Cov { X i ( t ) , X k ( t ′ ) } ∣ ≤ C X κ i k \lvert\operatorname{Cov}\{X_i(t),X_k(t')\}\rvert\leq C_X\kappa_{ik} ∣ Cov { X i ( t ) , X k ( t ′ )}∣ ≤ C X κ ik , where ( κ i k ) (\kappa_{ik}) ( κ ik ) is symmetric and nonnegative, κ i i = 1 \kappa_{ii}=1 κ ii = 1 , and sup i ∑ k κ i k ≤ Θ \sup_i\sum_k\kappa_{ik}\leq\Theta sup i ∑ k κ ik ≤ Θ . Let J ≥ 1 J\geq1 J ≥ 1 . For q ≤ p q\leq p q ≤ p , let B q B_q B q be deterministic functions satisfying
sup q ∥ B q ∥ ∞ ≤ C B , ∫ ∣ B q ∣ f T + ∫ B q 2 f T ≤ C B / J . \sup_q\lVert B_q\rVert_\infty\leq C_B,\qquad\int\lvert B_q\rvert f_T+\int B_q^2f_T\leq C_B/J. q sup ∥ B q ∥ ∞ ≤ C B , ∫ ∣ B q ∣ f T + ∫ B q 2 f T ≤ C B / J . Let w i q w_{iq} w i q be deterministic weights and let d i q d_{iq} d i q be deterministic uniformly bounded functions. Define
S q = ∑ i w i q ∑ j = 1 N i B q ( T i j ) { d i q ( T i j ) + X i ( T i j ) + σ i ( T i j ) ε i j } − ∑ i w i q N i ∫ B q ( t ) d i q ( t ) f T ( t ) d t . (29) S_q=\sum_i w_{iq}\sum_{j=1}^{N_i}B_q(T_{ij})\{d_{iq}(T_{ij})+X_i(T_{ij})+\sigma_i(T_{ij})\varepsilon_{ij}\}-\sum_iw_{iq}N_i\int B_q(t)d_{iq}(t)f_T(t)\,dt. \tag*{(29)} S q = i ∑ w i q j = 1 ∑ N i B q ( T ij ) { d i q ( T ij ) + X i ( T ij ) + σ i ( T ij ) ε ij } − i ∑ w i q N i ∫ B q ( t ) d i q ( t ) f T ( t ) d t . (29)
Put
M 1 = M 1 , n , H q = ∑ i w i q 2 N i , F q = ∑ i w i q 2 N i 2 , E q = max i w i q 2 N i , w ∗ = max i , q ∣ w i q ∣ , M_1=M_{1,n},\qquad H_q=\sum_i w_{iq}^2N_i,\qquad F_q=\sum_i w_{iq}^2N_i^2,\qquad E_q=\max_i w_{iq}^2N_i,\qquad w_*=\max_{i,q}|w_{iq}|, M 1 = M 1 , n , H q = i ∑ w i q 2 N i , F q = i ∑ w i q 2 N i 2 , E q = i max w i q 2 N i , w ∗ = i , q max ∣ w i q ∣ ,
and
V q = H q J + F q − H q + ( Θ − 1 ) F q J 2 , ℓ = log ( e + p ) , L = { log ( e + M 1 + p ) } 1 / 2 . V_q=\frac{H_q}{J}+\frac{F_q-H_q+(\Theta-1)F_q}{J^2},\qquad\ell=\log(e+p),\qquad L=\{\log(e+M_1+p)\}^{1/2}. V q = J H q + J 2 F q − H q + ( Θ − 1 ) F q , ℓ = log ( e + p ) , L = { log ( e + M 1 + p ) } 1/2 .
For arrays with M 1 + p → ∞ M_1+p\to\infty M 1 + p → ∞ ,
max q ≤ p ∣ S q ∣ = O p ( max q V q ℓ + ℓ max q Θ E q / J + w ∗ ℓ L ) . \max_{q\le p}|S_q|=O_p\left(\max_q\sqrt{V_q\ell}+\ell\max_q\sqrt{\Theta E_q/J}+w_*\ell L\right). q ≤ p max ∣ S q ∣ = O p ( q max V q ℓ + ℓ q max Θ E q / J + w ∗ ℓ L ) .
The assertion permits fixed p p p .
Proof. We first justify the use of the joint linear sub-Gaussian condition for integrated functionals. For deterministic c i ∈ L 2 ( f T ) c_i\in L^2(f_T) c i ∈ L 2 ( f T ) , consider I = ∑ i ∫ c i ( t ) X i ( t ) f T ( t ) d t I=\sum_i\int c_i(t)X_i(t)f_T(t)\,dt I = ∑ i ∫ c i ( t ) X i ( t ) f T ( t ) d t . Introduce auxiliary independent samples from f T f_T f T , independent of the field, and approximate each integral by an empirical average of size r r r . Conditional on the field these averages are unbiased. The expected squared approximation error, averaged over the auxiliary samples and the field, is at most ( C / r ) ∑ i ∥ c i ∥ L 2 ( f T ) 2 (C/r)\sum_i\|c_i\|_{L^2(f_T)}^2 ( C / r ) ∑ i ∥ c i ∥ L 2 ( f T ) 2 . Consequently, for every r r r , deterministic sampling points can be selected so that the resulting finite-point linear combinations converge to I I I in L 2 ( P ) L^2(P) L 2 ( P ) . The original joint variance-relative condition implies a centered linear MGF bound with a uniform constant. Passing along an almost surely convergent subsequence and using Fatou’s lemma, together with convergence of variances, proves the same variance-relative MGF bound for I I I . No sample-path continuity is used.
Decompose the latent part of S q S_q S q as G q + R q G_q+R_q G q + R q , where
G q = ∑ i w i q N i ∫ B q ( t ) X i ( t ) f T ( t ) d t G_q=\sum_i w_{iq}N_i\int B_q(t)X_i(t)f_T(t)\,dt G q = i ∑ w i q N i ∫ B q ( t ) X i ( t ) f T ( t ) d t
and
R q = ∑ i w i q { ∑ j = 1 N i B q ( T i j ) X i ( T i j ) − N i ∫ B q ( t ) X i ( t ) f T ( t ) d t } . R_q=\sum_i w_{iq}\left\{\sum_{j=1}^{N_i}B_q(T_{ij})X_i(T_{ij})-N_i\int B_q(t)X_i(t)f_T(t)\,dt\right\}. R q = i ∑ w i q { j = 1 ∑ N i B q ( T ij ) X i ( T ij ) − N i ∫ B q ( t ) X i ( t ) f T ( t ) d t } .
The covariance bound, the L 1 L^1 L 1 support bound, and 2 ∣ a b ∣ ≤ a 2 + b 2 2|ab|\le a^2+b^2 2∣ ab ∣ ≤ a 2 + b 2 imply Var ( G q ) ≤ C Θ F q / J 2 ≤ C V q \operatorname{Var}(G_q)\le C\Theta F_q/J^2\le CV_q Var ( G q ) ≤ C Θ F q / J 2 ≤ C V q , where the last inequality follows from J ≥ 1 J\ge1 J ≥ 1 . The integrated linear MGF bound therefore gives, for x ≥ 1 x\ge1 x ≥ 1 ,
P ( max q ∣ G q ∣ > C max q V q x ) ≤ C p e − c x . P\left(\max_q|G_q|>C\max_q\sqrt{V_qx}\right)\le Cpe^{-cx}. P ( q max ∣ G q ∣ > C q max V q x ) ≤ C p e − c x .
Conditional on the latent information H n \mathcal{H}_n H n , all observation times remain independent. Moreover,
Var ( R q ∣ H n ) ≤ Q q : = ∑ i w i q 2 N i ∫ B q ( t ) 2 X i ( t ) 2 f T ( t ) d t . \operatorname{Var}(R_q\mid\mathcal{H}_n)\le Q_q:=\sum_i w_{iq}^2N_i\int B_q(t)^2X_i(t)^2f_T(t)\,dt. Var ( R q ∣ H n ) ≤ Q q := i ∑ w i q 2 N i ∫ B q ( t ) 2 X i ( t ) 2 f T ( t ) d t .
Let Z q = ( ∣ w i q ∣ N i B q X i ) i Z_q = (|w_{iq}|\sqrt{N_i}B_qX_i)_i Z q = ( ∣ w i q ∣ N i B q X i ) i be a random element of H = ⨁ i L 2 ( f T ) \mathbb{H} = \bigoplus_i L^2(f_T) H = ⨁ i L 2 ( f T ) , and let Σ q \Sigma_q Σ q denote its covariance operator. Then Q q = ∥ Z q ∥ 2 Q_q = \|Z_q\|^2 Q q = ∥ Z q ∥ 2 and tr ( Σ q ) ≤ C H q / J \operatorname{tr}(\Sigma_q) \le CH_q/J tr ( Σ q ) ≤ C H q / J . For v = ( v i ) i ∈ H v = (v_i)_i \in\mathbb{H} v = ( v i ) i ∈ H , Cauchy–Schwarz gives ∫ ∣ B q v i ∣ f T ≤ C J − 1 / 2 ∥ v i ∥ L 2 ( f T ) \int|B_qv_i|f_T \le CJ^{-1/2}\|v_i\|_{L^2(f_T)} ∫ ∣ B q v i ∣ f T ≤ C J − 1/2 ∥ v i ∥ L 2 ( f T ) . Thus, by the covariance bound and its row-sum bound,
⟨ v , Σ q v ⟩ ≤ C J ∑ i , k κ i k ∣ w i q ∣ N i ∣ w k q ∣ N k ∥ v i ∥ ∥ v k ∥ ≤ C Θ E q J ∑ i ∥ v i ∥ 2 . \langle v,\Sigma_qv\rangle\le\frac{C}{J}\sum_{i,k}\kappa_{ik}|w_{iq}|\sqrt{N_i}|w_{kq}|\sqrt{N_k}\|v_i\|\|v_k\| \le\frac{C\Theta E_q}{J}\sum_i\|v_i\|^2. ⟨ v , Σ q v ⟩ ≤ J C i , k ∑ κ ik ∣ w i q ∣ N i ∣ w k q ∣ N k ∥ v i ∥∥ v k ∥ ≤ J C Θ E q i ∑ ∥ v i ∥ 2 .
Hence ∥ Σ q ∥ o p ≤ C Θ E q / J \|\Sigma_q\|_{\mathrm{op}} \le C\Theta E_q/J ∥ Σ q ∥ op ≤ C Θ E q / J . The integrated linear MGF argument established above applies to every ⟨ v , Z q ⟩ \langle v,Z_q\rangle ⟨ v , Z q ⟩ . Lemma S.2.8 and a union bound therefore yield
P ( ∃ q : Q q > C { H q / J + Θ E q x / J } ) ≤ p e − x . P\left(\exists q: Q_q > C\{H_q/J+\Theta E_qx/J\}\right) \le pe^{-x}. P ( ∃ q : Q q > C { H q / J + Θ E q x / J } ) ≤ p e − x .
Set L 0 = C 0 { log ( e + M 1 + p ) } 1 / 2 L_0=C_0\{\log(e+M_1+p)\}^{1/2} L 0 = C 0 { log ( e + M 1 + p ) } 1/2 and let X i ( L 0 ) X_i^{(L_0)} X i ( L 0 ) be X i X_i X i clipped pointwise to [ − L 0 , L 0 ] [-L_0,L_0] [ − L 0 , L 0 ] . Define R q ( L 0 ) R_q^{(L_0)} R q ( L 0 ) by replacing both the sampled values and the integral in R q R_q R q by their clipped counterparts. Conditional on H n \mathcal{H}_n H n , its centered summands are independent, bounded by C w ∗ L 0 Cw_*L_0 C w ∗ L 0 , and have total variance at most Q q Q_q Q q . Conditional Bernstein’s inequality [4 Theorem 2.10] , followed by the preceding norm bound, gives
P ( max q ∣ R q ( L 0 ) ∣ > C { max q H q x / J + x max q Θ E q / J + w ∗ L 0 x } ) ≤ C p e − c x . P\left(\max_q|R_q^{(L_0)}|>C\left\{\max_q\sqrt{H_qx/J}+x\max_q\sqrt{\Theta E_q/J}+w_*L_0x\right\}\right)\le Cpe^{-cx}. P ( q max ∣ R q ( L 0 ) ∣ > C { q max H q x / J + x q max Θ E q / J + w ∗ L 0 x } ) ≤ C p e − c x .
The clipping replacement is uniform over coordinates without a factor p p p . Indeed, boundedness of every B q B_q B q implies
max q ∣ R q − R q ( L 0 ) ∣ ≤ C w ∗ [ ∑ i , j ∣ X i ( T i j ) − X i ( L 0 ) ( T i j ) ∣ + ∑ i N i ∫ ∣ X i ( t ) − X i ( L 0 ) ( t ) ∣ f T ( t ) d t ] . \max_q|R_q-R_q^{(L_0)}|\le Cw_*\left[\sum_{i,j}|X_i(T_{ij})-X_i^{(L_0)}(T_{ij})|+\sum_iN_i\int|X_i(t)-X_i^{(L_0)}(t)|f_T(t)\,dt\right]. q max ∣ R q − R q ( L 0 ) ∣ ≤ C w ∗ [ i , j ∑ ∣ X i ( T ij ) − X i ( L 0 ) ( T ij ) ∣ + i ∑ N i ∫ ∣ X i ( t ) − X i ( L 0 ) ( t ) ∣ f T ( t ) d t ] .
Pointwise sub-Gaussian tails and exogeneity of the times give E max q ∣ R q − R q ( L 0 ) ∣ ≤ C w ∗ M 1 e − c L 0 2 = o ( w ∗ ) E\max_q|R_q-R_q^{(L_0)}|\le Cw_*M_1e^{-cL_0^2}=o(w_*) E max q ∣ R q − R q ( L 0 ) ∣ ≤ C w ∗ M 1 e − c L 0 2 = o ( w ∗ ) when C 0 C_0 C 0 is sufficiently large. Markov’s inequality makes this replacement negligible. No supremum-norm tail assumption on the sample paths, and no union bound over a continuum of times, is used.
The centered deterministic-design contribution is a sum of independent bounded variables with variance at most C H q / J CH_q/J C H q / J . For the measurement-error contribution, independence over ( i , j ) (i,j) ( i , j ) , bounded σ i \sigma_i σ i , the support bound on B q B_q B q , and the sub-Gaussian error moments imply the Bernstein moment bounds with variance proxy C H q / J CH_q/J C H q / J and scale C w ∗ Cw_* C w ∗ . A union bound therefore controls these two contributions by C { max q H q x / J + w ∗ x } C\{\max_q\sqrt{H_qx/J}+w_*x\} C { max q H q x / J + w ∗ x } outside an event of probability at most C p e − c x Cpe^{-cx} C p e − c x .
Finally, fix ε > 0 \varepsilon>0 ε > 0 and take x = C ε log ( e + p ) x=C_\varepsilon\log(e+p) x = C ε log ( e + p ) with C ε C_\varepsilon C ε sufficiently large. The preceding exceptional probabilities are at most ε + o ( 1 ) \varepsilon+o(1) ε + o ( 1 ) . Since H q / J ≤ V q H_q/J\le V_q H q / J ≤ V q , combining all contributions proves the asserted O p O_p O p bound, including when p p p is fixed. □
Corollary S.2.10. Under Assumptions 2.1, 2.2, 3.1 and 3.2, apply Lemma S.2.9 with p = p n p=p_n p = p n , M 1 = M 1 , n M_1=M_{1,n} M 1 = M 1 , n , and κ i k = β X ( ∥ S i − S k ∥ 2 ) 1 / 2 \kappa_{ik}=\beta_X(\|S_i-S_k\|_2)^{1/2} κ ik = β X ( ∥ S i − S k ∥ 2 ) 1/2 for i ≠ k i\ne k i = k , κ i i = 1 \kappa_{ii}=1 κ ii = 1 . Let D n = δ n − d D_n=\delta_n^{-d} D n = δ n − d , N ∗ = max i N i N_*=\max_iN_i N ∗ = max i N i , Λ n = w ∗ N ∗ \Lambda_n=w_*N_* Λ n = w ∗ N ∗ . Use b n b_n b n from Section 3.1 with this value of p n p_n p n . Then Lemma S.2.9 gives
max q ≤ p n ∣ S q ∣ = O p ( max q V q log ( e + p n ) + D n Λ n b n d log ( e + p n ) ) . \max_{q\le p_n}|S_q|=O_p\left(\max_q\sqrt{V_q\log(e+p_n)}+D_n\Lambda_nb_n^{d}\log(e+p_n)\right). q ≤ p n max ∣ S q ∣ = O p ( q max V q log ( e + p n ) + D n Λ n b n d log ( e + p n ) ) .
Proof. Rio’s covariance inequality [35 Theorem 1.1] and Assumption 3.2 give the required covariance bound. Lemma S.2.1 and α β > 2 d \alpha_{\beta}>2d α β > 2 d give Θ n ≤ C D n \Theta_n\le CD_n Θ n ≤ C D n . In the stated spatial regimes, D n ≥ c > 0 D_n\ge c>0 D n ≥ c > 0 . Since E q ≤ w ∗ 2 N ∗ E_q\le w_*^2N_* E q ≤ w ∗ 2 N ∗ and J ≥ 1 J\ge1 J ≥ 1 ,
max q Θ n E q / J ≤ C w ∗ D n N ∗ / J ≤ C D n Λ n . \max_q\sqrt{\Theta_nE_q/J}\le Cw_*\sqrt{D_nN_*/J}\le CD_n\Lambda_n. q max Θ n E q / J ≤ C w ∗ D n N ∗ / J ≤ C D n Λ n .
Also, M 1 , n ≤ m n N ∗ = Θ n m e f f , n N ∗ ≤ C D n N ∗ m e f f , n M_{1,n}\le m_nN_*=\Theta_nm_{\mathrm{eff},n}N_*\le CD_nN_*m_{\mathrm{eff},n} M 1 , n ≤ m n N ∗ = Θ n m eff , n N ∗ ≤ C D n N ∗ m eff , n . Write R n = e + p n + m e f f , n R_n=e+p_n+m_{\mathrm{eff},n} R n = e + p n + m eff , n and ϵ 0 = 2 d / α β > 0 \epsilon_0=2d/\alpha_\beta>0 ϵ 0 = 2 d / α β > 0 , so b n d = R n ϵ 0 b_n^d=R_n^{\epsilon_0} b n d = R n ϵ 0 . Using log R ≤ C ϵ 0 R ϵ 0 \log R\le C_{\epsilon_0}R^{\epsilon_0} log R ≤ C ϵ 0 R ϵ 0 ,
log ( e + M 1 , n + p n ) ≤ C { log R n + log ( 1 + D n N ∗ ) } ≤ C D n N ∗ b n d . \log(e+M_{1,n}+p_n)\le C\{\log R_n+\log(1+D_nN_*)\}\le CD_nN_*b_n^d. log ( e + M 1 , n + p n ) ≤ C { log R n + log ( 1 + D n N ∗ )} ≤ C D n N ∗ b n d .
Because D n N ∗ ≥ c D_nN_*\ge c D n N ∗ ≥ c and b n ≥ 1 b_n\ge1 b n ≥ 1 , it follows that
w ∗ { log ( e + M 1 , n + p n ) } 1 / 2 ≤ C D n Λ n b n d . w_*\{\log(e+M_{1,n}+p_n)\}^{1/2}\le CD_n\Lambda_nb_n^d. w ∗ { log ( e + M 1 , n + p n ) } 1/2 ≤ C D n Λ n b n d .
Substitution in Lemma S.2.9 proves the result.
Lemma S.2.11. Under Assumptions 2.1 and 3.4,
sup t ∈ [ 0 , 1 ] ∣ m ( t ) − m J m ( t ) ∣ ≤ C J m − q m ∗ . \sup_{t\in[0,1]}|m(t)-m_{J_m}(t)|\le CJ_m^{-q_m^*}. t ∈ [ 0 , 1 ] sup ∣ m ( t ) − m J m ( t ) ∣ ≤ C J m − q m ∗ . Proof. The spline approximation theorem for quasi-uniform meshes [12 Chapter XII] gives a vector β ~ m , J m ∈ R K m \widetilde{\boldsymbol{\beta}}_{m,J_m}\in\mathbb{R}^{K_m} β m , J m ∈ R K m such that m ~ J m ( t ) : = B m ( t ) ⊤ β ~ m , J m \widetilde m_{J_m}(t):=\boldsymbol{B}_m(t)^\top\widetilde{\boldsymbol{\beta}}_{m,J_m} m J m ( t ) := B m ( t ) ⊤ β m , J m satisfies
sup t ∈ [ 0 , 1 ] ∣ m ( t ) − m ~ J m ( t ) ∣ ≤ C J m − q m ∗ . (30) \sup_{t\in[0,1]}|m(t)-\widetilde m_{J_m}(t)|\le CJ_m^{-q_m^*}. \tag*{(30)} t ∈ [ 0 , 1 ] sup ∣ m ( t ) − m J m ( t ) ∣ ≤ C J m − q m ∗ . (30) Since β m , J m = Q m , J m − 1 r m , J m \boldsymbol{\beta}_{m,J_m}=Q_{m,J_m}^{-1}r_{m,J_m} β m , J m = Q m , J m − 1 r m , J m ,
β m , J m − β ~ m , J m = Q m , J m − 1 E [ B m ( T ) { m ( T ) − m ~ J m ( T ) } ] . \boldsymbol{\beta}_{m,J_m}-\widetilde{\boldsymbol{\beta}}_{m,J_m}
=Q_{m,J_m}^{-1}\mathbb{E}\left[\boldsymbol{B}_m(T)\{m(T)-\widetilde m_{J_m}(T)\}\right]. β m , J m − β m , J m = Q m , J m − 1 E [ B m ( T ) { m ( T ) − m J m ( T )} ] . Thus, for each t t t ,
∣ m J m ( t ) − m ~ J m ( t ) ∣ = ∣ B m ( t ) ⊤ Q m , J m − 1 E [ B m ( T ) { m ( T ) − m ~ J m ( T ) } ] ∣ ≤ ∥ B m ( t ) ∥ 1 ∥ Q m , J m − 1 ∥ ∞ ∥ E [ B m ( T ) { m ( T ) − m ~ J m ( T ) } ] ∥ ∞ . \begin{aligned}
|m_{J_m}(t)-\widetilde m_{J_m}(t)|
&=\left|\boldsymbol{B}_m(t)^\top Q_{m,J_m}^{-1}\mathbb{E}\left[\boldsymbol{B}_m(T)\{m(T)-\widetilde m_{J_m}(T)\}\right]\right|\\
&\le\|\boldsymbol{B}_m(t)\|_1\|Q_{m,J_m}^{-1}\|_\infty
\left\|\mathbb{E}\left[\boldsymbol{B}_m(T)\{m(T)-\widetilde m_{J_m}(T)\}\right]\right\|_\infty.
\end{aligned} ∣ m J m ( t ) − m J m ( t ) ∣ = B m ( t ) ⊤ Q m , J m − 1 E [ B m ( T ) { m ( T ) − m J m ( T )} ] ≤ ∥ B m ( t ) ∥ 1 ∥ Q m , J m − 1 ∥ ∞ ∥ E [ B m ( T ) { m ( T ) − m J m ( T )} ] ∥ ∞ . By Lemmas S.2.5 and S.2.6, ∥ B m ( t ) ∥ 1 ≤ 1 \|\boldsymbol{B}_m(t)\|_1\le1 ∥ B m ( t ) ∥ 1 ≤ 1 and ∥ Q m , J m − 1 ∥ ∞ ≤ C J m \|Q_{m,J_m}^{-1}\|_\infty\le CJ_m ∥ Q m , J m − 1 ∥ ∞ ≤ C J m . For each coordinate a a a ,
∣ E [ B a , m ( T ) { m ( T ) − m ~ J m ( T ) } ] ∣ ≤ sup u ∣ m ( u ) − m ~ J m ( u ) ∣ E B a , m ( T ) . \left|\mathbb{E}\left[B_{a,m}(T)\{m(T)-\widetilde m_{J_m}(T)\}\right]\right|
\le\sup_u|m(u)-\widetilde m_{J_m}(u)|\mathbb{E}B_{a,m}(T). ∣ E [ B a , m ( T ) { m ( T ) − m J m ( T )} ] ∣ ≤ u sup ∣ m ( u ) − m J m ( u ) ∣ E B a , m ( T ) . Because B a , m B_{a,m} B a , m is supported on an interval of length O ( J m − 1 ) O(J_m^{-1}) O ( J m − 1 ) and f T f_T f T is bounded, E B a , m ( T ) ≤ C J m − 1 \mathbb{E}B_{a,m}(T)\le CJ_m^{-1} E B a , m ( T ) ≤ C J m − 1 . Therefore,
sup t ∈ [ 0 , 1 ] ∣ m J m ( t ) − m ~ J m ( t ) ∣ ≤ C J m ⋅ J m − 1 ⋅ J m − q m ∗ ≤ C J m − q m ∗ . \sup_{t\in[0,1]}|m_{J_m}(t)-\widetilde m_{J_m}(t)|
\le CJ_m\cdot J_m^{-1}\cdot J_m^{-q_m^*}\le CJ_m^{-q_m^*}. t ∈ [ 0 , 1 ] sup ∣ m J m ( t ) − m J m ( t ) ∣ ≤ C J m ⋅ J m − 1 ⋅ J m − q m ∗ ≤ C J m − q m ∗ . Combining this bound with (S.2.5) proves the lemma.
Lemma S.2.12. Under Assumption 2.1,
∥ Q ^ m , J m − Q m , J m ∥ max = O p ( log K m M 1 , n J m + log K m M 1 , n ) . \left\|\widehat Q_{m,J_m}-Q_{m,J_m}\right\|_{\max}
=O_p\left(\sqrt{\frac{\log K_m}{M_{1,n}J_m}}+\frac{\log K_m}{M_{1,n}}\right). Q m , J m − Q m , J m m a x = O p ( M 1 , n J m log K m + M 1 , n log K m ) .
Consequently, if
J m log K m M 1 , n J m + J m log K m M 1 , n → 0 , (31) J_m\sqrt{\frac{\log K_m}{M_{1,n}J_m}}+J_m\frac{\log K_m}{M_{1,n}}\to0, \tag*{(31)} J m M 1 , n J m log K m + J m M 1 , n log K m → 0 , (31)
then ∥ Q ^ m , J m − 1 ∥ ∞ = O p ( J m ) \lVert\widehat{Q}^{-1}_{m,J_m}\rVert_\infty=O_p(J_m) ∥ Q m , J m − 1 ∥ ∞ = O p ( J m ) , and
∥ Q ^ m , J m − 1 − Q m , J m − 1 ∥ ∞ = O p ( J m 2 log K m M 1 , n J m + J m 2 log K m M 1 , n ) . \left\lVert\widehat{Q}^{-1}_{m,J_m}-Q^{-1}_{m,J_m}\right\rVert_\infty
=O_p\left(J_m^2\sqrt{\frac{\log K_m}{M_{1,n}J_m}}+J_m^2\frac{\log K_m}{M_{1,n}}\right). Q m , J m − 1 − Q m , J m − 1 ∞ = O p ( J m 2 M 1 , n J m log K m + J m 2 M 1 , n log K m ) .
Proof. Fix a coordinate pair ( a , b ) (a,b) ( a , b ) .
( Q ^ m , J m − Q m , J m ) a b = 1 M 1 , n ∑ i = 1 m n ∑ j = 1 N i [ B a , m ( T i j ) B b , m ( T i j ) − E { B a , m ( T ) B b , m ( T ) } ] . \left(\widehat{Q}_{m,J_m}-Q_{m,J_m}\right)_{ab}
=\frac{1}{M_{1,n}}\sum_{i=1}^{m_n}\sum_{j=1}^{N_i}\left[B_{a,m}(T_{ij})B_{b,m}(T_{ij})-\mathbb{E}\left\{B_{a,m}(T)B_{b,m}(T)\right\}\right]. ( Q m , J m − Q m , J m ) ab = M 1 , n 1 i = 1 ∑ m n j = 1 ∑ N i [ B a , m ( T ij ) B b , m ( T ij ) − E { B a , m ( T ) B b , m ( T ) } ] .
The summands are independent over all ( i , j ) (i,j) ( i , j ) because the time points are i.i.d., and each is bounded in absolute value by 1. If ∣ a − b ∣ > p − 1 |a-b|>p-1 ∣ a − b ∣ > p − 1 , local support makes the product identically zero. Otherwise,
Var { B a , m ( T ) B b , m ( T ) } ≤ E { B a , m ( T ) 2 B b , m ( T ) 2 } ≤ C J m − 1 , \operatorname{Var}\left\{B_{a,m}(T)B_{b,m}(T)\right\}\leq\mathbb{E}\left\{B_{a,m}(T)^2B_{b,m}(T)^2\right\}\leq CJ_m^{-1}, Var { B a , m ( T ) B b , m ( T ) } ≤ E { B a , m ( T ) 2 B b , m ( T ) 2 } ≤ C J m − 1 ,
because the overlap support has length O ( J m − 1 ) O(J_m^{-1}) O ( J m − 1 ) . Thus, Bernstein’s inequality [4 Corollary 2.11] gives, for every x > 0 x>0 x > 0 ,
P ( ∣ ( Q ^ m , J m − Q m , J m ) a b ∣ > x ) ≤ 2 exp { − C M 1 , n x 2 J m − 1 + x } . \mathbb{P}\left(\left|\left(\widehat{Q}_{m,J_m}-Q_{m,J_m}\right)_{ab}\right|>x\right)
\leq2\exp\left\{-\frac{CM_{1,n}x^2}{J_m^{-1}+x}\right\}. P ( ( Q m , J m − Q m , J m ) ab > x ) ≤ 2 exp { − J m − 1 + x C M 1 , n x 2 } .
There are only O ( K m ) O(K_m) O ( K m ) potentially nonzero entries, so take
x = A ( log K m M 1 , n J m + log K m M 1 , n ) x=A\left(\sqrt{\frac{\log K_m}{M_{1,n}J_m}}+\frac{\log K_m}{M_{1,n}}\right) x = A ( M 1 , n J m log K m + M 1 , n log K m )
and apply a union bound to obtain
P ( ∥ Q ^ m , J m − Q m , J m ∥ max > x ) ≤ C K m e − c min ( A 2 , A ) log K m . \mathbb{P}\left(\left\lVert\widehat{Q}_{m,J_m}-Q_{m,J_m}\right\rVert_{\max}>x\right)
\leq CK_m e^{-c\min(A^2,A)\log K_m}. P ( Q m , J m − Q m , J m m a x > x ) ≤ C K m e − c m i n ( A 2 , A ) l o g K m .
For each ε > 0 \varepsilon>0 ε > 0 , the last bound is at most ε \varepsilon ε for sufficiently large A A A , uniformly for K m ≥ 2 K_m\geq2 K m ≥ 2 . This proves the max-norm bound. Because the Gram matrix has fixed bandwidth, its ℓ ∞ \ell_\infty ℓ ∞ operator norm is bounded by a constant multiple of the max norm; hence ∥ Q ^ m , J m − Q m , J m ∥ ∞ = o p ( J m − 1 ) \lVert\widehat{Q}_{m,J_m}-Q_{m,J_m}\rVert_\infty=o_p(J_m^{-1}) ∥ Q m , J m − Q m , J m ∥ ∞ = o p ( J m − 1 ) under (S.2.6). The Neumann-series inversion identity then gives
Q ^ m , J m − 1 = Q m , J m − 1 { I + ( Q ^ m , J m − Q m , J m ) Q m , J m − 1 } − 1 , \widehat{Q}^{-1}_{m,J_m}=Q^{-1}_{m,J_m}\left\{I+\left(\widehat{Q}_{m,J_m}-Q_{m,J_m}\right)Q^{-1}_{m,J_m}\right\}^{-1}, Q m , J m − 1 = Q m , J m − 1 { I + ( Q m , J m − Q m , J m ) Q m , J m − 1 } − 1 ,
which implies ∥ Q ^ m , J m − 1 ∥ ∞ = O p ( J m ) \lVert\widehat{Q}^{-1}_{m,J_m}\rVert_\infty=O_p(J_m) ∥ Q m , J m − 1 ∥ ∞ = O p ( J m ) . Finally,
Q ^ m , J m − 1 − Q m , J m − 1 = Q m , J m − 1 ( Q m , J m − Q ^ m , J m ) Q ^ m , J m − 1 , \widehat{Q}^{-1}_{m,J_m}-Q^{-1}_{m,J_m}
=Q^{-1}_{m,J_m}\left(Q_{m,J_m}-\widehat{Q}_{m,J_m}\right)\widehat{Q}^{-1}_{m,J_m}, Q m , J m − 1 − Q m , J m − 1 = Q m , J m − 1 ( Q m , J m − Q m , J m ) Q m , J m − 1 ,
and the preceding bounds yield the last display.
Lemma S.2.13. Under Assumption 2.1 , E ( ξ i , J m ) = 0 \mathbb{E}(\xi_{i,J_m})=0 E ( ξ i , J m ) = 0 for 1 ≤ i ≤ m n 1\le i\le m_n 1 ≤ i ≤ m n .
Proof. The exogeneity in Assumption 2.1, the pointwise centering of X X X , and the centering and independence of the measurement errors imply E ( Y i j ∣ T i j ) = m ( T i j ) \mathbb{E}(Y_{ij}\mid T_{ij})=m(T_{ij}) E ( Y ij ∣ T ij ) = m ( T ij ) . Consequently,
E ( ξ i , J m ) = N i E [ B m ( T ) { m ( T ) − m J m ( T ) } ] = 0 \mathbb{E}(\xi_{i,J_m})=N_i\mathbb{E}\left[B_m(T)\{m(T)-m_{J_m}(T)\}\right]=0 E ( ξ i , J m ) = N i E [ B m ( T ) { m ( T ) − m J m ( T )} ] = 0 by equation (3) .
Lemma S.2.14. Under Assumptions 2.1, 2.2, 3.1, 3.2 and 3.4, for every 1 ≤ a ≤ K m 1\le a\le K_m 1 ≤ a ≤ K m ,
Var ( ξ i , J m , a ) ≤ C { N i J m + N i ( N i − 1 ) J m 2 } . \operatorname{Var}(\xi_{i,J_m,a})\le C\left\{\frac{N_i}{J_m}+\frac{N_i(N_i-1)}{J_m^2}\right\}. Var ( ξ i , J m , a ) ≤ C { J m N i + J m 2 N i ( N i − 1 ) } . Define
v m , n : = 1 M 1 , n J m + M 2 , n + ( Θ n − 1 ) ( M 1 , n + M 2 , n ) M 1 , n 2 J m 2 . (32) v_{m,n}:=\frac{1}{M_{1,n}J_m}+\frac{M_{2,n}+(\Theta_n-1)(M_{1,n}+M_{2,n})}{M_{1,n}^2J_m^2}. \tag*{(32)} v m , n := M 1 , n J m 1 + M 1 , n 2 J m 2 M 2 , n + ( Θ n − 1 ) ( M 1 , n + M 2 , n ) . (32) Then
Var ( 1 M 1 , n ∑ i = 1 m n ξ i , J m , a ) ≤ C v m , n . \operatorname{Var}\left(\frac{1}{M_{1,n}}\sum_{i=1}^{m_n}\xi_{i,J_m,a}\right)\le Cv_{m,n}. Var ( M 1 , n 1 i = 1 ∑ m n ξ i , J m , a ) ≤ C v m , n . Proof. Fix a a a , write B a ( t ) = B a , m ( t ) B_a(t)=B_{a,m}(t) B a ( t ) = B a , m ( t ) , and set R i j = Y i j − m J m ( T i j ) R_{ij}=Y_{ij}-m_{J_m}(T_{ij}) R ij = Y ij − m J m ( T ij ) . By Lemma S.2.13, the score is centered. Exogeneity and the uniform moment bounds give E { B a ( T i j ) 2 R i j 2 } ≤ C J m − 1 \mathbb{E}\{B_a(T_{ij})^2R_{ij}^2\}\le CJ_m^{-1} E { B a ( T ij ) 2 R ij 2 } ≤ C J m − 1 . For j ≠ ℓ j\ne\ell j = ℓ , independence of the two observation times and bounded conditional second moments give
∣ E { B a ( T i j ) B a ( T i ℓ ) R i j R i ℓ } ∣ ≤ C { E B a ( T ) } 2 ≤ C J m − 2 . \left|\mathbb{E}\{B_a(T_{ij})B_a(T_{i\ell})R_{ij}R_{i\ell}\}\right|\le C\{\mathbb{E}B_a(T)\}^2\le CJ_m^{-2}. ∣ E { B a ( T ij ) B a ( T i ℓ ) R ij R i ℓ } ∣ ≤ C { E B a ( T ) } 2 ≤ C J m − 2 . This proves the same-site variance bound. For i ≠ k i\ne k i = k , projection orthogonality and independence of the observation times and errors from the latent process remove all cross-site terms except the latent-field product. Rio’s covariance inequality [35 Theorem 1.1] and Assumption 3.2 yield
∣ Cov ( ξ i , J m , a , ξ k , J m , a ) ∣ ≤ C N i N k J m − 2 β X ( ∥ S i − S k ∥ 2 ) 1 / 2 . \left|\operatorname{Cov}(\xi_{i,J_m,a},\xi_{k,J_m,a})\right|\le CN_iN_kJ_m^{-2}\beta_X(\|S_i-S_k\|_2)^{1/2}. ∣ Cov ( ξ i , J m , a , ξ k , J m , a ) ∣ ≤ C N i N k J m − 2 β X ( ∥ S i − S k ∥ 2 ) 1/2 . Using 2 N i N k ≤ N i 2 + N k 2 2N_iN_k\le N_i^2+N_k^2 2 N i N k ≤ N i 2 + N k 2 ,
∑ i ≠ k N i N k β X ( ∥ S i − S k ∥ 2 ) 1 / 2 ≤ ( Θ n − 1 ) ∑ i N i 2 = ( Θ n − 1 ) ( M 1 , n + M 2 , n ) . \sum_{i\ne k}N_iN_k\beta_X(\|S_i-S_k\|_2)^{1/2}\le(\Theta_n-1)\sum_iN_i^2=(\Theta_n-1)(M_{1,n}+M_{2,n}). i = k ∑ N i N k β X ( ∥ S i − S k ∥ 2 ) 1/2 ≤ ( Θ n − 1 ) i ∑ N i 2 = ( Θ n − 1 ) ( M 1 , n + M 2 , n ) . Adding the diagonal and off-diagonal same-site terms gives
C M 1 , n 2 { M 1 , n J m + M 2 , n + ( Θ n − 1 ) ( M 1 , n + M 2 , n ) J m 2 } = C v m , n , \frac{C}{M_{1,n}^2}\left\{\frac{M_{1,n}}{J_m}+\frac{M_{2,n}+(\Theta_n-1)(M_{1,n}+M_{2,n})}{J_m^2}\right\}=Cv_{m,n}, M 1 , n 2 C { J m M 1 , n + J m 2 M 2 , n + ( Θ n − 1 ) ( M 1 , n + M 2 , n ) } = C v m , n , for the full normalized variance.
Lemma S.2.15. Under Assumptions 2.1, 2.2, 3.1, 3.2, 3.4 and 3.5, take p n = K m p_n=K_m p n = K m in the definition of b n b_n b n . Then
∥ 1 M 1 , n ∑ i = 1 m n ξ i , J m ∥ ∞ = O p ( v m , n log K m + δ n − d b n d log K m m n ) . (33) \left\|\frac{1}{M_{1,n}}\sum_{i=1}^{m_n}\xi_{i,J_m}\right\|_\infty
=O_p\left(\sqrt{v_{m,n}\log K_m}+\frac{\delta_n^{-d}b_n^d\log K_m}{m_n}\right). \tag*{(33)} M 1 , n 1 i = 1 ∑ m n ξ i , J m ∞ = O p ( v m , n log K m + m n δ n − d b n d log K m ) . (33)
where v m , n v_{m,n} v m , n is defined in (S.2.7). If δ n − d b n d J m { log K m / [ m n ( Θ n + J m / N n ) ] } 1 / 2 → 0 \delta_n^{-d}b_n^dJ_m\left\{\log K_m/\left[m_n\left(\Theta_n+J_m/N_n\right)\right]\right\}^{1/2}\to0 δ n − d b n d J m { log K m / [ m n ( Θ n + J m / N n ) ] } 1/2 → 0 , then Assumption 3.5 further gives
J m ∥ 1 M 1 , n ∑ i = 1 m n ξ ˉ i , J m ∥ ∞ = O p [ log K m { J m m n N n + 1 m e f f , n } ] . (34) J_m\left\|\frac{1}{M_{1,n}}\sum_{i=1}^{m_n}\bar{\xi}_{i,J_m}\right\|_\infty=O_p\left[\sqrt{\log K_m\left\{\frac{J_m}{m_nN_n}+\frac{1}{m_{\mathrm{eff},n}}\right\}}\right]. \tag*{(34)} J m M 1 , n 1 i = 1 ∑ m n ξ ˉ i , J m ∞ = O p [ log K m { m n N n J m + m eff , n 1 } ] . (34)
Proof. In Corollary S.2.10, take J = J m J=J_m J = J m , w i , a = M 1 , n − 1 w_{i,a}=M_{1,n}^{-1} w i , a = M 1 , n − 1 , B a = B a , m B_a=B_{a,m} B a = B a , m and d i , a = m − m J m d_{i,a}=m-m_{J_m} d i , a = m − m J m . The score is centered by Lemma S.2.13, and the residual is uniformly bounded by Lemma S.2.11. Here H a = M 1 , n − 1 H_a=M_{1,n}^{-1} H a = M 1 , n − 1 , F a = ( M 1 , n + M 2 , n ) / M 1 , n 2 F_a=(M_{1,n}+M_{2,n})/M_{1,n}^2 F a = ( M 1 , n + M 2 , n ) / M 1 , n 2 , and V a = v m , n V_a=v_{m,n} V a = v m , n . By Assumption 3.5, w ∗ max i N i ≤ C / m n w_*\max_i N_i\le C/m_n w ∗ max i N i ≤ C / m n , proving (S.2.8). Moreover, Assumption 3.5 implies M 1 , n + M 2 , n = ∑ i N i 2 ≤ C M 1 , n 2 / m n M_{1,n}+M_{2,n}=\sum_iN_i^2\le CM_{1,n}^2/m_n M 1 , n + M 2 , n = ∑ i N i 2 ≤ C M 1 , n 2 / m n . Applying this reduction to (S.2.7), multiplying (S.2.8) by J m J_m J m , and using the additional condition in the lemma prove (S.2.9).
For an interior point u u u , use the weights W i ( u ) W_i(u) W i ( u ) and local observation mass M 1 , n ( u ) M_{1,n}(u) M 1 , n ( u ) from Section 2.2.1, and write
m ˉ n ( u , t ) : = 1 M 1 , n ( u ) ∑ i = 1 m n W i ( u ) N i m ( S i , t ) , \bar{m}_n(u,t):=\frac{1}{M_{1,n}(u)}\sum_{i=1}^{m_n}W_i(u)N_i m(S_i,t), m ˉ n ( u , t ) := M 1 , n ( u ) 1 i = 1 ∑ m n W i ( u ) N i m ( S i , t ) ,
β m , J m l o c ( u ) : = Q m , J m − 1 E { B m ( T ) m ˉ n ( u , T ) } , m J m l o c ( u , t ) : = B m ( t ) T β m , J m l o c ( u ) , \beta^{\mathrm{loc}}_{m,J_m}(u):=Q_{m,J_m}^{-1}\mathbb{E}\left\{B_m(T)\bar{m}_n(u,T)\right\},\qquad m^{\mathrm{loc}}_{J_m}(u,t):=B_m(t)^{\mathsf{T}}\beta^{\mathrm{loc}}_{m,J_m}(u), β m , J m loc ( u ) := Q m , J m − 1 E { B m ( T ) m ˉ n ( u , T ) } , m J m loc ( u , t ) := B m ( t ) T β m , J m loc ( u ) ,
and let
ξ i , J m ( u ) : = W i ( u ) ∑ j = 1 N i B m ( T i j ) { Y i j − m J m l o c ( u , T i j ) } , ξ ˉ i , J m ( u ) : = ξ i , J m ( u ) − E { ξ i , J m ( u ) } . \xi_{i,J_m}(u):=W_i(u)\sum_{j=1}^{N_i}B_m(T_{ij})\left\{Y_{ij}-m^{\mathrm{loc}}_{J_m}(u,T_{ij})\right\},\qquad\bar{\xi}_{i,J_m}(u):=\xi_{i,J_m}(u)-\mathbb{E}\left\{\xi_{i,J_m}(u)\right\}. ξ i , J m ( u ) := W i ( u ) j = 1 ∑ N i B m ( T ij ) { Y ij − m J m loc ( u , T ij ) } , ξ ˉ i , J m ( u ) := ξ i , J m ( u ) − E { ξ i , J m ( u ) } .
These quantities are used only in the deterministic and stochastic decompositions below.
Lemma S.2.16. Under Assumptions 2.1 and 3.9,
sup t ∈ [ 0 , 1 ] ∣ m ˉ n ( u 0 , t ) − M ( u 0 , t ) ∣ ≤ C ( a n γ u + L n − 1 ) . (35) \sup_{t\in[0,1]}\left|\bar{m}_n(u_0,t)-M(u_0,t)\right|\le C\left(a_n^{\gamma_u}+L_n^{-1}\right). \tag*{(35)} t ∈ [ 0 , 1 ] sup ∣ m ˉ n ( u 0 , t ) − M ( u 0 , t ) ∣ ≤ C ( a n γ u + L n − 1 ) . (35) sup t ∈ [ 0 , 1 ] ∣ m J m l o c ( u 0 , t ) − m ˉ n ( u 0 , t ) ∣ ≤ C J m − q m ∗ . (36) \sup_{t\in[0,1]}\left|m^{\mathrm{loc}}_{J_m}(u_0,t)-\bar{m}_n(u_0,t)\right|\le CJ_m^{-q_m^*}. \tag*{(36)} t ∈ [ 0 , 1 ] sup m J m loc ( u 0 , t ) − m ˉ n ( u 0 , t ) ≤ C J m − q m ∗ . (36) and therefore
sup t ∈ [ 0 , 1 ] ∣ m J m l o c ( u 0 , t ) − M ( u 0 , t ) ∣ ≤ C ( J m − q m ∗ + a n γ u + L n − 1 ) . (37) \sup_{t\in[0,1]}\left|m^{\mathrm{loc}}_{J_m}(u_0,t)-M(u_0,t)\right|\le C\left(J_m^{-q_m^*}+a_n^{\gamma_u}+L_n^{-1}\right). \tag*{(37)} t ∈ [ 0 , 1 ] sup m J m loc ( u 0 , t ) − M ( u 0 , t ) ≤ C ( J m − q m ∗ + a n γ u + L n − 1 ) . (37)
Proof. To prove (S.2.10), write
∣ m ˉ n ( u 0 , t ) − M ( u 0 , t ) ∣ ≤ 1 M 1 , n ( u 0 ) ∑ i = 1 m n W i ( u 0 ) N i ∣ m ( S i , t ) − M ( U i , t ) ∣ + 1 M 1 , n ( u 0 ) ∑ i = 1 m n W i ( u 0 ) N i ∣ M ( U i , t ) − M ( u 0 , t ) ∣ . \begin{aligned}
\left|\bar{m}_n(u_0,t)-M(u_0,t)\right|
&\le\frac{1}{M_{1,n}(u_0)}\sum_{i=1}^{m_n}W_i(u_0)N_i\left|m(S_i,t)-M(U_i,t)\right|\\
&\quad+\frac{1}{M_{1,n}(u_0)}\sum_{i=1}^{m_n}W_i(u_0)N_i\left|M(U_i,t)-M(u_0,t)\right|.
\end{aligned} ∣ m ˉ n ( u 0 , t ) − M ( u 0 , t ) ∣ ≤ M 1 , n ( u 0 ) 1 i = 1 ∑ m n W i ( u 0 ) N i ∣ m ( S i , t ) − M ( U i , t ) ∣ + M 1 , n ( u 0 ) 1 i = 1 ∑ m n W i ( u 0 ) N i ∣ M ( U i , t ) − M ( u 0 , t ) ∣ .
The first term is bounded by C L n − 1 CL_n^{-1} C L n − 1 by Assumption 3.9(iii). For the second term, W i ( u 0 ) ≠ 0 W_i(u_0)\ne0 W i ( u 0 ) = 0 implies ∥ U i − u 0 ∥ 2 ≤ C a n \lVert U_i-u_0\rVert_2\le Ca_n ∥ U i − u 0 ∥ 2 ≤ C a n , and Assumption 3.9(i) bounds it by C a n γ u Ca_n^{\gamma_u} C a n γ u . Thus, (S.2.10) follows.
For (S.2.11), Assumption 3.9(ii),(iv) imply that m ˉ n ( u 0 , ⋅ ) \bar{m}_n(u_0,\cdot) m ˉ n ( u 0 , ⋅ ) is uniformly bounded in C q m , ν m [ 0 , 1 ] C^{q_m,\nu_m}[0,1] C q m , ν m [ 0 , 1 ] , with time-spline order p ≥ q m ∗ + 1 p \ge q_m^*+1 p ≥ q m ∗ + 1 . The deterministic spline approximation theorem (again de Boor [12 ] ) gives a spline m ~ J m \widetilde{m}_{J_m} m J m such that sup t ∈ [ 0 , 1 ] ∣ m ˉ n ( u 0 , t ) − m ~ J m ( t ) ∣ ≤ C J m − q m ∗ \sup_{t\in[0,1]}|\bar{m}_n(u_0,t)-\widetilde{m}_{J_m}(t)| \le CJ_m^{-q_m^*} sup t ∈ [ 0 , 1 ] ∣ m ˉ n ( u 0 , t ) − m J m ( t ) ∣ ≤ C J m − q m ∗ . Projection onto the spline space under the population Gram inner product preserves this order by the argument used for Lemma S.2.11; therefore, (S.2.11) follows. Finally, (S.2.12) follows by combining (S.2.10) and (S.2.11).
Lemma S.2.17. Under Assumption 2.1,
∥ Q ^ m , J m ( u 0 ) − Q m , J m ∥ max = O p ( log K m M 1 , n ( u 0 ) J m + log K m M 1 , n ( u 0 ) ) . \left\|\widehat{Q}_{m,J_m}(u_0)-Q_{m,J_m}\right\|_{\max}
=O_p\left(\sqrt{\frac{\log K_m}{M_{1,n}(u_0)J_m}}+\frac{\log K_m}{M_{1,n}(u_0)}\right). Q m , J m ( u 0 ) − Q m , J m m a x = O p ( M 1 , n ( u 0 ) J m log K m + M 1 , n ( u 0 ) log K m ) . If
J m ( log K m M 1 , n ( u 0 ) J m + log K m M 1 , n ( u 0 ) ) → 0 , (38) J_m\left(\sqrt{\frac{\log K_m}{M_{1,n}(u_0)J_m}}+\frac{\log K_m}{M_{1,n}(u_0)}\right)\to0, \tag*{(38)} J m ( M 1 , n ( u 0 ) J m log K m + M 1 , n ( u 0 ) log K m ) → 0 , (38) then
∥ Q ^ m , J m ( u 0 ) − 1 ∥ ∞ = O p ( J m ) , \left\|\widehat{Q}_{m,J_m}(u_0)^{-1}\right\|_{\infty}=O_p(J_m), Q m , J m ( u 0 ) − 1 ∞ = O p ( J m ) , ∥ Q ^ m , J m ( u 0 ) − 1 − Q m , J m − 1 ∥ ∞ = O p ( J m 2 log K m M 1 , n ( u 0 ) J m + J m 2 log K m M 1 , n ( u 0 ) ) . \left\|\widehat{Q}_{m,J_m}(u_0)^{-1}-Q_{m,J_m}^{-1}\right\|_{\infty}
=O_p\left(J_m^2\sqrt{\frac{\log K_m}{M_{1,n}(u_0)J_m}}+J_m^2\frac{\log K_m}{M_{1,n}(u_0)}\right). Q m , J m ( u 0 ) − 1 − Q m , J m − 1 ∞ = O p ( J m 2 M 1 , n ( u 0 ) J m log K m + J m 2 M 1 , n ( u 0 ) log K m ) . Proof. Fix coordinates ( a , b ) (a,b) ( a , b ) . Then
{ Q ^ m , J m ( u 0 ) − Q m , J m } a b = 1 M 1 , n ( u 0 ) ∑ i = 1 m n ∑ j = 1 N i W i ( u 0 ) [ B a , m ( T i j ) B b , m ( T i j ) − E { B a , m ( T ) B b , m ( T ) } ] . \left\{\widehat{Q}_{m,J_m}(u_0)-Q_{m,J_m}\right\}_{ab}
=\frac{1}{M_{1,n}(u_0)}\sum_{i=1}^{m_n}\sum_{j=1}^{N_i}W_i(u_0)
\left[B_{a,m}(T_{ij})B_{b,m}(T_{ij})-\mathbb{E}\{B_{a,m}(T)B_{b,m}(T)\}\right]. { Q m , J m ( u 0 ) − Q m , J m } ab = M 1 , n ( u 0 ) 1 i = 1 ∑ m n j = 1 ∑ N i W i ( u 0 ) [ B a , m ( T ij ) B b , m ( T ij ) − E { B a , m ( T ) B b , m ( T )} ] . The summands are independent over ( i , j ) (i,j) ( i , j ) , uniformly bounded, and have variance bounded by C W i ( u 0 ) 2 J m − 1 CW_i(u_0)^2J_m^{-1} C W i ( u 0 ) 2 J m − 1 when the supports overlap and by 0 otherwise. Since ∑ i W i ( u 0 ) 2 N i ≤ C M 1 , n ( u 0 ) \sum_i W_i(u_0)^2N_i\le CM_{1,n}(u_0) ∑ i W i ( u 0 ) 2 N i ≤ C M 1 , n ( u 0 ) , Bernstein’s inequality (Boucheron et al., [4 ] ) and a union bound over the O ( K m ) O(K_m) O ( K m ) nonzero entries yield the displayed max-norm rate. The inverse bound follows from the same Neumann-series argument as in Lemma S.2.12.
Lemma S.2.18. Assume Assumptions 2.1, 2.2, 3.1 to 3.3, 3.9 and 3.10. Define
A 1 , n ( u 0 ) : = ∑ i W i ( u 0 ) 2 N i , A_{1,n}(u_0):=\sum_i W_i(u_0)^2N_i, A 1 , n ( u 0 ) := i ∑ W i ( u 0 ) 2 N i , A 2 , n ( u 0 ) : = ∑ i W i ( u 0 ) 2 N i ( N i − 1 ) , A_{2,n}(u_0):=\sum_i W_i(u_0)^2N_i(N_i-1), A 2 , n ( u 0 ) := i ∑ W i ( u 0 ) 2 N i ( N i − 1 ) , A 0 , n ( u 0 ) : = ∑ i W i ( u 0 ) 2 N i 2 = A 1 , n ( u 0 ) + A 2 , n ( u 0 ) , A_{0,n}(u_0):=\sum_i W_i(u_0)^2N_i^2=A_{1,n}(u_0)+A_{2,n}(u_0), A 0 , n ( u 0 ) := i ∑ W i ( u 0 ) 2 N i 2 = A 1 , n ( u 0 ) + A 2 , n ( u 0 ) ,
and
v l o c , n ( u 0 ) : = 1 M 1 , n ( u 0 ) 2 { A 1 , n ( u 0 ) J m + A 2 , n ( u 0 ) + ( Θ n − 1 ) A 0 , n ( u 0 ) J m 2 } . (39) v_{\mathrm{loc},n}(u_0) := \frac{1}{M_{1,n}(u_0)^2}\left\{\frac{A_{1,n}(u_0)}{J_m}+\frac{A_{2,n}(u_0)+(\Theta_n-1)A_{0,n}(u_0)}{J_m^2}\right\}. \tag*{(39)} v loc , n ( u 0 ) := M 1 , n ( u 0 ) 2 1 { J m A 1 , n ( u 0 ) + J m 2 A 2 , n ( u 0 ) + ( Θ n − 1 ) A 0 , n ( u 0 ) } . (39)
For every coordinate a a a ,
Var { 1 M 1 , n ( u 0 ) ∑ i ξ ˉ i , J m , a ( u 0 ) } ≤ C v l o c , n ( u 0 ) , \operatorname{Var}\left\{\frac{1}{M_{1,n}(u_0)}\sum_i \bar{\xi}_{i,J_m,a}(u_0)\right\}\le C v_{\mathrm{loc},n}(u_0), Var { M 1 , n ( u 0 ) 1 i ∑ ξ ˉ i , J m , a ( u 0 ) } ≤ C v loc , n ( u 0 ) ,
Moreover,
v l o c , n ( u 0 ) ≤ C { 1 m n a n d N ˉ n J m + 1 m e f f , n a n d J m 2 } . (40) v_{\mathrm{loc},n}(u_0)\le C\left\{\frac{1}{m_n a_n^d\bar{N}_nJ_m}+\frac{1}{m_{\mathrm{eff},n}a_n^dJ_m^2}\right\}. \tag*{(40)} v loc , n ( u 0 ) ≤ C { m n a n d N ˉ n J m 1 + m eff , n a n d J m 2 1 } . (40)
With p n = K m p_n=K_m p n = K m in the definition of b n b_n b n ,
∥ 1 M 1 , n ( u 0 ) ∑ i ξ ˉ i , J m ( u 0 ) ∥ ∞ = O p [ log K m { 1 m n a n d N ˉ n J m + 1 m e f f , n a n d J m 2 } + δ n − d b n d log K m m n a n d ] . (41) \left\|\frac{1}{M_{1,n}(u_0)}\sum_i\bar{\xi}_{i,J_m}(u_0)\right\|_{\infty}
=O_p\left[\sqrt{\log K_m\left\{\frac{1}{m_na_n^d\bar{N}_nJ_m}+\frac{1}{m_{\mathrm{eff},n}a_n^dJ_m^2}\right\}+\frac{\delta_n^{-d}b_n^d\log K_m}{m_na_n^d}}\right]. \tag*{(41)} M 1 , n ( u 0 ) 1 i ∑ ξ ˉ i , J m ( u 0 ) ∞ = O p [ log K m { m n a n d N ˉ n J m 1 + m eff , n a n d J m 2 1 } + m n a n d δ n − d b n d log K m ] . (41)
If δ n − d b n d J m { a n − d log K m / [ m n ( Θ n + J m / N ˉ n ) ] } 1 / 2 → 0 \delta_n^{-d}b_n^dJ_m\{a_n^{-d}\log K_m/[m_n(\Theta_n+J_m/\bar{N}_n)]\}^{1/2}\to0 δ n − d b n d J m { a n − d log K m / [ m n ( Θ n + J m / N ˉ n )] } 1/2 → 0 , then
J m ∥ 1 M 1 , n ( u 0 ) ∑ i ξ ˉ i , J m ( u 0 ) ∥ ∞ = O p [ log K m { J m m n a n d N ˉ n + 1 m e f f , n a n d } ] . (42) J_m\left\|\frac{1}{M_{1,n}(u_0)}\sum_i\bar{\xi}_{i,J_m}(u_0)\right\|_{\infty}
=O_p\left[\sqrt{\log K_m\left\{\frac{J_m}{m_na_n^d\bar{N}_n}+\frac{1}{m_{\mathrm{eff},n}a_n^d}\right\}}\right]. \tag*{(42)} J m M 1 , n ( u 0 ) 1 i ∑ ξ ˉ i , J m ( u 0 ) ∞ = O p [ log K m { m n a n d N ˉ n J m + m eff , n a n d 1 } ] . (42)
Proof. Fix a a a . The deterministic residual m ( S i , t ) − m J m l o c ( u 0 , t ) m(S_i,t)-m^{\mathrm{loc}}_{J_m}(u_0,t) m ( S i , t ) − m J m loc ( u 0 , t ) is uniformly bounded by Assumption 3.9(iv) and Lemma S.2.16. The same-site diagonal/off-diagonal calculation gives
Var { ξ ˉ i , J m , a ( u 0 ) } ≤ C W i ( u 0 ) 2 { N i J m + N i ( N i − 1 ) J m 2 } . \operatorname{Var}\{\bar{\xi}_{i,J_m,a}(u_0)\}\le CW_i(u_0)^2\left\{\frac{N_i}{J_m}+\frac{N_i(N_i-1)}{J_m^2}\right\}. Var { ξ ˉ i , J m , a ( u 0 )} ≤ C W i ( u 0 ) 2 { J m N i + J m 2 N i ( N i − 1 ) } .
For i ≠ k i\ne k i = k , centering does not change covariance, and only the latent-field product remains. Thus
∣ Cov { ξ ˉ i , J m , a ( u 0 ) , ξ ˉ k , J m , a ( u 0 ) } ∣ ≤ C W i ( u 0 ) W k ( u 0 ) N i N k J m − 2 β X ( ∥ S i − S k ∥ 2 ) 1 / 2 . \left|\operatorname{Cov}\{\bar{\xi}_{i,J_m,a}(u_0),\bar{\xi}_{k,J_m,a}(u_0)\}\right|
\le CW_i(u_0)W_k(u_0)N_iN_kJ_m^{-2}\beta_X(\lVert S_i-S_k\rVert_2)^{1/2}. Cov { ξ ˉ i , J m , a ( u 0 ) , ξ ˉ k , J m , a ( u 0 )} ≤ C W i ( u 0 ) W k ( u 0 ) N i N k J m − 2 β X (∥ S i − S k ∥ 2 ) 1/2 .
Symmetrization gives
∑ i ≠ k W i W k N i N k β X ( ∥ S i − S k ∥ 2 ) 1 / 2 ≤ ( Θ n − 1 ) ∑ i W i 2 N i 2 . \sum_{i\ne k}W_iW_kN_iN_k\beta_X(\lVert S_i-S_k\rVert_2)^{1/2}
\le(\Theta_n-1)\sum_iW_i^2N_i^2. i = k ∑ W i W k N i N k β X (∥ S i − S k ∥ 2 ) 1/2 ≤ ( Θ n − 1 ) i ∑ W i 2 N i 2 .
This proves the variance bound with (S.2.14).
Balanced counts and Assumption 3.10 imply
M 1 , n ( u 0 ) ≍ m n a n d N ˉ n , A 1 , n ( u 0 ) ≤ C m n a n d N ˉ n , A 0 , n ( u 0 ) ≤ C m n a n d N ˉ n 2 . M_{1,n}(u_0)\asymp m_na_n^d\bar{N}_n,\qquad A_{1,n}(u_0)\le Cm_na_n^d\bar{N}_n,\qquad A_{0,n}(u_0)\le Cm_na_n^d\bar{N}_n^2. M 1 , n ( u 0 ) ≍ m n a n d N ˉ n , A 1 , n ( u 0 ) ≤ C m n a n d N ˉ n , A 0 , n ( u 0 ) ≤ C m n a n d N ˉ n 2 .
Since A 2 , n ≤ A 0 , n A_{2,n}\le A_{0,n} A 2 , n ≤ A 0 , n and m n = Θ n m e f f , n m_n=\Theta_nm_{\mathrm{eff},n} m n = Θ n m eff , n , these bounds give (S.2.15).
For the maximal inequality, apply Corollary S.2.10 with J = J m J=J_m J = J m and
w i , a = W i ( u 0 ) M 1 , n ( u 0 ) , B a = B a , m , d i , a ( t ) = m ( S i , t ) − m J m l o c ( u 0 , t ) . w_{i,a}=\frac{W_i(u_0)}{M_{1,n}(u_0)},\qquad B_a=B_{a,m},\qquad d_{i,a}(t)=m(S_i,t)-m^{\mathrm{loc}}_{J_m}(u_0,t). w i , a = M 1 , n ( u 0 ) W i ( u 0 ) , B a = B a , m , d i , a ( t ) = m ( S i , t ) − m J m loc ( u 0 , t ) .
Here H a = A 1 , n ( u 0 ) / M 1 , n ( u 0 ) 2 H_a=A_{1,n}(u_0)/M_{1,n}(u_0)^2 H a = A 1 , n ( u 0 ) / M 1 , n ( u 0 ) 2 , F a = A 0 , n ( u 0 ) / M 1 , n ( u 0 ) 2 F_a=A_{0,n}(u_0)/M_{1,n}(u_0)^2 F a = A 0 , n ( u 0 ) / M 1 , n ( u 0 ) 2 , and V a = v l o c , n ( u 0 ) V_a=v_{\mathrm{loc},n}(u_0) V a = v loc , n ( u 0 ) . Balanced counts and local mass give w ∗ max i N i ≤ C ( m n a n d ) − 1 w_*\max_iN_i\le C(m_na_n^d)^{-1} w ∗ max i N i ≤ C ( m n a n d ) − 1 . Substitution gives (S.2.16). Multiplication by J m J_m J m and the additional condition in the lemma give (S.2.17).
Lemma S.2.19. Under Assumptions 2.1, 3.9, 3.14 and 3.15 and Assumption 3.3, use Q M , J s , J m Q_{M,J_s,J_m} Q M , J s , J m , r M , J s , J m r_{M,J_s,J_m} r M , J s , J m , θ M , J s , J m ∘ \theta^\circ_{M,J_s,J_m} θ M , J s , J m ∘ , and M J s , J m ∘ M^\circ_{J_s,J_m} M J s , J m ∘ from Section 2.2.2. Then
sup u ∈ [ 0 , 1 ] d sup t ∈ [ 0 , 1 ] ∣ M J s , J m ∘ ( u , t ) − M ( u , t ) ∣ ≤ C { J s − q u ∗ + J m − q m ∗ + L n − 1 } . (43) \sup_{u\in[0,1]^d}\sup_{t\in[0,1]}\left|M^\circ_{J_s,J_m}(u,t)-M(u,t)\right|\leq C\left\{J_s^{-q_u^*}+J_m^{-q_m^*}+L_n^{-1}\right\}. \tag*{(43)} u ∈ [ 0 , 1 ] d sup t ∈ [ 0 , 1 ] sup M J s , J m ∘ ( u , t ) − M ( u , t ) ≤ C { J s − q u ∗ + J m − q m ∗ + L n − 1 } . (43) Furthermore, ∥ Q M , J s , J m − 1 ∥ ∞ ≤ C J s d J m \left\|Q^{-1}_{M,J_s,J_m}\right\|_\infty\leq CJ_s^dJ_m Q M , J s , J m − 1 ∞ ≤ C J s d J m , and the population normal equation is
∑ i N i ∫ 0 1 Z J s , J m ( U i , x ) { m ( S i , x ) − M J s , J m ∘ ( U i , x ) } f T ( x ) d x = 0. (44) \sum_i N_i\int_0^1 Z_{J_s,J_m}(U_i,x)\left\{m(S_i,x)-M^\circ_{J_s,J_m}(U_i,x)\right\}f_T(x)\,dx=0. \tag*{(44)} i ∑ N i ∫ 0 1 Z J s , J m ( U i , x ) { m ( S i , x ) − M J s , J m ∘ ( U i , x ) } f T ( x ) d x = 0. (44)
Proof. Tensor-product spline approximation ([36 Chapter 12] ) gives a spline M ~ J s , J m ( u , t ) = Z J s , J m ( u , t ) ⊤ θ ~ \widetilde M_{J_s,J_m}(u,t)=Z_{J_s,J_m}(u,t)^\top\widetilde\theta M J s , J m ( u , t ) = Z J s , J m ( u , t ) ⊤ θ satisfying
sup u , t ∣ M ~ J s , J m ( u , t ) − M ( u , t ) ∣ ≤ C { J s − q u ∗ + J m − q m ∗ } . \sup_{u,t}\left|\widetilde M_{J_s,J_m}(u,t)-M(u,t)\right|\leq C\left\{J_s^{-q_u^*}+J_m^{-q_m^*}\right\}. u , t sup M J s , J m ( u , t ) − M ( u , t ) ≤ C { J s − q u ∗ + J m − q m ∗ } .
Together with Assumption 3.9(iii),
max i sup t ∣ m ( S i , t ) − M ~ J s , J m ( U i , t ) ∣ ≤ C R n , R n : = J s − q u ∗ + J m − q m ∗ + L n − 1 . \max_i\sup_t\left|m(S_i,t)-\widetilde M_{J_s,J_m}(U_i,t)\right|\leq CR_n,\qquad R_n:=J_s^{-q_u^*}+J_m^{-q_m^*}+L_n^{-1}. i max t sup m ( S i , t ) − M J s , J m ( U i , t ) ≤ C R n , R n := J s − q u ∗ + J m − q m ∗ + L n − 1 .
The normal equations imply
θ M , J s , J m ∘ − θ ~ = Q M , J s , J m − 1 1 M 1 , n ∑ i N i ∫ Z J s , J m ( U i , x ) { m ( S i , x ) − M ~ J s , J m ( U i , x ) } f T ( x ) d x . \theta^\circ_{M,J_s,J_m}-\widetilde\theta=Q^{-1}_{M,J_s,J_m}\frac{1}{M_{1,n}}\sum_iN_i\int Z_{J_s,J_m}(U_i,x)\left\{m(S_i,x)-\widetilde M_{J_s,J_m}(U_i,x)\right\}f_T(x)\,dx. θ M , J s , J m ∘ − θ = Q M , J s , J m − 1 M 1 , n 1 i ∑ N i ∫ Z J s , J m ( U i , x ) { m ( S i , x ) − M J s , J m ( U i , x ) } f T ( x ) d x .
By Lemma S.2.7, M 1 , n − 1 ∑ i N i A a ( U i ) ≤ C J s − d M_{1,n}^{-1}\sum_iN_iA_a(U_i)\leq CJ_s^{-d} M 1 , n − 1 ∑ i N i A a ( U i ) ≤ C J s − d , while E B b , m ( T ) ≤ C J m − 1 \operatorname{E}B_{b,m}(T)\leq CJ_m^{-1} E B b , m ( T ) ≤ C J m − 1 . Hence the moment vector on the right is bounded coordinatewise by C J s − d J m − 1 R n CJ_s^{-d}J_m^{-1}R_n C J s − d J m − 1 R n .
Let
Q s , J s , n : = 1 M 1 , n ∑ i N i A s ( U i ) A s ( U i ) ⊤ . Q_{s,J_s,n}:=\frac{1}{M_{1,n}}\sum_iN_iA_s(U_i)A_s(U_i)^\top. Q s , J s , n := M 1 , n 1 i ∑ N i A s ( U i ) A s ( U i ) ⊤ .
Equation (S.2.3) gives C J s − d ∥ v ∥ 2 2 ≤ v ⊤ Q s , J s , n v ≤ C J s − d ∥ v ∥ 2 2 CJ_s^{-d}\lVert v\rVert_2^2\leq v^\top Q_{s,J_s,n}v\leq CJ_s^{-d}\lVert v\rVert_2^2 C J s − d ∥ v ∥ 2 2 ≤ v ⊤ Q s , J s , n v ≤ C J s − d ∥ v ∥ 2 2 . Set H s , n = J s d Q s , J s , n H_{s,n}=J_s^dQ_{s,J_s,n} H s , n = J s d Q s , J s , n . Its spectrum lies in a fixed interval [ c , C ] ⊂ ( 0 , ∞ ) [c,C]\subset(0,\infty) [ c , C ] ⊂ ( 0 , ∞ ) . Let G s , n G_{s,n} G s , n be its sparsity graph. Local support of fixed-order tensor-product B-splines implies that every graph edge changes each tensor index by at most a fixed number; consequently, every graph ball of radius r r r contains at most C ( 1 + r ) d C(1+r)^d C ( 1 + r ) d vertices.
We use the polynomial approximation argument for sparse matrix functions in Benzi and Razouk [1 ] . Because x ↦ x − 1 x\mapsto x^{-1} x ↦ x − 1 is analytic on [ c , C ] [c,C] [ c , C ] , there are constants C 0 < ∞ C_0<\infty C 0 < ∞ , ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) , and polynomials P k P_k P k of degree at most k k k such that
sup x ∈ [ c , C ] ∣ x − 1 − P k ( x ) ∣ ≤ C 0 ρ k . \sup_{x\in[c,C]}\left|x^{-1}-P_k(x)\right|\leq C_0\rho^k. x ∈ [ c , C ] sup x − 1 − P k ( x ) ≤ C 0 ρ k .
If the graph distance between basis indices a a a and a ′ a' a ′ is r > k r>k r > k , then { P k ( H s , n ) } a a ′ = 0 \{P_k(H_{s,n})\}_{aa'}=0 { P k ( H s , n ) } a a ′ = 0 , because every power of degree at most k k k connects only vertices at graph distance at most k k k . Taking k = r − 1 k=r-1 k = r − 1 and applying the spectral calculus gives ∣ ( H s , n − 1 ) a a ′ ∣ ≤ C 0 ρ r − 1 \left|(H_{s,n}^{-1})_{aa'}\right|\leq C_0\rho^{r-1} ( H s , n − 1 ) a a ′ ≤ C 0 ρ r − 1 , r ≥ 1 r\geq1 r ≥ 1 . The diagonal entries are bounded by ∥ H s , n − 1 ∥ 2 ≤ c − 1 \lVert H_{s,n}^{-1}\rVert_2\leq c^{-1} ∥ H s , n − 1 ∥ 2 ≤ c − 1 , and entries between distinct connected components are zero. Summing over graph-distance shells therefore yields
∥ H s , n − 1 ∥ ∞ ≤ C ∑ r = 0 ∞ ( 1 + r ) d ρ r ≤ C , ∥ Q s , J s , n − 1 ∥ ∞ ≤ C J s d . \left\|H_{s,n}^{-1}\right\|_\infty\leq C\sum_{r=0}^{\infty}(1+r)^d\rho^r\leq C,\qquad\left\|Q_{s,J_s,n}^{-1}\right\|_\infty\leq CJ_s^d. H s , n − 1 ∞ ≤ C r = 0 ∑ ∞ ( 1 + r ) d ρ r ≤ C , Q s , J s , n − 1 ∞ ≤ C J s d .
Independence of T T T from the site design gives Q M , J s , J m = Q s , J s , n ⊗ Q m , J m Q_{M,J_s,J_m}=Q_{s,J_s,n}\otimes Q_{m,J_m} Q M , J s , J m = Q s , J s , n ⊗ Q m , J m . Thus, by Lemma S.2.6 ,
∥ Q M , J s , J m − 1 ∥ ∞ = ∥ Q s , J s , n − 1 ∥ ∞ ∥ Q m , J m − 1 ∥ ∞ ≤ C J s d J m . \left\|Q^{-1}_{M,J_s,J_m}\right\|_\infty=\left\|Q^{-1}_{s,J_s,n}\right\|_\infty\left\|Q^{-1}_{m,J_m}\right\|_\infty\le C J_s^d J_m. Q M , J s , J m − 1 ∞ = Q s , J s , n − 1 ∞ Q m , J m − 1 ∞ ≤ C J s d J m .
Combining this inverse bound with the coordinatewise moment bound proves (S.2.18). Equation (S.2.19) is the defining normal equation.
Lemma S.2.20. Assume Assumptions 2.1, 2.2, 3.1, 3.2, 3.9, 3.14 and 3.15 and Assumption 3.3. Let
p M , n : = K s K m , r M , Q , n : = J s − d log p M , n M 1 , n J m + log p M , n M 1 , n . p_{M,n}:=K_sK_m,\qquad r_{M,Q,n}:=\sqrt{\frac{J_s^{-d}\log p_{M,n}}{M_{1,n}J_m}+\frac{\log p_{M,n}}{M_{1,n}}}. p M , n := K s K m , r M , Q , n := M 1 , n J m J s − d log p M , n + M 1 , n log p M , n . and take p n = p M , n p_n=p_{M,n} p n = p M , n in the definition of b n b_n b n . Then ∥ Q ^ M , J s , J m − Q M , J s , J m ∥ max = O p ( r M , Q , n ) \left\|\widehat{Q}_{M,J_s,J_m}-Q_{M,J_s,J_m}\right\|_{\max}=O_p(r_{M,Q,n}) Q M , J s , J m − Q M , J s , J m m a x = O p ( r M , Q , n ) . If J s d J m r M , Q , n → 0 J_s^dJ_mr_{M,Q,n}\to0 J s d J m r M , Q , n → 0 , then
∥ Q ^ M , J s , J m − 1 ∥ ∞ = O p ( J s d J m ) , ∥ Q ^ M , J s , J m − 1 − Q M , J s , J m − 1 ∥ ∞ = O p { ( J s d J m ) 2 r M , Q , n } = o p ( J s d J m ) . \left\|\widehat{Q}^{-1}_{M,J_s,J_m}\right\|_\infty=O_p(J_s^dJ_m),\qquad\left\|\widehat{Q}^{-1}_{M,J_s,J_m}-Q^{-1}_{M,J_s,J_m}\right\|_\infty=O_p\left\{(J_s^dJ_m)^2r_{M,Q,n}\right\}=o_p(J_s^dJ_m). Q M , J s , J m − 1 ∞ = O p ( J s d J m ) , Q M , J s , J m − 1 − Q M , J s , J m − 1 ∞ = O p { ( J s d J m ) 2 r M , Q , n } = o p ( J s d J m ) . Use the centered score ω i , J s , J m \omega_{i,J_s,J_m} ω i , J s , J m from § 2.2.2 . For q = ( a , b ) q=(a,b) q = ( a , b ) , set
v s p , n , q : = 1 M 1 , n 2 [ 1 J m ∑ i A a ( U i ) 2 N i + 1 J m 2 { ∑ i A a ( U i ) 2 N i ( N i − 1 ) + ( Θ n − 1 ) ∑ i A a ( U i ) 2 N i 2 } ] . (45) v_{\mathrm{sp},n,q}:=\frac{1}{M_{1,n}^2}\left[\frac{1}{J_m}\sum_i A_a(U_i)^2N_i+\frac{1}{J_m^2}\left\{\sum_i A_a(U_i)^2N_i(N_i-1)+(\Theta_n-1)\sum_i A_a(U_i)^2N_i^2\right\}\right]. \tag*{(45)} v sp , n , q := M 1 , n 2 1 [ J m 1 i ∑ A a ( U i ) 2 N i + J m 2 1 { i ∑ A a ( U i ) 2 N i ( N i − 1 ) + ( Θ n − 1 ) i ∑ A a ( U i ) 2 N i 2 } ] . (45) The full normalized score variance is bounded by C v s p , n , q Cv_{\mathrm{sp},n,q} C v sp , n , q , and
v s p , n , q ≤ C J s − d { 1 m n N ‾ n J m + 1 m e f f , n J m 2 } . (46) v_{\mathrm{sp},n,q}\le CJ_s^{-d}\left\{\frac{1}{m_n\overline{N}_nJ_m}+\frac{1}{m_{\mathrm{eff},n}J_m^2}\right\}. \tag*{(46)} v sp , n , q ≤ C J s − d { m n N n J m 1 + m eff , n J m 2 1 } . (46) Moreover,
∥ 1 M 1 , n ∑ i ω i , J s , J m ∥ ∞ = O p [ J s − d log p M , n { 1 m n N ‾ n J m + 1 m e f f , n J m 2 } + δ n − d b n d log p M , n m n ] . (47) \left\|\frac{1}{M_{1,n}}\sum_i\omega_{i,J_s,J_m}\right\|_\infty
=O_p\left[\sqrt{J_s^{-d}\log p_{M,n}\left\{\frac{1}{m_n\overline{N}_nJ_m}+\frac{1}{m_{\mathrm{eff},n}J_m^2}\right\}}+\frac{\delta_n^{-d}b_n^d\log p_{M,n}}{m_n}\right]. \tag*{(47)} M 1 , n 1 i ∑ ω i , J s , J m ∞ = O p [ J s − d log p M , n { m n N n J m 1 + m eff , n J m 2 1 } + m n δ n − d b n d log p M , n ] . (47) If δ n − d b n d J m { J s d log ( K s K m ) / [ m n ( Θ n + J m / N ‾ n ) ] } 1 / 2 → 0 \delta_n^{-d}b_n^dJ_m\left\{J_s^d\log(K_sK_m)/[m_n(\Theta_n+J_m/\overline{N}_n)]\right\}^{1/2}\to0 δ n − d b n d J m { J s d log ( K s K m ) / [ m n ( Θ n + J m / N n )] } 1/2 → 0 , then
J s d J m ∥ 1 M 1 , n ∑ i ω i , J s , J m ∥ ∞ = O p [ J s d log p M , n { J m m n N ‾ n + 1 m e f f , n } ] . (48) J_s^dJ_m\left\|\frac{1}{M_{1,n}}\sum_i\omega_{i,J_s,J_m}\right\|_\infty
=O_p\left[\sqrt{J_s^d\log p_{M,n}\left\{\frac{J_m}{m_n\overline{N}_n}+\frac{1}{m_{\mathrm{eff},n}}\right\}}\right]. \tag*{(48)} J s d J m M 1 , n 1 i ∑ ω i , J s , J m ∞ = O p [ J s d log p M , n { m n N n J m + m eff , n 1 } ] . (48)
Proof. For a nonzero Gram coordinate, the summand is bounded and its time-factor variance is O ( J m − 1 ) O(J_m^{-1}) O ( J m − 1 ) . By (S.2.4), the weighted sum of the spatial basis values is O ( J s − d ) O(J_s^{-d}) O ( J s − d ) . Hence a normalized Gram coordinate has variance at most C J s − d / ( M 1 , n J m ) CJ_s^{-d}/(M_{1,n}J_m) C J s − d / ( M 1 , n J m ) . Bernstein’s inequality [4 Corollary 2.11] and a union bound over the O ( p M , n ) O(p_{M,n}) O ( p M , n ) nonzero entries give the stated max-norm rate. The fixed number of nonzero entries in each row,
Lemma S.2.19 , the Neumann-series identity, and the resolvent identity give the inverse bounds.
For one score coordinate q = ( a , b ) q=(a,b) q = ( a , b ) , the deterministic residual is uniformly bounded by Lemma S.2.19 . The same-site calculation gives
Var ( ω i , q ) ≤ C A a ( U i ) 2 { N i J m + N i ( N i − 1 ) J m 2 } . \operatorname{Var}(\omega_{i,q}) \le C A_a(U_i)^2 \left\{ \frac{N_i}{J_m} + \frac{N_i(N_i-1)}{J_m^2} \right\}. Var ( ω i , q ) ≤ C A a ( U i ) 2 { J m N i + J m 2 N i ( N i − 1 ) } .
For i ≠ k i \ne k i = k , only the latent-field cross product remains. Rio [35 ] gives
∣ Cov ( ω i , q , ω k , q ) ∣ ≤ C ∣ A a ( U i ) A a ( U k ) ∣ N i N k J m − 2 β X ( ∥ S i − S k ∥ 2 ) 1 / 2 . \left|\operatorname{Cov}(\omega_{i,q},\omega_{k,q})\right| \le C\left|A_a(U_i)A_a(U_k)\right|N_iN_kJ_m^{-2}\beta_X\left(\left\|S_i-S_k\right\|_2\right)^{1/2}. ∣ Cov ( ω i , q , ω k , q ) ∣ ≤ C ∣ A a ( U i ) A a ( U k ) ∣ N i N k J m − 2 β X ( ∥ S i − S k ∥ 2 ) 1/2 .
Symmetrization bounds the cross-site sum by ( Θ n − 1 ) ∑ i A a ( U i ) 2 N i 2 (\Theta_n-1)\sum_i A_a(U_i)^2N_i^2 ( Θ n − 1 ) ∑ i A a ( U i ) 2 N i 2 . This proves the variance bound with (S.2.20).
By (S.2.4), ∑ i A a ( U i ) 2 N i ≤ C M 1 , n J s − d \sum_i A_a(U_i)^2N_i \le CM_{1,n}J_s^{-d} ∑ i A a ( U i ) 2 N i ≤ C M 1 , n J s − d . Balanced counts then give ∑ i A a ( U i ) 2 N i 2 ≤ C N ‾ n M 1 , n J s − d \sum_i A_a(U_i)^2N_i^2 \le C\overline{N}_nM_{1,n}J_s^{-d} ∑ i A a ( U i ) 2 N i 2 ≤ C N n M 1 , n J s − d , and the same bound applies to the N i ( N i − 1 ) N_i(N_i-1) N i ( N i − 1 ) sum. Since M 1 , n ≍ m n N ‾ n M_{1,n}\asymp m_n\overline{N}_n M 1 , n ≍ m n N n and m n = Θ n m e f f , n m_n=\Theta_nm_{\mathrm{eff},n} m n = Θ n m eff , n , (S.2.21) follows.
Finally, apply Corollary S.2.10 with J = J m J=J_m J = J m and
w i , q = A a ( U i ) M 1 , n , B q = B b , m , d i , q ( t ) = m ( S i , t ) − M J s , J m ∘ ( U i , t ) . w_{i,q}=\frac{A_a(U_i)}{M_{1,n}}, \qquad B_q=B_{b,m}, \qquad d_{i,q}(t)=m(S_i,t)-M^\circ_{J_s,J_m}(U_i,t). w i , q = M 1 , n A a ( U i ) , B q = B b , m , d i , q ( t ) = m ( S i , t ) − M J s , J m ∘ ( U i , t ) .
The preceding spatial-load bounds give H q ≤ C J s − d / ( m n N ‾ n ) H_q\le CJ_s^{-d}/(m_n\overline{N}_n) H q ≤ C J s − d / ( m n N n ) , F q ≤ C J s − d / m n F_q\le CJ_s^{-d}/m_n F q ≤ C J s − d / m n , and V q = v s p , n , q V_q=v_{\mathrm{sp},n,q} V q = v sp , n , q . Bounded spatial B-splines and balanced counts give w ∗ max i N i ≤ C / m n w_*\max_iN_i\le C/m_n w ∗ max i N i ≤ C / m n . These bounds yield (S.2.22). Multiplication by J s d J m J_s^dJ_m J s d J m and the additional condition in the lemma give (S.2.23)$. □
Proofs of main theorems
Proof of Theorem 3.6. By Lemma S.2.11,
sup t ∣ m ^ J m ( t ) − m ( t ) ∣ ≤ sup t ∣ m ^ J m ( t ) − m J m ( t ) ∣ + C J m − q m ∗ . \sup_t \lvert\widehat{m}_{J_m}(t)-m(t)\rvert\le\sup_t \lvert\widehat{m}_{J_m}(t)-m_{J_m}(t)\rvert+ C J_m^{-q_m^*}. t sup ∣ m J m ( t ) − m ( t )∣ ≤ t sup ∣ m J m ( t ) − m J m ( t )∣ + C J m − q m ∗ .
The normal equations and equation (4) give
β ^ m , J m − β m , J m = Q m , J m − 1 1 M 1 , n ∑ i ξ i , J m + ( Q ^ m , J m − 1 − Q m , J m − 1 ) 1 M 1 , n ∑ i ξ i , J m . \widehat{\beta}_{m,J_m}-\beta_{m,J_m}=Q_{m,J_m}^{-1}\frac{1}{M_{1,n}}\sum_i \xi_{i,J_m}+(\widehat{Q}_{m,J_m}^{-1}-Q_{m,J_m}^{-1})\frac{1}{M_{1,n}}\sum_i \xi_{i,J_m}. β m , J m − β m , J m = Q m , J m − 1 M 1 , n 1 i ∑ ξ i , J m + ( Q m , J m − 1 − Q m , J m − 1 ) M 1 , n 1 i ∑ ξ i , J m .
Because sup t ∥ B m ( t ) ∥ 1 = 1 \sup_t\lVert B_m(t)\rVert_1=1 sup t ∥ B m ( t ) ∥ 1 = 1 and ∥ Q m , J m − 1 ∥ ∞ ≤ C J m \lVert Q_{m,J_m}^{-1}\rVert_\infty\le C J_m ∥ Q m , J m − 1 ∥ ∞ ≤ C J m , Lemma S.2.15 yields the leading stochastic bound
O p ( J m v m , n log K m + J m δ n d b d log K m m n ) . O_p\left(J_m\sqrt{v_{m,n}\log K_m}+\frac{J_m\delta_n^{d}b_d\log K_m}{m_n}\right). O p ( J m v m , n log K m + m n J m δ n d b d log K m ) .
Using (S.2.7),
J m 2 v m , n = J m M 1 , n + M 2 , n + ( Θ n − 1 ) ( M 1 , n + M 2 , n ) M 1 , n 2 , J_m^2v_{m,n}=\frac{J_m}{M_{1,n}}+\frac{M_{2,n}+(\Theta_n-1)(M_{1,n}+M_{2,n})}{M_{1,n}^2}, J m 2 v m , n = M 1 , n J m + M 1 , n 2 M 2 , n + ( Θ n − 1 ) ( M 1 , n + M 2 , n ) ,
which is the stochastic term in equation (12) . By Lemma S.2.12 and (S.2.6), ∥ Q ^ m , J m − 1 − Q m , J m − 1 ∥ ∞ = o p ( J m ) \lVert\widehat{Q}_{m,J_m}^{-1}-Q_{m,J_m}^{-1}\rVert_\infty=o_p(J_m) ∥ Q m , J m − 1 − Q m , J m − 1 ∥ ∞ = o p ( J m ) ; hence its product with the score is lower order. Adding the approximation bias proves equation (12) .
Proof of Corollary 3.7. Under the additional condition in Corollary 3.7, use the score bound in (S.2.9) and the bias bound in Lemma S.2.11. The inverse-Gram argument used for Theorem 3.6 then gives equation (13) . Under Assumption 3.3, suppress logarithmic and lower-order terms. In the sparse time regime, balance
J m − q m ∗ ≍ ( J m m n N ‾ n ) 1 / 2 , J_m^{-q_m^*}\asymp\left(\frac{J_m}{m_n\overline{N}_n}\right)^{1/2}, J m − q m ∗ ≍ ( m n N n J m ) 1/2 ,
which gives J m ≍ ( m n N ‾ n ) 1 / ( 2 q m ∗ + 1 ) J_m\asymp(m_n\overline{N}_n)^{1/(2q_m^*+1)} J m ≍ ( m n N n ) 1/ ( 2 q m ∗ + 1 ) and rate ( m n N ‾ n ) − q m ∗ / ( 2 q m ∗ + 1 ) (m_n\overline{N}_n)^{-q_m^*/(2q_m^*+1)} ( m n N n ) − q m ∗ / ( 2 q m ∗ + 1 ) . This component meets the curve-level term m e f f , n − 1 / 2 m_{\mathrm{eff},n}^{-1/2} m eff , n − 1/2 when
( m n N ‾ n ) − q m ∗ / ( 2 q m ∗ + 1 ) ≍ m e f f , n − 1 / 2 . (m_n\overline{N}_n)^{-q_m^*/(2q_m^*+1)}\asymp m_{\mathrm{eff},n}^{-1/2}. ( m n N n ) − q m ∗ / ( 2 q m ∗ + 1 ) ≍ m eff , n − 1/2 .
Since m n = Θ n m e f f , n m_n=\Theta_nm_{\mathrm{eff},n} m n = Θ n m eff , n , this is equivalent to N ‾ n ≍ m e f f , n 1 / ( 2 q m ∗ ) / Θ n \overline{N}_n\asymp m_{\mathrm{eff},n}^{1/(2q_m^*)}/\Theta_n N n ≍ m eff , n 1/ ( 2 q m ∗ ) / Θ n . Above the phase transition boundary, balancing J m − q m ∗ J_m^{-q_m^*} J m − q m ∗ with m e f f , n − 1 / 2 m_{\mathrm{eff},n}^{-1/2} m eff , n − 1/2 gives J m ≍ m e f f , n 1 / ( 2 q m ∗ ) J_m\asymp m_{\mathrm{eff},n}^{1/(2q_m^*)} J m ≍ m eff , n 1/ ( 2 q m ∗ ) and the stated dense-time rate.
Derivation of the local effective-sample-size equivalence. Under Assumptions 3.3 and 3.10,
M 1 , n ( u 0 ) ≍ N ‾ n ∑ i W i ( u 0 ) ≍ N ‾ n m n a n d ∑ i W i ( u 0 ) N i 2 ≍ N ‾ n 2 ∑ i W i ( u 0 ) 2 ≍ N ‾ n 2 m n a n d . M_{1,n}(u_0)\asymp\overline{N}_n\sum_i W_i(u_0)\asymp\overline{N}_nm_na_n^d\sum_i W_i(u_0)N_i^2\asymp\overline{N}_n^2\sum_i W_i(u_0)^2\asymp\overline{N}_n^2m_na_n^d. M 1 , n ( u 0 ) ≍ N n i ∑ W i ( u 0 ) ≍ N n m n a n d i ∑ W i ( u 0 ) N i 2 ≍ N n 2 i ∑ W i ( u 0 ) 2 ≍ N n 2 m n a n d .
Consequently, the definition gives m u 0 , e f f , n ≍ m n a n d / Θ n = m e f f , n a n d m_{u_0,\mathrm{eff},n}\asymp m_na_n^d/\Theta_n=m_{\mathrm{eff},n}a_n^d m u 0 , eff , n ≍ m n a n d / Θ n = m eff , n a n d .
Proof of Theorem 3.11 . By Lemma S.2.16 ,
sup t ∣ m J m l o c ( u 0 , t ) − M ( u 0 , t ) ∣ ≤ C { J m − q m ∗ + a n γ u + L n − 1 } . \sup_t \left|m^{\mathrm{loc}}_{J_m}(u_0,t)-M(u_0,t)\right| \le C\left\{J_m^{-q_m^*}+a_n^{\gamma_u}+L_n^{-1}\right\}. t sup m J m loc ( u 0 , t ) − M ( u 0 , t ) ≤ C { J m − q m ∗ + a n γ u + L n − 1 } .
The local normal equations give
β ^ m , J m ( u 0 ) − β m , J m l o c ( u 0 ) = Q m , J m − 1 1 M 1 , n ( u 0 ) ∑ i ξ ˉ i , J m ( u 0 ) + { Q ^ m , J m ( u 0 ) − 1 − Q m , J m − 1 } 1 M 1 , n ( u 0 ) ∑ i ξ ˉ i , J m ( u 0 ) . \begin{aligned}
\widehat{\beta}_{m,J_m}(u_0)-\beta^{\mathrm{loc}}_{m,J_m}(u_0)=Q^{-1}_{m,J_m}\frac{1}{M_{1,n}(u_0)}\sum_i\bar{\xi}_{i,J_m}(u_0) \\
&+\left\{\widehat{Q}_{m,J_m}(u_0)^{-1}-Q^{-1}_{m,J_m}\right\}\frac{1}{M_{1,n}(u_0)}\sum_i\bar{\xi}_{i,J_m}(u_0).
\end{aligned} β m , J m ( u 0 ) − β m , J m loc ( u 0 ) = Q m , J m − 1 M 1 , n ( u 0 ) 1 i ∑ ξ ˉ i , J m ( u 0 ) + { Q m , J m ( u 0 ) − 1 − Q m , J m − 1 } M 1 , n ( u 0 ) 1 i ∑ ξ ˉ i , J m ( u 0 ) .
Since sup t ∥ B m ( t ) ∥ 1 = 1 \sup_t\lVert B_m(t)\rVert_1=1 sup t ∥ B m ( t ) ∥ 1 = 1 and ∥ Q m , J m − 1 ∥ ∞ ≤ C J m \lVert Q^{-1}_{m,J_m}\rVert_\infty\le CJ_m ∥ Q m , J m − 1 ∥ ∞ ≤ C J m , (S.2.16) gives
O p [ log K m { J m m n a n d N n + 1 m e f f , n a n d } + J m δ n − d b n d log K m m n a n d ] . O_p\left[\sqrt{\log K_m\left\{\frac{J_m}{m_na_n^dN_n}+\frac{1}{m_{\mathrm{eff},n}a_n^d}\right\}+\frac{J_m\delta_n^{-d}b_n^d\log K_m}{m_na_n^d}}\right]. O p [ log K m { m n a n d N n J m + m eff , n a n d 1 } + m n a n d J m δ n − d b n d log K m ] .
By Lemma S.2.17 and (S.2.13), the inverse-Gram difference is o p ( J m ) o_p(J_m) o p ( J m ) , so its score product is lower order. Adding the deterministic bias proves equation (15) ; under the additional condition in Corollary 3.12 , (S.2.17) gives equation (16) .
Proof of Theorem 3.13 . Use ℓ n \ell_n ℓ n and U n \mathcal{U}_n U n from the theorem, and write η n = a n 2 d + 2 / J m \eta_n=a_n^{2d+2}/J_m η n = a n 2 d + 2 / J m for the mesh size. Then
∣ U n ∣ ≤ C η n − d , log { e + K m ∣ U n ∣ } ≤ C ℓ n . \lvert\mathcal{U}_n\rvert\le C\eta_n^{-d},\qquad\log\{e+K_m\lvert\mathcal{U}_n\rvert\}\le C\ell_n. ∣ U n ∣ ≤ C η n − d , log { e + K m ∣ U n ∣} ≤ C ℓ n .
The uniform local-mass and balanced-count assumptions imply inf v ∈ U n m v , e f f , n ≳ m e f f , n a n d \inf_{v\in\mathcal{U}_n}m_{v,\mathrm{eff},n}\gtrsim m_{\mathrm{eff},n}a_n^d inf v ∈ U n m v , eff , n ≳ m eff , n a n d . Apply Corollary S.2.10 to the joint coordinates q = ( v , a ) q=(v,a) q = ( v , a ) with J = J m J=J_m J = J m and
w i , q = W i ( v ) M 1 , n ( v ) , B q = B a , m , d i , q ( t ) = m ( S i , t ) − m J m l o c ( v , t ) . w_{i,q}=\frac{W_i(v)}{M_{1,n}(v)},\qquad B_q=B_{a,m},\qquad d_{i,q}(t)=m(S_i,t)-m^{\mathrm{loc}}_{J_m}(v,t). w i , q = M 1 , n ( v ) W i ( v ) , B q = B a , m , d i , q ( t ) = m ( S i , t ) − m J m loc ( v , t ) .
The uniform first- and second-order kernel-mass bounds in Theorem 3.13 , together with balanced counts, give w ∗ max i N i ≤ C ( m n a n d ) − 1 w_*\max_iN_i\le C(m_na_n^d)^{-1} w ∗ max i N i ≤ C ( m n a n d ) − 1 uniformly over the grid. The bounds for H q H_q H q , F q F_q F q and V q V_q V q in Lemma S.2.18 hold uniformly as well. Applying the corollary once with p n = K m ∣ U n ∣ p_n=K_m\lvert\mathcal{U}_n\rvert p n = K m ∣ U n ∣ gives
max v ∈ U n ∥ 1 M 1 , n ( v ) ∑ i ξ ˉ i , J m ( v ) ∥ ∞ = O p [ ℓ n { 1 m n a n d N n J m + 1 m e f f , n a n d J m 2 } + δ n − d b n d ℓ n m n a n d ] . \max_{v\in\mathcal{U}_n}\left\lVert\frac{1}{M_{1,n}(v)}\sum_i\bar{\xi}_{i,J_m}(v)\right\rVert_\infty
=O_p\left[\sqrt{\ell_n\left\{\frac{1}{m_na_n^dN_nJ_m}+\frac{1}{m_{\mathrm{eff},n}a_n^dJ_m^2}\right\}+\frac{\delta_n^{-d}b_n^d\ell_n}{m_na_n^d}}\right]. v ∈ U n max M 1 , n ( v ) 1 i ∑ ξ ˉ i , J m ( v ) ∞ = O p [ ℓ n { m n a n d N n J m 1 + m eff , n a n d J m 2 1 } + m n a n d δ n − d b n d ℓ n ] .
Bernstein’s inequality [4 Corollary 2.11] and a union bound for the empirical Gram entries over the same grid give
max v ∈ U n ∥ Q ^ m , J m ( v ) − Q m , J m ( v ) ∥ max = O p ( ℓ n { inf u ∈ U ∘ M 1 , n ( u ) } J m + ℓ n inf u ∈ U ∘ M 1 , n ( u ) ) . \max_{v\in\mathcal{U}_n}\left\lVert\widehat{Q}_{m,J_m}(v)-Q_{m,J_m}(v)\right\rVert_{\max}
=O_p\left(\sqrt{\frac{\ell_n}{\{\inf_{u\in\mathcal{U}_\circ}M_{1,n}(u)\}J_m}+\frac{\ell_n}{\inf_{u\in\mathcal{U}_\circ}M_{1,n}(u)}}\right). v ∈ U n max Q m , J m ( v ) − Q m , J m ( v ) m a x = O p ( { inf u ∈ U ∘ M 1 , n ( u )} J m ℓ n + inf u ∈ U ∘ M 1 , n ( u ) ℓ n ) .
Because these Gram matrices have fixed bandwidth, the uniform Gram condition in Theorem 3.13 and the Neumann-series argument imply $\max_{v\in\mathcal{U}_n}\lVert\widehat{Q}_{m,J_m}(v)^{-1}\rVert_\infty=O_p(J_m), with the inverse-Gram product lower order uniformly on the grid. Combining this fact, the preceding score bound, and Lemma S.2.16 proves the asserted rate on U n U_n U n .
It remains to pass from the grid to all u ∈ U ∘ u \in U_\circ u ∈ U ∘ . For each u u u , choose u ♯ ∈ U n u^\sharp\in U_n u ♯ ∈ U n with ∥ u − u ♯ ∥ 2 ≤ C η n \lVert u-u^\sharp\rVert_2 \le C\eta_n ∥ u − u ♯ ∥ 2 ≤ C η n . Uniform local mass and balanced counts give inf u ∈ U ∘ M 1 , n ( u ) ≥ c a n d M 1 , n \inf_{u\in U_\circ} M_{1,n}(u) \ge c a_n^d M_{1,n} inf u ∈ U ∘ M 1 , n ( u ) ≥ c a n d M 1 , n , while the Lipschitz property of K K K yields max i ∣ W i ( u ) − W i ( u ♯ ) ∣ ≤ C η n / a n \max_i \lvert W_i(u)-W_i(u^\sharp)\rvert\le C\eta_n/a_n max i ∣ W i ( u ) − W i ( u ♯ )∣ ≤ C η n / a n . Moreover, Assumption 3.2 and Markov’s inequality give M 1 , n − 1 ∑ i = 1 m n ∑ j = 1 N i ∣ Y i j ∣ = O p ( 1 ) M_{1,n}^{-1}\sum_{i=1}^{m_n}\sum_{j=1}^{N_i}\lvert Y_{ij}\rvert=O_p(1) M 1 , n − 1 ∑ i = 1 m n ∑ j = 1 N i ∣ Y ij ∣ = O p ( 1 ) . Applying these bounds to the unnormalized Gram and moment sums, and also to the change in their normalizing local masses, gives
sup u ∈ U ∘ ∥ Q ^ m , J m ( u ) − Q ^ m , J m ( u ♯ ) ∥ max = O p ( a n / J m ) , \sup_{u\in U_\circ}\left\lVert\widehat{Q}_{m,J_m}(u)-\widehat{Q}_{m,J_m}(u^\sharp)\right\rVert_{\max}=O_p(a_n/J_m), u ∈ U ∘ sup Q m , J m ( u ) − Q m , J m ( u ♯ ) m a x = O p ( a n / J m ) ,
sup u ∈ U ∘ ∥ r ^ m , J m ( u ) − r ^ m , J m ( u ♯ ) ∥ ∞ = O p ( a n / J m ) . \sup_{u\in U_\circ}\left\lVert\widehat{r}_{m,J_m}(u)-\widehat{r}_{m,J_m}(u^\sharp)\right\rVert_\infty=O_p(a_n/J_m). u ∈ U ∘ sup r m , J m ( u ) − r m , J m ( u ♯ ) ∞ = O p ( a n / J m ) .
Because every time-spline Gram matrix has fixed bandwidth, the first display also holds in the ℓ ∞ \ell_\infty ℓ ∞ operator norm. Indeed, the first normalization term is bounded by C η n / a n d + 1 C\eta_n/a_n^{d+1} C η n / a n d + 1 , and the largest term for the moment-vector normalization is O p ( η n / a n 2 d + 1 ) O_p(\eta_n/a_n^{2d+1}) O p ( η n / a n 2 d + 1 ) ; the definition η n = a n 2 d + 2 / J m \eta_n=a_n^{2d+2}/J_m η n = a n 2 d + 2 / J m makes both O p ( a n / J m ) O_p(a_n/J_m) O p ( a n / J m ) or smaller.
The Gram interpolation bound and the grid inverse bound imply
sup u ∈ U ∘ ∥ Q ^ m , J m ( u ) − 1 ∥ ∞ = O p ( J m ) . \sup_{u\in U_\circ}\left\lVert\widehat{Q}_{m,J_m}(u)^{-1}\right\rVert_\infty=O_p(J_m). u ∈ U ∘ sup Q m , J m ( u ) − 1 ∞ = O p ( J m ) .
Let ρ n \rho_n ρ n denote the stochastic terms on the right-hand side of equation (3.10). The grid normal equations give max v ∈ U n ∥ β ^ m , J m ( v ) ∥ ∞ = O p ( 1 + ρ n ) \max_{v\in U_n}\lVert\widehat{\beta}_{m,J_m}(v)\rVert_\infty=O_p(1+\rho_n) max v ∈ U n ∥ β m , J m ( v ) ∥ ∞ = O p ( 1 + ρ n ) . Indeed, the population local-projection coefficients are uniformly bounded because their moment vectors are O ( J m − 1 ) O(J_m^{-1}) O ( J m − 1 ) coordinatewise and ∥ Q m , J m − 1 ∥ ∞ = O ( J m ) \lVert Q_{m,J_m}^{-1}\rVert_\infty=O(J_m) ∥ Q m , J m − 1 ∥ ∞ = O ( J m ) ; the grid score bound controls the remaining coefficient error. The identity
β ^ m , J m ( u ) − β ^ m , J m ( u ♯ ) = Q ^ m , J m ( u ) − 1 [ r ^ m , J m ( u ) − r ^ m , J m ( u ♯ ) + { Q ^ m , J m ( u ♯ ) − Q ^ m , J m ( u ) } β ^ m , J m ( u ♯ ) ] \widehat{\beta}_{m,J_m}(u)-\widehat{\beta}_{m,J_m}(u^\sharp)
=\widehat{Q}_{m,J_m}(u)^{-1}\left[\widehat{r}_{m,J_m}(u)-\widehat{r}_{m,J_m}(u^\sharp)
+\left\{\widehat{Q}_{m,J_m}(u^\sharp)-\widehat{Q}_{m,J_m}(u)\right\}\widehat{\beta}_{m,J_m}(u^\sharp)\right] β m , J m ( u ) − β m , J m ( u ♯ ) = Q m , J m ( u ) − 1 [ r m , J m ( u ) − r m , J m ( u ♯ ) + { Q m , J m ( u ♯ ) − Q m , J m ( u ) } β m , J m ( u ♯ ) ]
therefore yields
sup u ∈ U ∘ sup t ∈ [ 0 , 1 ] ∣ M ^ J m ( u , t ) − M ^ J m ( u ♯ , t ) ∣ = O p { a n ( 1 + ρ n ) } . \sup_{u\in U_\circ}\sup_{t\in[0,1]}\left\lvert\widehat{M}_{J_m}(u,t)-\widehat{M}_{J_m}(u^\sharp,t)\right\rvert=O_p\{a_n(1+\rho_n)\}. u ∈ U ∘ sup t ∈ [ 0 , 1 ] sup M J m ( u , t ) − M J m ( u ♯ , t ) = O p { a n ( 1 + ρ n )} .
Since a n ≤ a n γ u a_n\le a_n^{\gamma_u} a n ≤ a n γ u and a n ρ n ≤ ρ n a_n\rho_n\le\rho_n a n ρ n ≤ ρ n , this interpolation term is absorbed by the displayed rate. Finally, sup t ∈ [ 0 , 1 ] ∣ M ( u , t ) − M ( u ♯ , t ) ∣ ≤ C η n γ u ≤ C a n γ u \sup_{t\in[0,1]}\lvert M(u,t)-M(u^\sharp,t)\rvert\le C\eta_n^{\gamma_u}\le Ca_n^{\gamma_u} sup t ∈ [ 0 , 1 ] ∣ M ( u , t ) − M ( u ♯ , t )∣ ≤ C η n γ u ≤ C a n γ u by Assumption 3.9(i). Combining the grid and interpolation bounds proves equation (3.10).
It remains to verify the inverse-Gram assertion away from the net. The grid Gram bound, fixed bandwidth, and the interpolation bound above imply
sup u ∈ U ∘ ∥ Q ^ m , J m ( u ) − Q m , J m ∥ ∞ = o p ( J m − 1 ) , \sup_{u\in U_\circ}\left\lVert\widehat{Q}_{m,J_m}(u)-Q_{m,J_m}\right\rVert_\infty=o_p\left(J_m^{-1}\right), u ∈ U ∘ sup Q m , J m ( u ) − Q m , J m ∞ = o p ( J m − 1 ) ,
because the uniform Gram condition controls the grid deviation and a n / J m = o ( J m − 1 ) a_n/J_m=o(J_m^{-1}) a n / J m = o ( J m − 1 ) . The resolvent identity and the uniform inverse bound therefore give
sup u ∈ U ∘ ∥ Q ^ m , J m − 1 ( u ) − Q m , J m − 1 ∥ ∞ = o p ( J m ) . \sup_{u\in\mathcal{U}_{\circ}}\left\|\widehat{Q}_{m,J_m}^{-1}(u)-Q_{m,J_m}^{-1}\right\|_{\infty}=o_p(J_m). u ∈ U ∘ sup Q m , J m − 1 ( u ) − Q m , J m − 1 ∞ = o p ( J m ) .
Define S n ( u ) : = r ^ m , J m ( u ) − Q ^ m , J m ( u ) β m , J m l o c ( u ) S_n(u):=\widehat{r}_{m,J_m}(u)-\widehat{Q}_{m,J_m}(u)\beta_{m,J_m}^{\mathrm{loc}}(u) S n ( u ) := r m , J m ( u ) − Q m , J m ( u ) β m , J m loc ( u ) . This is the normalized local score. For λ i ( u ) : = W i ( u ) N i / M 1 , n ( u ) \lambda_i(u):=W_i(u)N_i/M_{1,n}(u) λ i ( u ) := W i ( u ) N i / M 1 , n ( u ) , the normalized-weight calculation used above gives sup u ∑ i ∣ λ i ( u ) − λ i ( u ♯ ) ∣ ≤ C η n / a n d + 1 = O ( a n / J m ) \sup_u\sum_i|\lambda_i(u)-\lambda_i(u^\sharp)|\le C\eta_n/a_n^{d+1}=O(a_n/J_m) sup u ∑ i ∣ λ i ( u ) − λ i ( u ♯ ) ∣ ≤ C η n / a n d + 1 = O ( a n / J m ) . Since the population moment coordinates are O ( J m − 1 ) O(J_m^{-1}) O ( J m − 1 ) and ∥ Q m , J m − 1 ∥ ∞ = O ( J m ) \|Q_{m,J_m}^{-1}\|_{\infty}=O(J_m) ∥ Q m , J m − 1 ∥ ∞ = O ( J m ) , it follows that
sup u ∈ U ∘ ∥ β m , J m l o c ( u ) − β m , J m l o c ( u ♯ ) ∥ ∞ = O ( a n / J m ) . \sup_{u\in\mathcal{U}_{\circ}}\left\|\beta_{m,J_m}^{\mathrm{loc}}(u)-\beta_{m,J_m}^{\mathrm{loc}}(u^\sharp)\right\|_{\infty}=O(a_n/J_m). u ∈ U ∘ sup β m , J m loc ( u ) − β m , J m loc ( u ♯ ) ∞ = O ( a n / J m ) .
Together with those interpolation bounds and the grid score bound, this yields
sup u ∈ U ∘ ∥ S n ( u ) ∥ ∞ = O p { ( ρ n + a n ) / J m } . \sup_{u\in\mathcal{U}_{\circ}}\|S_n(u)\|_{\infty}=O_p\{(\rho_n+a_n)/J_m\}. u ∈ U ∘ sup ∥ S n ( u ) ∥ ∞ = O p {( ρ n + a n ) / J m } .
Since sup t ∥ B m ( t ) ∥ 1 = 1 \sup_t\|B_m(t)\|_1=1 sup t ∥ B m ( t ) ∥ 1 = 1 , the off-net inverse-Gram product is consequently o p ( ρ n + a n ) o_p(\rho_n+a_n) o p ( ρ n + a n ) , which is absorbed by equation (18) because a n ≤ a n γ u a_n\le a_n^{\gamma_u} a n ≤ a n γ u .
Proof of Corollary 3.12. Suppress logarithmic and lower-order terms. The observation-limited balance
J m − q m ∗ ≍ ( J m m n a n d N ‾ n ) 1 / 2 J_m^{-q_m^*}\asymp\left(\frac{J_m}{m_na_n^d\overline{N}_n}\right)^{1/2} J m − q m ∗ ≍ ( m n a n d N n J m ) 1/2
gives J m ≍ ( m n a n d N ‾ n ) 1 / ( 2 q m ∗ + 1 ) J_m\asymp(m_na_n^d\overline{N}_n)^{1/(2q_m^*+1)} J m ≍ ( m n a n d N n ) 1/ ( 2 q m ∗ + 1 ) and rate ( m n a n d N ‾ n ) − q m ∗ / ( 2 q m ∗ + 1 ) (m_na_n^d\overline{N}_n)^{-q_m^*/(2q_m^*+1)} ( m n a n d N n ) − q m ∗ / ( 2 q m ∗ + 1 ) . Equating this rate with the curve-level term ( m e f f , n a n d ) − 1 / 2 (m_{\mathrm{eff},n}a_n^d)^{-1/2} ( m eff , n a n d ) − 1/2 and using m n = Θ n m e f f , n m_n=\Theta_n m_{\mathrm{eff},n} m n = Θ n m eff , n gives the boundary in equation (17) . Above it, balancing the time-spline bias with the curve-level term gives J m ≍ ( m e f f , n a n d ) 1 / ( 2 q m ∗ ) J_m\asymp(m_{\mathrm{eff},n}a_n^d)^{1/(2q_m^*)} J m ≍ ( m eff , n a n d ) 1/ ( 2 q m ∗ ) .
Proof of Theorem 3.16. By Lemma S.2.19,
sup u , t ∣ M J s , J m ∘ ( u , t ) − M ( u , t ) ∣ ≤ C { J s − q u ∗ + J m − q m ∗ + L n − 1 } . \sup_{u,t}|M_{J_s,J_m}^{\circ}(u,t)-M(u,t)|\le C\{J_s^{-q_u^*}+J_m^{-q_m^*}+L_n^{-1}\}. u , t sup ∣ M J s , J m ∘ ( u , t ) − M ( u , t ) ∣ ≤ C { J s − q u ∗ + J m − q m ∗ + L n − 1 } .
The population normal equation and the least-squares normal equations give
θ ^ − θ ∘ = Q − 1 1 M 1 , n ∑ i ω i + ( Q ^ − 1 − Q − 1 ) 1 M 1 , n ∑ i ω i . \widehat{\theta}-\theta^\circ=Q^{-1}\frac{1}{M_{1,n}}\sum_i\omega_i+(\widehat{Q}^{-1}-Q^{-1})\frac{1}{M_{1,n}}\sum_i\omega_i. θ − θ ∘ = Q − 1 M 1 , n 1 i ∑ ω i + ( Q − 1 − Q − 1 ) M 1 , n 1 i ∑ ω i .
Since tensor-product B-splines form a partition of unity and ∥ Q − 1 ∥ ∞ ≤ C J s d J m \|Q^{-1}\|_{\infty}\le CJ_s^dJ_m ∥ Q − 1 ∥ ∞ ≤ C J s d J m , (S.2.22) gives the leading stochastic term
O p [ J s d log ( K s , K m ) { J m m n N ‾ n + 1 m e f f , n } + J s d J m δ n − d b n d log ( K s K m ) m n ] . (49) O_p\left[\sqrt{J_s^d\log(K_s,K_m)\left\{\frac{J_m}{m_n\overline{N}_n}+\frac{1}{m_{\mathrm{eff},n}}\right\}}+\frac{J_s^dJ_m\delta_n^{-d}b_n^d\log(K_sK_m)}{m_n}\right]. \tag*{(49)} O p [ J s d log ( K s , K m ) { m n N n J m + m eff , n 1 } + m n J s d J m δ n − d b n d log ( K s K m ) ] . (49)
The inverse-difference bound in Lemma S.2.20 is o p ( J s d J m ) o_p(J_s^dJ_m) o p ( J s d J m ) under the displayed Gram condition, so its score product is lower order. Adding the deterministic approximation proves equation (19) . Under the additional condition in Corollary 3.17, {(S.2.23)} gives the same bound without the final remainder term. Boundary knots on the full cube avoid any interior restriction.
Proof of Corollary 3.17. Fix J s J_s J s and suppress logarithmic and lower-order terms. In the observation-limited regime, balance
J m − q m ∗ ≲ ( J s d J m m n N ‾ n ) 1 / 2 . J_m^{-q_m^*} \lesssim\left(\frac{J_s^dJ_m}{m_n\overline{N}_n}\right)^{1/2}. J m − q m ∗ ≲ ( m n N n J s d J m ) 1/2 .
which gives J m ≲ ( m n N ‾ n / J s d ) 1 / ( 2 q m ∗ + 1 ) J_m \lesssim(m_n\overline{N}_n/J_s^d)^{1/(2q_m^*+1)} J m ≲ ( m n N n / J s d ) 1/ ( 2 q m ∗ + 1 ) and the stated rate. Equating it with the curve-level term ( J s d / m e f f , n ) 1 / 2 (J_s^d/m_{\mathrm{eff},n})^{1/2} ( J s d / m eff , n ) 1/2 yields N ‾ n ≲ ( m e f f , n / J s d ) 1 / ( 2 q m ∗ ) / Θ n \overline{N}_n \lesssim(m_{\mathrm{eff},n}/J_s^d)^{1/(2q_m^*)}/\Theta_n N n ≲ ( m eff , n / J s d ) 1/ ( 2 q m ∗ ) / Θ n . Above the boundary, balancing the time bias with the curve-level term gives J m ≲ ( m e f f , n / J s d ) 1 / ( 2 q m ∗ ) J_m \lesssim(m_{\mathrm{eff},n}/J_s^d)^{1/(2q_m^*)} J m ≲ ( m eff , n / J s d ) 1/ ( 2 q m ∗ ) .
For joint sparse-time minimization, balance
J s − q u ∗ , J m − q m ∗ , ( J s d J m m n N ‾ n ) 1 / 2 . J_s^{-q_u^*}, \qquad J_m^{-q_m^*}, \qquad\left(\frac{J_s^dJ_m}{m_n\overline{N}_n}\right)^{1/2}. J s − q u ∗ , J m − q m ∗ , ( m n N n J s d J m ) 1/2 .
Writing the common rate as R R R gives J s ≲ R − 1 / q u ∗ J_s \lesssim R^{-1/q_u^*} J s ≲ R − 1/ q u ∗ and J m ≲ R − 1 / q m ∗ J_m \lesssim R^{-1/q_m^*} J m ≲ R − 1/ q m ∗ , hence R 2 + d / q u ∗ + 1 / q m ∗ ≲ ( m n N ‾ n ) − 1 R^{2+d/q_u^*+1/q_m^*} \lesssim(m_n\overline{N}_n)^{-1} R 2 + d / q u ∗ + 1/ q m ∗ ≲ ( m n N n ) − 1 , which is the stated anisotropic exponent. In the dense-time regime, balancing J s − q u ∗ J_s^{-q_u^*} J s − q u ∗ with ( J s d / m e f f , n ) 1 / 2 (J_s^d/m_{\mathrm{eff},n})^{1/2} ( J s d / m eff , n ) 1/2 gives the stated spatial rate.
Derivation of the local-kernel and spatial-spline comparison. Set a n ≍ J s − 1 a_n \asymp J_s^{-1} a n ≍ J s − 1 . Apart from logarithmic factors, the squared stochastic terms in equation (15) become
J m m n a n d N ‾ n + 1 m e f f , n a n d ≲ J s d ( J m m n N ‾ n + 1 m e f f , n ) . \frac{J_m}{m_na_n^d\overline{N}_n}+\frac{1}{m_{\mathrm{eff},n}a_n^d}
\lesssim J_s^d\left(\frac{J_m}{m_n\overline{N}_n}+\frac{1}{m_{\mathrm{eff},n}}\right). m n a n d N n J m + m eff , n a n d 1 ≲ J s d ( m n N n J m + m eff , n 1 ) .
For the uniform local rate, ℓ n = log K m + log ( a n − 1 ) ≲ log ( K s K m ) \ell_n=\log K_m+\log(a_n^{-1})\lesssim\log(K_sK_m) ℓ n = log K m + log ( a n − 1 ) ≲ log ( K s K m ) because K s ≍ J s d K_s\asymp J_s^d K s ≍ J s d . The comparison concerns the leading stochastic terms after the respective remainder conditions are imposed; b n b_n b n depends on the coordinate dimension and need not have the same order for the two procedures. The local phase transition boundary likewise becomes N ‾ n ≲ ( m e f f , n / J s d ) 1 / ( 2 q m ∗ ) / Θ n \overline{N}_n\lesssim(m_{\mathrm{eff},n}/J_s^d)^{1/(2q_m^*)}/\Theta_n N n ≲ ( m eff , n / J s d ) 1/ ( 2 q m ∗ ) / Θ n , the spatial-spline phase transition boundary. The leading stochastic terms therefore agree at matched spatial resolution, while the spatial approximation terms remain a n q u ∗ a_n^{q_u^*} a n q u ∗ and J s − q u ∗ J_s^{-q_u^*} J s − q u ∗ , respectively.