Abstract
In this paper a k-nearest neighbor type estimator of the marginal density function for a random field which evolves with time is considered. Considering dependence, the consistency and asymptotic distribution are studied for the stationary and nonstationary cases. In particular, the parametric rate of convergence \(\sqrt{T}\) is proven when the random field is stationary. The performance of the estimator is shown by applying our procedure to a real data example.
Similar content being viewed by others
References
Blanke D (2004) Adaptive sampling schemes for density estimation. J Stat Plan Inference 136(9):2898–2917
Blanke D, Bosq D (1997) Accurate rates of density estimators for continuous-time processes. Stat Probab Lett 33(2):185–191
Carbon M, Hallin M, Tran L (1996) Kernel density estimation for random fields: the L 1 theory. J Nonparametr Stat 6(2–3):157–170
Carbon M, Hallin M, Tran L (1997) Kernel density estimation for random fields (density estimation for random fields). Stat Probab Lett 36(2):115–125
Castellana JV, Leadbetter MR (1986) On smoothed probability density estimation for stationary processes. Stoch Process Appl 21(2):179–193
Chao–Gan Y, Yu-Feng Z (2010a) DPARSF: a MATLAB toolbox for “pipeline” data analysis of resting–state fMRI. Frontiers Syst Neurosci 4:1–7
Chao–Gan Y, Yu-Feng Z (2010b). http://www.nitrc.org
Doukhan P, Leon J, Portal F (1984) Vitesse de convergence dans le théoréme central limite pour des variables aleatoires mtlangeantes a valeurs dans un espace de Hilbert. C R Acad Sci Paris 289:305–308
Doukhan P, Louhichi S (1999) A new dependence condition and applications to moment inequalities. Stoch Process Appl 84:313–342
Doukhan P, Neumann M (2006) Probability and moment inequalities for sums of weakly dependent random variables, with applications. Stoch Process Appl 117:878–903
Fox MD, Raichle ME (2007) Spontaneous fluctuations in brain activity observed with functional networks. Nat Rev Neurosci 8:700–711
Geman D, Horowitz J (1980) Occupation densities. Ann Probab 8(1):1–67
Hallin M, Lu Z, Tran L (2001) Density estimation for spatial linear processes. Bernoulli 7(4):657–668
Hallin M, Lu Z, Tran L (2004) Kernel density estimation for spatial processes: the L 1 theory. Ann Stat 88:61–75
Kutoyants Y (2004) On invariant density estimation for ergodic diffusion processes. SORT 28(2):111–124
Labrador B (2008) Strong pointwise consistency of the k T -occupation time density estimator. Stat Probab Lett 78(9):1128–1137
Llop P, Forzani L, Fraiman R (2011) On local times, density estimation and supervised classification from functional data. J Multivar Anal 102(1):73–86
Nguyen H (1979) Density estimation in a continuous-time stationary Markov process. Ann Stat 7(2):341–348
Neumann M, Paparoditis E (2008) Goodness-of-t tests for Markovian time series models: central limit theory and bootstrap approximations. Bernoulli 14(1):14–46
Robinson PM (1983) Nonparametric estimators for time series. J Time Ser Anal 4:185–206
Rosenblatt M (1956) A central limit theorem and a strong mixing condition. Proc Natl Acad Sci USA 42:43–47
Rosenblatt M (1970) Density estimates and Markov sequences. In: Nonparametric techniques in statistical inference. Cambridge University Press, Cambridge, pp 199–210
Tang X, Liu Y, Zhang J, Kainz W (2008) In: Advances in spatio-temporal analysis. ISPRS, vol 5
Tran LT (1990) Kernel density estimation on random fields. J Multivar Anal 34(1):37–53
Tran L, Yakowitz S (1993) Nearest neighbour estimators for random fields. J Multivar Anal 44(1):23–46
Acknowledgements
We are most grateful to Daniel Fraiman for his very helpful insights to analyze the brain fMRI data. We would also like to thank Roberto Scotto for helpful discussions. The authors were supported by PICT2008-0921, PICT2008-0622, PI 62-309 and PIP 112-200801-0218.
Author information
Authors and Affiliations
Corresponding author
Appendices
Appendix A: Auxiliary results
Theorem 4
(Doukhan and Neumann 2006)
Suppose that X 1,X 2,…,X T are real-valued random variables defined on a probability space \((\varOmega, \mathcal{A}, P)\) with \(\mathbb {E} (X_{t} )=0\) and P(|X t |≤M)=1, for all t=1,…,T and some M<∞. Let \(S_{T} = \sum_{t=1}^{T} X_{t}\) and ϕ:ℕ2→ℝ+ be one of the following functions:
-
ϕ(u,v)=2v;
-
ϕ(u,v)=u+v;
-
ϕ(u,v)=uv;
-
ϕ(u,v)=ρ(u+v)+(1−ρ)uv, for some ρ∈(0,1).
We assume that there exist constants K,L 1,L 2<∞,μ≥0 and a nonincreasing sequence of real coefficients {α(n)} n≥1 such that, for all u-tuples (t 1,…,t u ) and all v-tuples (l 1,…,l v ) with 1≤t 1≤t u <t u +r=l 1≤l v ≤∞ the following inequality is fulfilled:
where
Then,
where Σ T can be chosen as any number greater than or equal to \(\sigma_{T}^{2}=\operatorname{var} (S_{T} )\) and
Theorem 5
(Doukhan et al. 1984)
Let X 1,X 2,…,X T be a sequence of α-mixing random variables verifying \(\mathbb{E} (X_{t} )=0\) and |X t |≤1, for all t=1,…,T. Let \(S_{T} = \sum_{t=1}^{T} X_{t}\) and denote γ=2/(1−θ) and \(\sigma= \sup_{1\le t \le T}\{\mathbb{E} (|X_{t}|^{\gamma } )^{1/\gamma}\}\). Then, there exist constants C 1 and C 2 which depend only on the mixing coefficients, such that for 0<θ<1,
where C 2,T =C 2 if T 1/2 σ≤1 and C 2,T =C 2 T 1/4 σ 1/2 if T 1/2 σ>1.
Theorem 6
(Robinson 1983)
Let \(\{V_{tT}\}_{t=1}^{T}\) be a triangular array of zero mean random variables and {b T } T≥1 a sequence of positive constants such that:
-
(i)
for each T, V tT , t=1,…,T are identically distributed and α-mixing with the mixing coefficients α(r) verifying
$$N \sum_{r=N}^{\infty} \alpha(r) \to0 \quad \textit{as } N \to \infty; $$ -
(ii)
there exists M>0 such that P(|V tT |≤M)=1 for all t=1,…,T;
-
(iii)
b T →0 and Tb T →∞ as T→∞;
-
(iv)
there exists σ 2>0 such that \(\mathbb{E} (V_{tT}^{2} )/b_{T} \to\sigma^{2}\) as T→∞;
-
(v)
\(\mathbb{E} (\vert V_{tT} V_{(t+s)T}\vert ) \le C b_{T}^{2}\) for s≥1, t=1,…,T where C is independent of T.
Then
Appendix B: Proof of main results
Proof of Theorem 1
Let x∈ℝ be fixed. By definition of complete convergence we need to show that for all ϵ>0,
Using the definition of the estimator \(\widehat{f}_{\mathcal{X}}(x)\), it is enough to prove that
where for \(\epsilon_{T} \doteq\frac{\epsilon}{v_{T}}\), the sets A T and B T are defined by
and
To prove the left-side inequality of (6) (the proof of the right-side inequality is identical and it will be omitted), let us define \(a_{T}=a_{T}(x) \doteq\frac{k_{T}}{2T|{\mathbf {S}}|(f_{\mathcal{X} }(x)+\epsilon_{T})}\) so that \(A_{T} = \{h^{\mathcal{X}}_{T} < a_{T} \}\). From the equivalence
we have
Next we define \(\overline{Y}_{Tt}\doteq Y_{Tt}-\mathbb{E} (Y_{Tt} )\) and the probability \(p_{T} \doteq P(\mathcal{X}_{t}({\mathbf{s}}) \in(x-a_{T}, x+a_{T}))\). Since \(\mathbb{E} (Y_{Tt} ) = |\mathbf{S}| p_{T}\), using the definition of a T we get
By the Mean Value Theorem there exists x T ∈(x−a T ,x+a T ) for which \(\frac{p_{T}}{2a_{T}}=f_{\mathcal{X}}(x_{T})\). In addition, by condition H2, the definition of a T and the fact that ϵ T →0 we have
from which it follows that \(\frac{p_{T}}{2a_{T}} \le f_{\mathcal{X}}(x) + \frac {\epsilon_{T}}{2}\) and then
Therefore, with this inequality in (8) we get
(a) Weakly dependent case: in order to use a Bernstein type inequality in (9), we need the following Lemma which will be proved in Appendix B.
Lemma 4
Under H3 for \(\mathcal{X}({\mathbf{s}})\) we have
-
(i)
for any u-tuple (t 1,…,t u ) and any v-tuple (l 1,…,l v ), 1≤t 1≤t u <t u +r=l 1≤l v ≤T we have
$$\bigl \vert \operatorname{cov} (\overline{Y}_{Tt_1} \cdots \overline{Y}_{Tt_u},\overline{Y}_{Tl_1} \cdots \overline {Y}_{Tl_v} )\bigr \vert \le \bigl(2|\mathbf{S}| \bigr)^{u+v} \phi(u,v) \alpha(r), $$where 2|S| is such that \(\vert \overline {Y}_{Tt}\vert \le2|\mathbf{S}|\) for all t, α(r)→0 with ϕ any of the functions given in H3;
-
(ii)
for some constant C,
$$\operatorname{var} \Biggl(\sum_{t=1}^T \overline{Y}_{Tt} \Biggr) \le C T. $$
Therefore, Lemma 4(i) implies that the sequence \(\{\overline{Y}_{Tt} \}_{t=1}^{T}\) is \((\mathcal{G},\alpha ,\psi)\)-weakly with f:ℝu→ℝ and g:ℝv→ℝ given by \(f(\overline{Y}_{Tt_{1}}, \ldots, \overline{Y}_{Tt_{u}})= \overline{Y}_{Tt_{1}} \cdots \overline{Y}_{Tt_{u}}\) and \(g(\overline{Y}_{Tl_{1}}, \ldots, \overline{Y}_{Tl_{v}})= \overline{Y}_{Tl_{1}} \ldots\overline{Y}_{Tl_{v}}\), respectively, and \(\psi: \mathcal{G}^{2} \times \mathbb{N}^{2} \to\mathbb{R}^{+}\) defined by ψ(f,g,u,v)=(2|S|)u+v ϕ(u,v) with ϕ any of the four functions given in H3. Therefore, applying Theorem 4 with K=(2|S|)2, M=2|S|, Γ T =C and Σ T =CT in (9) we have
and the result follows since (2μ+3)/(μ+2)<2 and H4 \((\frac{v_{T}}{k_{T}})^{2} T \to0\) and, as a consequence, \(\frac {v_{T}}{k_{T}}\to0\).
(b) α-mixing case: under H1, H2, H3′ and H4′ for \(\mathcal{X} ({\mathbf{s}})\). Since \(\mathcal{X}_{t}\) verifies H5, \(\overline{Y}_{Tt}\) inherits the same condition and then we apply the Bernstein inequality given in Theorem 5 to the random variables \(Z_{Tt}\doteq\frac{\overline{Y}_{tT}}{2|\mathbf{S}|}\) with |Z Tt |≤1 and \(\mathbb{E} (Z_{Tt} ) = 0\). For \(\gamma= \frac {2}{1-\theta}\) with 0<θ<1, it is easy to verify that \(\mathbb{E} (|Z_{Tt}|^{\gamma} )^{1/\gamma} \le (2|\mathbf{S}|)^{1-\gamma}\) so that σ≤(2|S|)1−γ and then, since \(\frac {C_{2,T}}{T^{1/4}\sigma^{1/2}} \ge C\) in (9) we have
which is finite by H4′. □
Proof of Theorem 2
For a fixed x∈ℝ we define
From the definition of \(\widehat{f}_{\mathcal{X}}(x)\) and analogously to (7) we have
Then the proof will be completed if we prove:
-
(a)
\(\frac{S_{T} - \mathbb{E} (S_{T} )}{s_{T}} \to\mathcal{N} (0, 1 )\) in distribution;
-
(b)
\(\frac{k_{T} - T|\mathbf{S}|p_{T}}{s_{T}} \to \frac {2|\mathbf{S}|}{c_{o}}z\).
(a) It will a consequence of \(\frac{S_{T} - \mathbb {E} (S_{T} )}{\sqrt{T} a_{T} } = \frac{1}{\sqrt{T} a_{T} }\sum_{t=1}^{T} \overline{Y}_{tT} \to\mathcal{N} (0, c_{0}^{2} )\) and \(\frac {s_{T}^{2} }{T a_{T}^{2}} \to c_{0}^{2}\) where \(c_{0}^{2}\) is given in H6. To prove the second part observe that
Then,
Since a T ∼k T /T, from H6 we get
from which follows that \(I \to c_{0}^{2}\) as T→∞. On the other hand, for N integer we write,
where the term IV is considered zero if T−t<N. Since
and \(A_{t}({\mathbf{s}}) \doteq\{\mathcal{X}_{t}({\mathbf{s}}) \in (x-a_{T},x+a_{T})\} \in\mathcal{M}_{t}^{t}\) and \(A_{t+l}(\mathbf{r}) \doteq \{\mathcal{X}_{t+l}({\mathbf{s}}) \in(x-a_{T}, x+a_{T})\} \in\mathcal{M}_{t+l}^{t+l}\) for each fixed s and r, hypothesis H5 implies that
from which follows
On the other hand, from H7 we have
for some constant C which implies that \(\mathit{III} \le2C N a_{T}^{2}\) and hence
Let ϵ>0 fixed and \(N = \lfloor\frac{\epsilon }{a_{T}^{2}} \rfloor\). For this choice we get \(|\mathit{II}| \le C \epsilon+ \frac{C}{\epsilon} N \sum_{l=N}^{\infty} \alpha(l)\) for T large enough (since \(a_{T} \sim \frac{k_{T}}{T} \to0\)). From hypothesis H5 there exists N 0 such that if M≥N 0, \(M \sum_{l=M}^{\infty} \alpha(l) < \epsilon^{2}\). On the other hand, again since a T →0, there exists T 0 such that if T≥T 0, \(\lfloor \frac {\epsilon}{a_{T}^{2}} \rfloor \ge N_{0}\) and then \(\lfloor \frac{\epsilon}{a_{T}^{2}} \rfloor \sum_{l=\lfloor\frac{\epsilon}{a_{T}^{2}}\rfloor}^{\infty} \alpha(l) < \frac{\epsilon^{2}}{2}\) which implies that II≤Cϵ which implies
To prove that \(\frac{S_{T} - \mathbb{E} (S_{T} )}{\sqrt{T} a_{T} } = \frac{1}{\sqrt{T} a_{T} }\sum_{t=1}^{T} \overline{Y}_{tT} \to \mathcal{N} (0, c_{0}^{2} )\) we will show that the variables \(V_{Tt} \doteq\overline{Y}_{tT}\) verify the assumptions (i)–(v) of the Lindeberg version of the CLT given in Theorem 6 with \(b_{T} = a_{T}^{2}\). As before, since \(\mathcal{X}_{t}\) verifies H5, \(\overline {Y}_{Tt}\) inherits the same condition from which (i) follows. Condition (ii) holds for M=2|S|. (iii) follows from (4) (since k T →∞ and \(a_{T} \sim\frac{k_{T}}{T} \to0\)), H5 implies (iv) (see (11)) and H7 implies (v) (see (12)). Therefore, from Theorem 6 we get the result since \(\sigma^{2} = c^{2}_{0}\).
To prove (b) we use the definition of a T to get
By Taylor Theorem and since f has two derivatives bounded, for x ∗ between x and u we have
Therefore, in (14) we have
Finally, by hypothesis (4) and (13) \(s_{T}^{-1} T a_{T}^{3} \rightarrow0\) then from (11) we get (b). □
Proof of Theorem 3
This proof will be an immediate consequence of Theorem 1 and of the following lemma which was proved in Llop et al. (2011, Lemma 5, p. 85). □
Lemma 5
Assume H1–H3 and choose two sequences {k T } and {v T } of positive real numbers converging to infinity such that, for each fixed s, \(v_{T} (T/k_{T})|\bar{e}_{T}({\mathbf{s}})| \rightarrow0\) a.co. For this sequences, suppose that H4 hold. Then for each x∈ℝ
where \(\widehat{f}_{e}\) is the estimator of f e .
Appendix C: Proof of the auxiliary lemmas
Proof of Lemma 2
We need to prove that if α(r)≤aρ r with 0<ρ<1 and a>0 then, \(\sum_{j=0}^{\infty}(j+1)^{k} \alpha(j) \le L_{1} L_{2}^{k} (k!)^{\mu}, \forall k \ge0\). For that, let us suppose that α(r)≤aρ r for some 0<ρ<1 and a>0. If f (k) denotes the kth derivative of f then
Let C(k,j) be denote the combinatorial number . Since
we have
where in the last equality we have used the Binomial Theorem. Then, with this inequality in (15), we get
Therefore, taking \(L_{1} = \frac{a}{1-\rho}\), \(L_{2}=\frac{1}{1-\rho}\) and μ=1 we get the result. □
Proof of Lemma 3
For T and x fixed, the function
is strictly increasing with G(0)=0. On the other hand, due to the existence of local time we can write
then, G(r)→|S| when r→∞ and therefore, the existence and uniqueness of \(h^{\mathcal{X}}_{T}(x)\) is ensured. For a further reading on local times see Geman and Horowitz (1980). □
Proof of Lemma 4
To prove part (i) let us consider the u-tuple (t 1,…,t u ) and the v-tuple (l 1,…,l v ) for 1≤t 1≤t u <t u +r=l 1≤l v ≤T and let C(k,j) be denote the combinatorial number . Since \(\mathbb{E} (Y_{Tt} )= |\mathbf{S}| p_{T}\) then,
Therefore, taking absolute value and considering that |p T |≤1 we get
To compute
we will consider the following events:
and
In addition, we will denote
Then,
and similarly
Hence,
where in the last inequality we have used hypothesis H4 since, for each fixed s,r, since k≤u and m≤v,
with 1≤t 1≤t u <t u +r=l 1≤l v ≤T. Finally, with inequality (18) in (16) we get
To prove part (ii) of this lemma, first observe that since \(\operatorname{var} (Y_{Tt} ) \le\mathbb{E} (Y_{Tt}^{2} ) \le|\mathbf{S}|^{2} p_{T}^{2} \le|\mathbf{S}|^{2}\) we may write
□
Rights and permissions
About this article
Cite this article
Forzani, L., Fraiman, R. & Llop, P. Density estimation for spatial-temporal models. TEST 22, 321–342 (2013). https://doi.org/10.1007/s11749-012-0313-3
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11749-012-0313-3