Nearest-Neighbor Radii under Dependent Sampling

Fix x\in {\mathbbm R}^d. Since \mathcal X_{(k)}^n(x)\in {\rm supp}(\mu) almost surely and r_x =\inf\{\|y-x\|: y\in {\rm supp}(\mu)\}, we have

Thus, it remains to prove that

Fix \varepsilon>0, and define

Since B_{x,r_x+\varepsilon} intersects {\rm supp}(\mu) by definition of r_x, we have p_x>0.

We observe that the event \|\mathcal X_{(k)}^n(x)-x\|>r_x+\varepsilon occurs if and only if fewer than k observations in the n-th row fall inside B_{x,r_x+\varepsilon}. Then we have

We set

Then {\mathrm E}[Z_{n,i}]=0 and |Z_{n,i}|\le 1. Since Z_{n,i} is a measurable function of X_{n,i}, the sequence [Z_{n,i}]_{i=1}^n is again \alpha-mixing with the same mixing coefficients as [X_{n,i}]_{i=1}^n. Moreover, for any m\in\{1,\dots,n\} and any 0\le j\le n-1, the covariance inequality [davydov1968convergence, Corollary of Lemma 2.1] yields

Since \sum_{\ell=1}^\infty \alpha_\ell<\infty, we set M = 1+24\sum_{\ell=1}^\infty \alpha_\ell <\infty, then

Since k/n\to 0 and p_x>0, for all sufficiently large n, we have k-np_x \le -\frac{np_x}{2}. Then

We now apply Lemma 1 with m_n = \lfloor \frac{3}{64}\frac{np_x}{\log n}\rfloor. Then m_n\to\infty and m_n/n \to 0. Therefore for all sufficiently large n, we have 64M \le \frac{4}{3}p_x m_n and \frac{np_x}{2}>4m_n, and then the conditions of Lemma 1 are satisfied with S(n)=1 and \epsilon_n=np_x/2. Using \sigma_{n,m_n}^2\le M m_n, we obtain

By the choice of m_n, the exponential term is O(n^{-2}).

We now assume the mixing coefficients satisfy \alpha_m \lesssim m^{-(1+\gamma)} for some \gamma>1. Then

which is summable in n. Then

By the Borel–Cantelli lemma,

That is, for every fixed \varepsilon>0,

Let \varepsilon=1/j, j\in\mathbb N, and intersect the resulting probability-one events. Then we obtain

Together with the lower bound \|\mathcal X_{(k)}^n(x)-x\|\ge r_x a.s. for every n, we obtain

This completes the proof.

Fix r>0, and define

Let

By the definition of the k-nearest-neighbor radius,

We now use a blocking argument. Partition the row into alternating kept and discarded blocks of length b_n. Set M_n=\lfloor \frac{n}{2b_n}\rfloor, and for u=1,\dots,M_n, define the kept block

Let

Since S_n^{\rm keep}(r)\le N_n(x,r), we obtain

Consider the radius

and assume r_j\le r_0. By Assumption 2,

Since b_n/n \to 0 and M_n b_n \ge n/4 for all sufficiently large n,

Therefore, on the event \{S_n^{\rm keep}(r_j)<k\}, since \mu_j\ge 2k,

Then

Therefore, if we define

we have

Next we apply Lemma 1 for the block sequence [Z_u]_{u=1}^{M_n}. Since each Z_u is measurable with respect to the coordinates in the kept block I_u, and successive kept blocks are separated by a discarded block of length b_n, the sequence [Z_u]_{u=1}^{M_n} is \alpha-mixing with coefficients

In particular, \bar\alpha_1\le \alpha_{b_n}\le e^{-\gamma b_n}\le n^{-8}, as long as b_n is chosen with a sufficiently large constant times \log n. Also, 0\le W_u(r_j)\le b_n, so |Z_u| = |W_u(r_j)-{\mathrm E} W_u(r_j)| \le b_n. Thus the boundedness parameter in Lemma 1 is S(M_n)=b_n.

To bound the variance term \sigma^2_{M_n,1}, since W_u(r_j) is a sum of b_n Bernoulli indicators, we have W_u(r_j)^2\le b_n W_u(r_j), and therefore

Then for the block sequence with block parameter m=1,

We now apply Lemma 1 to [Z_u]_{u=1}^{M_n} with m=1 and threshold \varepsilon=\mu_j/2=M_n b_n p_{r_j}/2. Since \varepsilon\ge k\ge K_0\log n, and b_n\asymp \log n, we may choose K_0 sufficiently large so that \varepsilon>4b_n for all sufficiently large n. Then by Lemma 1,

Using \mu_j=M_n b_n p_{r_j}, we obtain \frac43 \mu_j b_n = \frac43 M_n b_n^2 p_{r_j}, so the denominator is bounded by

and then

Since M_n\ge n/(4b_n) for all large n,

Therefore, for some universal constant c_0>0,

Finally, using b_n\asymp \log n, \bar\alpha_1\le n^{-8}, and M_n\le n/(2b_n), we obtain 4M_n\bar\alpha_1 \le C n^{-7}, and then

after adjusting constants. This completes the proof.

Let

and choose J_0 maximal such that 2^{J_0}r_* \le r_0. Since the support of \mu has diameter at most D, we have R_{n,k}(x)\le D almost surely.

We decompose the moment over

and we obtain

We now apply Theorem 2. For every 0\le j\le J_0,

and then

We now bound the three terms on the right-hand side. For the first term, since k\ge K_0\log n, the quantity k/\log n is bounded below by a positive constant. Therefore

for some constant c_1>0. For the second term, since 2^{J_0}r_*\le r_0<2^{J_0+1}r_*, we have 2^{J_0}\asymp r_*^{-1}, where the constants depend only on r_0. Then

and then

which is negligible. The third term C D^p n^{-7} is negligible as well. Combining the above estimates yields

Since

we obtain

This completes the proof.

Let

By Assumption 3,

Therefore,

where,

By Markov’s inequality,

and

we have

On the event \{R_{n,k}(x)>r^-\}, we have R_{n,k}(x)^p\ge (r^-)^p, and then

By the definition of r^-, we obtain

The completes the proof.

The conclusion follows directly by combining the upper and lower bounds established in Theorems 3 and 4.