Abstract
We provide an improved version of the Darling–Erdős theorem for sums of i.i.d. random variables with mean zero and finite variance. We extend this result to multidimensional random vectors. Our proof is based on a new strong invariance principle in this setting which has other applications as well such as an integral test refinement of the multidimensional Hartman–Wintner LIL. We also identify a borderline situation where one has weak convergence to a shifted version of the standard limiting distribution in the classical Darling–Erdős theorem.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Let \(X, X_1, X_2, \ldots \) be i.i.d. random variables and set \(S_n=\sum _{j=1}^n X_j, n \ge 1.\) Further set \(Lt =\log _e(t \vee e), LL t = L(Lt)\) and \( LLL t = L( LL t), t \ge 0.\) In 1956, Darling and Erdős proved that under the assumption \(\mathbb {E}|X|^3 < \infty ,\) \(\mathbb {E}X^2=1\) and \(\mathbb {E}X=0,\) the following convergence in distribution result holds,
where \(a_n = \sqrt{2 LL n}, b_n = 2 LL n + LLL n{/}2 - \log (\pi ){/}2\) and \(\tilde{Y}\) is a random variable which has an extreme value distribution with distribution function \(y \mapsto \exp (-\exp (-y)).\)
The above third moment assumption was later relaxed in [14, 16] to \(\mathbb {E}|X|^{2 + \delta } < \infty \) for some \(\delta >0,\) but the question remained open whether a finite second moment would be already sufficient.
This was finally answered in [6], where it is shown that (1.1) holds if and only if
Moreover, it is shown in [6] that the above result holds more generally under the assumption of a finite second moment if one replaces the normalizers \(\sqrt{k}\) by \(\sqrt{B_k}\), where \(B_n=\sum _{j=1}^n \sigma _j^2\) and
for some \(p \ge 2.\) So we have under the classical assumption \(\mathbb {E}X^2 < \infty \) and \(\mathbb {E}X=0,\)
For some further related work on the classical Darling–Erdős theorem, the reader is referred to [2, 4, 10] and the references in these articles.
The Darling–Erdős theorem is also related to finding an integral test refining the Hartman–Wintner LIL, a problem which was already addressed by Feller in 1946. Here, one can relatively easily prove that the classical Kolmogorov–Erdős–Petrowski integral test for Brownian motion holds for sums of i.i.d. mean zero random variables if one has \(\mathbb {E}|X|^{2 + \delta } < \infty \) for some \(\delta >0\). In this case, one has for any non-decreasing function \(\phi {:}\,]0,\infty [ \rightarrow ]0,\infty [,\)
according as \(\sum _{n=1}^{\infty } n^{-1}\phi (n)\exp (-\phi ^2(n){/}2)\) is finite or infinite.
Feller proved that this result remains valid under the second moment assumption if one replaces \(\sqrt{n}\) by \(\sqrt{B_n}\) defined as above with \(p \ge 4\). Similarly, as for the Darling–Erdős theorem, this implies that the Kolmogorov–Erdős–Petrowski integral test holds in its original form if
The proof in [11] was based on a skillful double truncation argument which only worked for symmetric distributions. Finally, in [6] an extension of this argument to the general non-symmetric case was found so that we now know that most results in [11] are correct (see also [1] for more historical background).
There is still one question in the paper [11] which has not yet been addressed, namely whether it is possible to make the theorem “slightly more elegant” by replacing the sequence \(\sqrt{B_n}\) by \(\sqrt{n}\sigma _n.\) Feller writes that he “was unable to devise a proof simple enough to be justified by the slight improvement of the theorem” (see p. 632 in [11]). We believe that we have found a simple enough proof of Feller’s claim (see Step 3 in the proof of Theorem 2.3).
This leads to the following improved version of the Darling–Erdős theorem under the finite second moment assumption:
At the same time, we can show that there is a much wider choice for the truncation level in the definition of \(\sigma _n^2\). For instance, it is possible to define \(\sigma _n^2\) as \(\mathbb {E}X^2I\{|X| \le \sqrt{n}\}.\)
This improved version of the Darling–Erdős theorem will actually follow from a general result for d-dimensional random vectors which will be given in the following section.
2 Statement of Main Results
We now consider i.i.d. d-dimensional random vectors \(X, X_1, X_2, \ldots \) such that \(\mathbb {E}|X|^2 < \infty \) and \(\mathbb {E}X=0\), where we denote the Euclidean norm by \(|\cdot |\). The corresponding matrix norm will be denoted by \(\Vert \cdot \Vert \), that is, we set
for any (d, d)-matrix A. It is well known that \(\Vert A\Vert =\) the largest eigenvalue of A if A is symmetric and nonnegative definite.
Let again \(S_n:=\sum _{j=1}^n X_j, n \ge 1.\) Horváth [12] obtained in 1994 the following multidimensional version of the Darling–Erdős theorem assuming that \(\mathbb {E}|X|^{2+\delta } < \infty \) for some \(\delta > 0\) and that Cov(X) (\(=\) the covariance matrix of X) is equal to the d-dimensional identity matrix I,
where \(\tilde{Y}\) has the same distribution as in dimension 1,
and \(\varGamma (t), t > 0\) is the Gamma function. Recall that \(\varGamma (1{/}2)=\sqrt{\pi }\) so that this extends the one-dimensional Darling–Erdős theorem.
We are ready to formulate our general result. We consider non-decreasing sequences \(c_n\) of positive real numbers satisfying for large n,
where \(\epsilon _n \rightarrow 0.\)
Further, let for each \(n,\,\varGamma _n\) be the symmetric nonnegative definite matrix such that
If the covariance matrix of \(X=(X^{(1)},\ldots , X^{(d)})\) is positive definite, the matrices \(\varGamma _n\) will be invertible for large enough n. Replacing \(c_n\) by \(c_{n \vee n_0}\) for a suitable \(n_0 \ge 1\) if necessary, we can assume w.l.o.g. that all matrices \(\varGamma _n\) are invertible.
Theorem 2.1
Let \(X, X_n, n \ge 1\) be i.i.d. mean zero random vectors in \(\mathbb {R}^d\) such that \(E|X|^2 < \infty \) and \(\mathrm {Cov}(X) = I.\) Then, we have for any sequence \(\{c_n\}\) satisfying condition (2.2),
where \(\tilde{Y}{:}\,\varOmega \rightarrow \mathbb {R}\) is a random variable such that
Under the additional assumption
we also have
It is easy to see that condition (2.5) is satisfied if \(\mathbb {E}|X|^2 LL |X|< \infty .\) This latter condition, however, is more restrictive than (2.5).
It is natural to ask whether this condition is also necessary as in the one-dimensional case (see Theorem 2 in [6]). This question becomes much more involved in the multidimensional case, and we get a slightly weaker result, namely that the following condition
is necessary for (2.6).
To prove this result, we show that if condition (2.7) is not satisfied, then \(a_n \max _{1 \le k \le n} |S_k|{/}\sqrt{k} - b_{d,n}\) cannot converge in distribution to any variable of the form \(\tilde{Y}+c,\) where \(c \in \mathbb {R}.\)
If one allows this larger class of limiting distributions, condition (2.7) is optimal. There are examples where \(\mathbb {E}|X|^2I\{|X| \ge t\} = O(( LL t)^{-1})\) and \(a_n \max _{1 \le k \le n} |S_k|{/}\sqrt{k} - b_{d,n}\) converges in distribution to \(\tilde{Y}+c\) for some \(c \ne 0\) (see Theorem 6.1 below).
Theorem 2.2
Let \(X, X_n, n \ge 1\) be i.i.d. mean zero random vectors in \(\mathbb {R}^d\) such that \(E|X|^2 < \infty \) and \(\mathrm {Cov}(X) = I\) and suppose that there exists a \(c \in \mathbb {R}\) such that
where \(\tilde{Y}\) is as in Theorem 2.1. Then, condition (2.7) holds.
Our basic tool for proving the above results is a new strong invariance principle for sums of i.i.d. random vectors which is valid under a finite second moment assumption. If one has an approximation with an almost sure error term of order \(o(\sqrt{n{/} LL n})\), one can obtain the Darling–Erdős theorem directly from the normally distributed case. The problem is that it is impossible to get such an approximation under the sole assumption of a finite second moment. In [6], it was shown that one needs the “good” approximation of order \(o(\sqrt{n{/} LL n})\) only if the sums \(|S_n|\) are large and it was shown that in dimension 1 one can obtain approximations which are particularly efficient for the random subsequence where the sums are large. Using recent results on d-dimensional strong approximations (see [9, 15]), we are now able to obtain an analogue of the approximation in [6] in the d-dimensional setting [see Lemma 3.1 and relation (3.6) below]. As an additional new feature, we also show that an approximation by \(\varGamma _n \sum _{j=1}^n Z_j\) is possible where \(Z_j, j\ge 1\) are i.i.d. \(\mathcal {N}(0,I)\)-distributed random vectors with \(\mathcal {N}(0,\varSigma )\) denoting the d-dimensional normal distribution with mean zero and covariance matrix \(\varSigma .\) This type of approximation leads to the improved versions of the Darling–Erdős theorem and Feller’s integral test as indicated in Sect. 1.
Theorem 2.3
Let \(X, X_n, n \ge 1\) be i.i.d. mean zero random vectors in \(\mathbb {R}^d\) such that \(E|X|^2 < \infty \) and \(\mathrm {Cov}(X) = \varGamma ^2,\) where \(\varGamma \) is a symmetric nonnegative definite (d,d)-matrix. Let \(c_n\) be a non-decreasing sequence of positive real numbers satisfying condition (2.2) for large n and let \(\varGamma _n\) be defined as in (2.3). If the underlying p-space \((\varOmega , \mathcal {F}, \mathbb {P})\) is rich enough one can construct independent \(\mathcal {N}(0,I)\)-distributed random vectors \(Z_n, n \ge 1\) such that we have for the partial sums \(T_n:=\sum _{j=1}^n Z_j, n \ge 1,\)
-
(a)
\(S_n - \varGamma \;T_n = o(\sqrt{n LL n})\) as \(n \rightarrow \infty \) with prob. 1,
-
(b)
\(\mathbb {P}\{ |S_n - \varGamma _n T_n| \ge 2\sqrt{n}{{/}} LL n, |S_n| \ge \frac{4}{3}\Vert \varGamma \Vert \sqrt{n LL n} \hbox { infinitely often}\}=0,\) and
-
(c)
\(\mathbb {P}\{ |S_n - \varGamma _n T_n| \ge 2\sqrt{n}{/} LL n, |\varGamma T_n| \ge \frac{4}{3}\Vert \varGamma \Vert \sqrt{n LL n} \hbox { infinitely often}\}=0.\)
Combining our strong invariance principle with the Kolmogorov–Erdős–Petrowski integral test for d-dimensional Brownian motion, one obtains by the same arguments as in Section 5 of [6] the following result,
Theorem 2.4
Let \(X, X_n, n \ge 1\) be i.i.d. mean zero random vectors in \(\mathbb {R}^d\) such that \(E|X|^2 < \infty \) and \(\mathrm {Cov}(X) = \varGamma ^2,\) where \(\varGamma \) is a symmetric positive definite (d,d)-matrix. Let \(c_n\) be a non-decreasing sequence of positive real numbers satisfying condition (2.2) for large n and let \(\varGamma _n\) be defined as in (2.3). Then, we have for any non-decreasing function \(\phi {:}\,]0,\infty [ \rightarrow ]0,\infty [,\)
according as
Note that we can assume w.l.o.g. that all matrices \(\varGamma _n\) are invertible since they converge to \(\varGamma \) which is invertible.
Let \(\lambda _n (\varLambda _n)\) be the smallest (largest) eigenvalue of \(\varGamma _n, n \ge 1.\) Assuming that Cov(X) \(=I\), we can infer from Theorem 2.4,
and
which is the d-dimensional version of the result conjectured by Feller [11].
The proof of our strong invariance principle (\(=\) Theorem 2.3) will be given in Sect. 3. In the two subsequent Sects. 4 and 5, we will show how Theorems 2.1 and 2.2 follow from the strong invariance principle. In Sect. 6, we return to the real-valued case and show that if \(\mathbb {E}X^2I\{|X| \ge t\} \sim c( LL t)^{-1}\) that then (1.1) still remains valid if we replace \(\tilde{Y}\) by \(\tilde{Y}-c\). Finally, we answer a question which was posed in [13].
3 Proof of the Strong Invariance Principle
Our proof is divided into three steps.
Step 1 We recall a double truncation argument which goes back to Feller [11] for symmetric random variables. This was later extended to non-symmetric random variables in [6] and finally to random elements in Hilbert space in [7]. To formulate the relevant result, we need some extra notation. We set
and we denote the corresponding sums by \(S'_n, \overline{S}'_n, S''_n, \overline{S}''_n, S'''_n, \overline{S}'''_n.\)
Then, we have (see [7], Lemmas 11 and 12)
and
Step 2 Let \(\varSigma _n\) be the sequence of symmetric nonnegative definite matrices such that \(\varSigma _n^2\) is the covariance matrix of \(X'_n\) for \(n \ge 1.\) Furthermore, let A(t) be the symmetric nonnegative definite matrices satisfying
It is easy to see that \(A(t)^2, t \ge 0\) is monotone, that is, \(A(t)^2 - A(s)^2\) is nonnegative definite if \(0 \le s \le t.\)
This implies that \(A(t), t \ge 0\) is monotone as well (see Theorem V.1.9 in [3]). Consequently, \(A(c_n), n \ge 1\) is a monotone sequence of symmetric nonnegative definite matrices whenever \(c_n\) is non-decreasing. Moreover, \(A(c_n)\) converges to \(\varGamma \) if \(c_n \rightarrow \infty .\)
We have the following strong approximation result, where we set
Lemma 3.1
If the underlying p-space is rich enough, one can construct independent random vectors \(Z_i \sim \mathcal {N}(0, I)\) such that
and
Proof
-
(i)
We first show that one can construct independent \(\mathcal {N}(0,I)\)-distributed random vectors such that
$$\begin{aligned} \overline{S}'_n - \sum _{i=1}^n \varSigma _i Z_i = o\left( \sqrt{n}{/} LL n\right) \text { a.s. } \end{aligned}$$By Corollary 3.2 from [9] and the fact that \(\mathbb {E} | \overline{X}'_n |^3 \le 8 \mathbb {E} | X'_n |^3,\) it is enough to show
$$\begin{aligned} \sum _{n =1}^{\infty } \mathbb {E} \left| X'_n \right| ^3 {{\Big /}}\left( \frac{ \sqrt{ n } }{ LL n } \right) ^3 < \infty . \end{aligned}$$Using the simple inequality,
$$\begin{aligned} \mathbb {E} | X'_n |^3 \le \mathbb {E} |X|^{2 + \delta } I\{|X| \le \sqrt{n}\} \sqrt{n}^{1-\delta }( LL n)^{-5(1-\delta )},\quad 0< \delta <1, \end{aligned}$$we find (setting \(\delta =2{/}5\)) that the above series is
$$\begin{aligned} \le \sum _{n=1}^{\infty } \mathbb {E} |X|^{12{/}5} I\{|X| \le \sqrt{n}\} {/}\sqrt{n}^{12{/}5} \end{aligned}$$Using a standard argument (see for instance, the proof of part (a) of Lemma 3.3 in [9]), one can show that this last series is finite whenever \(\mathbb {E}|X|^2 < \infty .\)
-
(ii)
To complete the proof of (3.3), it is now sufficient to show that
$$\begin{aligned} \sum _{i=1}^n (\varSigma _i - \tilde{\varGamma }_i)Z_i = o(\sqrt{n}{/} LL n) \hbox { a.s.} \end{aligned}$$By a standard argument, this follows if
$$\begin{aligned} \sum _{n=1}^{\infty } \frac{ \mathbb {E} \left| \left( \varSigma _n - \tilde{\varGamma }_n \right) Z_n \right| ^2 }{ n{/}( LL n)^2 } < \infty . \end{aligned}$$(3.5)Since \(Z_n \sim \mathcal {N}(0, I)\), we have
$$\begin{aligned} \mathbb {E} \left| \left( \varSigma _n - \tilde{\varGamma }_n \right) Z_n \right| ^2 \le d \left\| \varSigma _n - \tilde{\varGamma }_n \right\| ^2 \le d\Vert \varSigma _n^2 -\tilde{ \varGamma }_n^2\Vert , \end{aligned}$$where we have used Theorem X.1.1 in [3] for the last inequality. From the definition of \(\varSigma _n\) and \(\tilde{\varGamma }_n\), it is obvious that
$$\begin{aligned} \langle x, (\tilde{\varGamma }_n ^2-\varSigma _n^2 )x \rangle = \left( \mathbb {E} \langle X, x\rangle I \left\{ | X| \le \sqrt{n}{/}( LL n)^5 \right\} \right) ^2,\quad x \in \mathbb {R}^d. \end{aligned}$$The last expression equals \( \left( \mathbb {E} \langle X, x\rangle I \{ | X| > \sqrt{n}{/}( LL n)^5 \} \right) ^2\) since \( \mathbb {E} \langle X, x\rangle =0.\) Hence \(\left\| \varSigma _n^2 - \tilde{\varGamma }_n^2 \right\| \le \left( \mathbb {E} | X| I \{ | X| > \sqrt{n}{/}( LL n)^5 \} \right) ^2 \le (\mathbb {E}|X|^2)^2 n^{-1} ( LL n)^{10} \). It is easy now to see that the series in (3.5) is finite.
-
(iii)
Finally note that
$$\begin{aligned} S_n - \sum _{i=1}^n \varGamma Z_i = (S_n - \overline{S}'_n) + \left( \overline{S}'_n -\sum _{i=1}^n \tilde{\varGamma }_i Z_i\right) + \sum _{i=1}^n (\tilde{\varGamma }_i -\varGamma )Z_i, \end{aligned}$$where the first two terms are of almost sure order \(o(\sqrt{n LL n})\) by (3.1) and (3.3), respectively. Since \(\tilde{\varGamma }_n \rightarrow \varGamma \) as \( n \rightarrow \infty \), we also have that
$$\begin{aligned} \sum _{i=1}^n (\tilde{\varGamma }_i -\varGamma )Z_i=o(\sqrt{n LL n}) \hbox { a.s.}, \end{aligned}$$and we can conclude that indeed \(S_n - \sum _{i=1}^n \varGamma Z_i = o(\sqrt{n LL n})\) a.s. Lemma 3.1 has been proven. \(\square \)
Step 3 Combining Lemma 3.1 with relations (3.1) and (3.2), we find that
We next show that
where
and \(c_n\) is an arbitrary non-decreasing sequence of positive real numbers satisfying condition (2.2) for large n.
Using that \(\{|\varGamma T_n| \ge \Vert \varGamma \Vert x\} \subset \{|T_n| \ge x\}, x >0\), we get from (3.6) and (3.7):
Further recall that \(S_n -\varGamma T_n = o(\sqrt{n LL n})\) a.s. (see Lemma 3.1). Consequently, we can infer from (3.8) that
We see that the proof of Theorem 2.3 is complete once we have established (3.7). Toward this end, we need the following inequality which is valid for normally distributed random vectors \(Y{:}\,\varOmega \rightarrow \mathbb {R}^d\) with mean zero and covariance matrix \(\varSigma \):
where \(\sigma ^2\) is the largest eigenvalue of \(\varSigma \) (see Lemma 4 in [7]).
From (3.10), we trivially get that
Though this last inequality is clearly suboptimal, it will nevertheless be more than sufficient for the proof of (3.7).
Proof of (3.7) To simplify notation, we set \(d_n:=\sqrt{n}{/}( LL n)^5, n \ge 1,\) \(n_k:=2^k\) and \(\ell _k:=[2^{k-1}{/}(Lk)^5], k \ge 0.\) By the Borel–Cantelli lemma, it is enough to show that
Set \(\tilde{\mathbb {N}}:= \tilde{\mathbb {N}}_1 \cap \tilde{\mathbb {N}}_2\), where
and
Then, it is easy to see that the above series is finite if
and
To bound the series in (3.12), we first note that employing the Lévy inequality for sums of independent symmetric random vectors, one obtains
where we have also used the fact that
This follows easily from the monotonicity of the sequence \(\varGamma _n.\)
To bound \(p_{k,1}\), we first note that by Theorem X.1.1 in [3] for \(1 \le j \le \ell _k,\)
Apply (3.11) with \(Y= \sum _{j=1}^{n_k} (\varGamma _{n_{k}} - \tilde{\varGamma }_j)Z_j.\) Then, clearly \(\mathbb {E}Y = 0\) and, moreover, by independence of the random vectors \(Z_j\), we have for \(k \in \tilde{\mathbb {N}},\)
We conclude that
Similarly, we obtain
It is now clear that the series in (3.12) is finite.
To show that the series in (3.13) is finite, we note that by (3.11) and the Lévy inequality,
where \(\eta =(16d)^{-1}\) and it is enough to check that
To verify that this series is finite, observe that by the argument used in (3.14) we have,
which implies
We conclude that
So the proof of (3.7) is complete if we show that
We need another lemma.
Lemma 3.2
Consider two sequences \(c_{k,i}, k \ge 1\) of positive real numbers satisfying for large enough k,
where \(0<\delta <1.\) Set \(\varGamma _{k,i} := A(c_{k,i}), i=1,2, k \ge 1.\) Then, we have,
Proof
Using the same argument as in (3.14), we have for large k,
where \(\beta _j :=\mathbb {E}|X|^2 I\{2^{j-1} < |X|^2 \le 2^j\}, j \ge 1\).
We can conclude that for some \(k_0 \ge 1\) and a suitable \(j_0 \ge 0,\)
where
and
It is easy to see that \(m_1(j) \ge j-3j^{\delta } \ge j{/}2\) and \(m_2(j) \le j + 4j^{\delta }\) for large j.
Consequently, we have for large j,
We obviously have \(\sum _{j=1}^{\infty } \beta _j < \infty \) (as \(\mathbb {E}|X|^2 < \infty \)). Combining relations (3.17) and (3.18), we obtain the assertion of the lemma. \(\square \)
We apply the above lemma with \(c_{k,1} = c_{n_k}, c_{k,2} = d_{\ell _k}, k \ge 1.\) From condition (2.2), we readily obtain that for large k,
where \(\epsilon '_k:=\epsilon _{n_k} \rightarrow 0\) so that condition (3.16) is satisfied for any \(\delta > 0\). This is also the case for the sequence \(c_{k,2}\). So we can choose \(\delta =\eta {/}2\), and it follows that
which shows that (3.15) holds.
4 Proof of Theorem 2.1
We first prove (2.4). Set \(k_n= [\exp ((Ln)^{\alpha })]\), where \(0< \alpha <1\). Then, it follows from the d-dimensional version of the Hartman–Wintner LIL that for any given \(\epsilon >0,\) with prob. 1,
where \(\lambda _k\) is the smallest eigenvalue of \(\varGamma _k\). As \(\lambda _k \nearrow 1,\) we can conclude that for large enough n,
which is \(\le \sqrt{2 LL n}\) if we choose \(\epsilon \) small enough. It follows that
So (2.4) holds if and only if
where \(K_n:=\{k_n+1,\ldots , n\}.\)
We split \(K_n\) into two random subsets:
In view of Theorem 2.3(b) (where we set \(\varGamma =I\)), there are with prob. 1 only finitely many k’s such that
where \(\lambda _k\) is again the smallest eigenvalue of \(\varGamma _k\).
As \(\lambda _k \nearrow 1,\) we can conclude that with prob. 1 there are only finitely many k’s such that
and it follows that
We see that (2.4) is equivalent to
From the definition of the sets \(K_{n,1}(\cdot )\), we easily get that
By Slutsky’s lemma (2.4) holds if and only if
Looking at Theorem 2.3(c), we can also conclude that
and the proof of (2.4) further reduces to showing
Using the same argument as in (4.1), we also see that
and we have shown that (2.4) holds if
This is the Darling–Erdős theorem for normally distributed random vectors which follows from (2.1). Thus, (2.4) has been proven.
We now turn to the proof of (2.6). By Slutsky’s lemma and (4.1), it is enough to show that
Using the triangular inequality, it is easy to see that
From (2.4), it follows that \( (\max _{1 \le k \le n} |\varGamma _k^{-1}S_k|{/}\sqrt{k}){/}\sqrt{ LL n}\) is stochastically bounded. By assumption (2.5), we also have that \(\Vert I -\varGamma _{k_n}\Vert = o(( LL n)^{-1})\). Recalling that \(a_n=\sqrt{2 LL n},\) we see that \(\varDelta _n \mathop {\rightarrow }\limits ^{\mathbb {P}} 0\) and our proof of Theorem 2.1 is complete. \(\square \)
5 Proof of Theorem 2.2
Using the same arguments as in the proof of Theorem 2.1, we can infer from (2.8) via relations (3.4) and (3.6) that
where \(T'_k =\sum _{j=1}^k \tilde{\varGamma }_j Z_j\) and the random vectors \(Z_j\) are i.i.d. with \(\mathcal {N}(0, I)\)-distribution and \(k_n \le \exp ((Ln)^{\alpha })\) for some \(0< \alpha <1.\)
Our first lemma gives an upper bound of \(\mathbb {P}\{M_n > t\}\) via the corresponding probability for the maximum of a subcollection of the random variables \(\{|T'_k|{/}\sqrt{k}{:}\,k_n \le k \le n\}\) (see Lemma 4.3 in [5] for a related result).
Let \(0< \xi < 1\) be fixed. Set
Then, \(m_{N} \le n \le m_{N+1}\). Also note that the sequence \(m_j\) depends on n and \(\xi \).
Next, set
so that \(j_n \sim \xi ^{-1}( LL n)^2\) and \(k_n \sim Ln\) as \(n \rightarrow \infty .\)
Finally to simplify notation, we set \(f_n(y)=(b_{d,n} + y){/}a_n, y \in \mathbb {R}\) so that
Lemma 5.1
Given \( 0< \delta < 1\), we have for \(y \in \mathbb {R}\) and \(n \ge n_0=n_0(\xi ,\delta ,y),\)
provided that \(0 < \xi \le \delta ^{3}{/}(36d).\)
Proof
Noting that
it is enough to show that
if \(\xi \) is sufficiently small.
Consider the following stopping time,
Then, it is obvious that the probability in (5.2) is bounded above by
Furthermore, we have for \(j_n +1 \le j \le N-1,\)
Next observe that
After some calculation, we find that for large enough n,
where the last inequality holds since \(\xi \le \delta {/}24.\) We trivially have by Markov’s inequality,
which is \(\le \delta \) by our condition on \(\xi .\)
It follows that
Concerning the second term in (5.3), simply note that
Arguing as above, we readily obtain,
Combining relations (5.4) and (5.5) and recalling (5.3), we see that
This implies (5.2) since
and the proof of Lemma 5.1 is complete. \(\square \)
We finally need the following lemma,
Lemma 5.2
Let Y be a d-dimensional random vector with distribution \(\mathcal {N}(0,\varSigma ),\) where \(d \ge 2.\) Assume that the largest eigenvalue of \(\varSigma \) is equal to 1 and has multiplicity \(d-1.\) Denote the remaining (smallest) eigenvalue of \(\varSigma \) by \(\sigma ^2\). Then, we have:
where \(Z{:}\,\varOmega \rightarrow \mathbb {R}^{d-1}\) has a normal\((0, I_{d-1})\)-distribution.
Proof
If \(d \ge 3\), Lemma 5.2 follows by integrating the inequality given in Lemma 1(a) of [7].
To prove Lemma 5.2 if \(d=2\), we proceed similarly as in [7]. Choose an orthonormal basis \(e_1, e_2\) of \(\mathbb {R}^2\) consisting of two eigenvectors corresponding to the eigenvalues 1 and \(\sigma ^2 \in ]0,1[\) of \(\varSigma .\) Then,
where \(\eta _i, 1 \le i \le 2 \) are independent standard normal random variables.
It is then obvious that
where \(R_1\) and \(R_2{/}\sigma ^2\) have Chi-square distributions with 1 degree of freedom. Denote the densities of \(R_1+R_2,\,R_1\) and \(R_2\) by \(h, h_1, h_2\).
Then, \(h_2(y)=h_1(y{/}\sigma ^2){/}\sigma ^2\) and
Using that \(h_1(y)=(2\pi )^{-1{/}2} y^{-1{/}2} e^{-y{/}2}, y > 0\), we can infer that
Employing the trivial inequality \(e^{-x{/}2}\le x^{-1{/}2}, x > 0,\) it follows that
We can conclude that for \(t \ge 0,\)
and Lemma 5.2 has been proven. \(\square \)
Recall that \(d_n=\sqrt{n}{/}( LL n)^5\) and \(\tilde{\varGamma }_n = A(d_n), n \ge 1.\) Let \(\{v_1, \ldots , v_d\}\) be an orthonormal basis of \(\mathbb {R}^d.\) Then, it is easy to see that
It is now obvious that condition (2.7) is equivalent to
Furthermore, \(\Vert I - \tilde{\varGamma }^2_n \Vert \) is equal to \(1 -\tilde{\lambda }^2_n,\) where \(\tilde{\lambda }_n\) is the smallest eigenvalue of \(\tilde{\varGamma }_n\) since \(\tilde{\varGamma }^2_n\) is symmetric and \(I -\tilde{\varGamma }^2_n \) is nonnegative definite. So it remains to be shown that (5.1) implies
or, equivalently, to show that if (5.6) does not hold, we cannot have (5.1).
To that end, we apply Lemma 5.1 with \(\delta =1{/}2\) and we get for \(y \in \mathbb {R},\)
Here we have used the monotonicity of the sequence \(\tilde{\varGamma }_k, k \ge 1\) which implies that \(\tilde{\varGamma }_n -\mathrm {Cov}( T'_{m_j}{/}\sqrt{m_j})\) is nonnegative definite for \(j_n \le j \le N.\) This allows us to conclude that for \(j_n \le j \le N,\)
Let \(D_n\) be the d-dimensional diagonal matrix with \(D_n(i,i)=1, 1 \le i \le d-1\) and \(D_n(d,d)=\tilde{\lambda }_n\). Then, clearly
and we can infer from Lemma 5.2 that
where \(Z'\) is a \((d-1)\)-dimensional normal mean zero random vector with covariance matrix equal to the identity matrix.
Using the fact that the square of the Euclidean norm of a d-dimensional \(\mathcal {N}(0,I)\)-distributed random vector X has a gamma distribution with parameters d / 2 and 2, one can show that there exist positive constants \(C_1(d), C_2(d)\) so that
(see Lemma 1 and Lemma 3 in [8], where more precise bounds are given if \(d \ge 3\). If \(d=1\), this follows directly from well-known bounds for the tail probabilities of the one-dimensional normal distribution. If \(d=2\), the random variable \(|X|^2\) has an exponential distribution and (5.9) is trivial).
We can conclude that for large enough n,
where we set \(C_3(d)= C_2(d-1){/}C_1(d).\) Returning to inequality (5.8) and noting that \(f_n(y-1{/}2) \ge \sqrt{\log \log n}\) if n is large, we get in this case,
Applying (5.9) once more, we find that
Recalling (5.7) we can conclude that if
we have for any \(y \in \mathbb {R},\)
Consequently, \(M_n\) cannot converge in distribution to any variable of the form \(\tilde{Y}+c\). \(\square \)
Remark
-
1.
Denote the distribution of \(a_n\max _{1 \le k \le n} |S_k|{/}\sqrt{k} -b_{d,n} \) by \(Q_n\). From (3.4) and (3.6), it follows that this sequence is tight if and only if the distributions of \(M_n\) form a tight sequence. The above argument actually shows that this last sequence cannot be tight if condition (2.7) is not satisfied. Moreover, it is not difficult to prove via Theorem 2.1 that (2.7) implies that the sequence \(\{Q_n{:}\,n \ge 1\}\) is tight. Thus, we have
$$\begin{aligned} \{Q_n{:}\,n \ge 1\} \hbox { is tight } \Longleftrightarrow (2.7). \end{aligned}$$ -
2.
Also note that
$$\begin{aligned} \mathbb {P}\left\{ a_n \max _{1 \le k \le n} |\varGamma _k^{-1}S_k|{/}\sqrt{k} - b_n \le 0\right\} \le \mathbb {P}\left\{ a_n \varLambda _n^{-1}\max _{1 \le k \le n} |S_k|{/}\sqrt{k} - b_n \le 0\right\} , \end{aligned}$$where \(\varLambda _n\) is the largest eigenvalue of \(\varGamma _n\) which in turn is defined as in (2.3) [here we can choose any sequence \(c_n\) satisfying condition (2.2)]. Using this inequality, one can show by the same argument as on p. 255 in [6] that (2.6) implies
$$\begin{aligned} 1 - \varLambda ^2_n = o \left( ( LL n)^{-1}\right) \hbox { as }n \rightarrow \infty . \end{aligned}$$This is of course weaker than (2.5) if \(d \ge 2.\)
6 Some Further Results
We first prove the following Darling–Erdős type theorem with a shifted limiting distribution.
Theorem 6.1
Let \(X, X_n, n \ge 1\) be i.i.d. real-valued random variables with \(\mathbb {E}X^2 =1\) and \(\mathbb {E}X=0.\) Assume that for some \(c>0,\)
Then, we have,
where \(\tilde{Y}\) and \(b_n\) are defined as in (1.1).
Proof
-
(i)
Set \(\sigma _n^2 =\mathbb {E}X^2I\{|X| \le \sqrt{n}\}\) and let \(1 \le k_n \le \exp ((Ln)^{\alpha })\) for some \(0< \alpha <1.\) Then, we have by Theorem 2.1 and the argument in (4.1),
$$\begin{aligned} a_n \max _{k_n \le k \le n} |S_k|{{/}}\sqrt{k}\sigma _k - b_{n} \mathop {\rightarrow }\limits ^{d} \tilde{Y}, \end{aligned}$$which trivially implies for any sequence \( \rho _n\) of positive real numbers converging to 1,
$$\begin{aligned} \rho _n a_n \max _{k_n \le k \le n} |S_k|{/}\sqrt{k}\sigma _k - \rho _n b_{n} \mathop {\rightarrow }\limits ^{d} \tilde{Y} \end{aligned}$$(6.1)Set \(k_n = [\exp ((Ln)^{\alpha }]\), where \(0< \alpha <1.\) Then, it is easy to see that
$$\begin{aligned}&\mathbb {P}\left\{ a_n \max _{1 \le k \le n} \frac{|S_k|}{\sqrt{k}} - b_n \le y\right\} \\&\quad \le \,\mathbb {P}\left\{ \sigma _{k_n} a_n \max _{k_n \le k \le n} \frac{|S_k|}{\sqrt{k}\sigma _k} - b_n \le y\right\} \\&\quad =\mathbb {P}\left\{ \sigma _{k_n} a_n \max _{k_n \le k \le n} \frac{|S_k|}{\sqrt{k}\sigma _k} - \sigma _{k_n} b_n \le y +(1-\sigma _{k_n})b_n\right\} \end{aligned}$$Noticing that \((1-\sigma _{k_n})b_n \sim (1 -\sigma ^2_{k_n})(2 LL n){/}(1+\sigma _{k_n})\sim c( LL k_n)^{-1} LL n\) (since \(\sigma ^2_{k_n} \rightarrow \mathbb {E}X^2=1\)), it is clear that \((1-\sigma _{k_n})b_n \rightarrow c{/}\alpha \) as \(n \rightarrow \infty .\) By (6.1) (with \(\rho _n=\sigma _{k_n}\)), this last sequence of probabilities converges to \(\mathbb {P}\{\tilde{Y} \le y +c{/}\alpha \}.\) Since this holds for any \(0< \alpha <1,\) it follows that
$$\begin{aligned} \limsup _{n \rightarrow \infty } \mathbb {P}\left\{ a_n \max _{1 \le k \le n} \frac{|S_k|}{\sqrt{k}} - b_n \le y\right\} \le \mathbb {P}\{\tilde{Y} \le y +c\}. \end{aligned}$$(6.2) -
(ii)
Similarly, we have,
$$\begin{aligned}&\mathbb {P}\left\{ a_n \max _{1 \le k \le n} \frac{|S_k|}{\sqrt{k}} - b_n \le y\right\} \\&\quad \ge \mathbb {P}\left\{ \sigma _{n} a_n \max _{1 \le k \le n} \frac{|S_k|}{\sqrt{k}\sigma _k} - \sigma _{n} b_n \le y +(1-\sigma _{n})b_n\right\} , \end{aligned}$$where \((1-\sigma _n)b_n \rightarrow c\) as \(n \rightarrow \infty .\) Applying (6.1) (with \(k_n=1\) and \(\rho _n=\sigma _n\)), we obtain that
$$\begin{aligned} \liminf _{n \rightarrow \infty } \mathbb {P}\left\{ a_n \max _{1 \le k \le n} \frac{|S_k|}{\sqrt{k}} - b_n \le y\right\} \ge \mathbb {P}\{\tilde{Y} \le y +c\} \end{aligned}$$and Theorem 6.1 has been proven. \(\square \)
We finally mention the following result for real-valued random variables given in [13] where it is shown that if \(\mathbb {E}X=0, \mathbb {E}X^2 =1\) and \(\mathbb {E}X^2 LL|X| < \infty \), then one has
The authors asked whether this result can hold under the finite second moment assumption.
Using Theorem 2.3 in combination with Theorem 1.1 in [13], we obtain the following general result:
where \(\sigma _n^2 = \mathbb {E}X^2 I\{|X| \le c_n\}\) and \(c_n\) is a non-decreasing sequence of positive real numbers satisfying condition (2.2). As in [6], this implies that (6.3) holds if and only if condition (2.5) is satisfied.
References
Bai, Z.D.: A theorem of Feller revisited. Ann. Probab. 17, 385–395 (1989)
Bertoin, J.: Darling–Erdős theorems for normalized sums of i.i.d. variables close to a stable law. Ann. Probab. 26, 832–852 (1998)
Bhatia, R.: Matrix Analysis. Springer, New York (1997)
Csörgő, M., Szyskowicz, B., Wang, Q.: Darling–Erdős theorem for self-normalized sums. Ann. Probab. 31, 676–692 (2003)
Darling, D.A., Erdős, P.: A limit theorem for the maximum of normalized sums of independent random variables. Duke Math. J. 23, 143–155 (1956)
Einmahl, U.: The Darling–Erdős theorem for sums of i.i.d. random variables. Probab. Theory Relat. Fields 82, 241–257 (1989)
Einmahl, U.: On the almost sure behavior of sums of iid random variables in Hilbert space. Ann. Probab. 19, 1227–1263 (1991)
Einmahl, U.: Exact convergence rates for the bounded law of the iterated logarithm in Hilbert space. Probab. Theory Relat. Fields 92, 177–194 (1992)
Einmahl, U.: A new strong invariance principle for sums of independent random vectors. J. Math. Sci. 163, 311–327 (2009)
Einmahl, U., Mason, D.M.: Darling–Erdős theorems for martingales. J. Theor. Probab. 2, 437–460 (1989)
Feller, W.: The law of the iterated logarithm for identically distributed random variables. Ann. Math. 47, 631–638 (1946)
Horváth, L.: Likelihood method for testing changes in the parameters of normal observations. Ann. Stat. 21(2), 671–680 (1994)
Khoshnevisan, D., Levin, D.A., Shi, Z.: An extreme-value analysis of the LIL for Brownian motion. Electron. Commun. Probab. 10, 196–206 (2005)
Oodaira, H.: Some limit theorems for the maximum of normalized sums of weakly dependent random variables. In: Proceedings of the Third Japan-USSR Symposium on Probability Theory, Lecture Notes in Mathematics, vol. 550, pp. 1–13. Springer, Berlin (1976)
Sakhanenko, A.I.: A new way to obtain estimates in the invariance principle. In: Proceedings of High Dimensional Probability II, Birkhäuser, Progress in Probability, vol. 47, pp. 223–245 (2000)
Shorack, G.: Extension of the Darling and Erdős theorem on the maximum of normalized sums. Ann. Probab. 7, 1092–1096 (1979)
Acknowledgements
The authors would like to thank the referee for some useful suggestions helping us to improve the presentation of our results.
Author information
Authors and Affiliations
Corresponding author
Additional information
G. Dierickx was supported by a Ph.D.-Fellowship of the Research Foundation Flanders-FWO.
Rights and permissions
About this article
Cite this article
Dierickx, G., Einmahl, U. A General Darling–Erdős Theorem in Euclidean Space. J Theor Probab 31, 1142–1165 (2018). https://doi.org/10.1007/s10959-016-0728-y
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10959-016-0728-y
Keywords
- Darling–Erdős theorem
- Extreme value distribution
- Hartman–Wintner LIL
- Integral test
- Strong invariance principle
- Multidimensional version
- Double truncation