Introduction
The law of the iterated logarithm (LIL) is one of the most important and precise limit theorems in probability theory. The strong law of large numbers establishes the almost sure convergence of normalized sums and the central limit theorem describes their limiting distribution after rescaling. The LIL tells us how large these fluctuations can be infinitely often. To see the basic idea, let S n = ∑ k = 1 n X k S_n = \sum_{k=1}^{n} X_k S n = ∑ k = 1 n X k , where ( X k ) (X_k) ( X k ) is a sequence of centered independent random variables. In the classical finite-variance case, the LIL uses the scale 2 s n 2 log log s n 2 \sqrt{2s_n^2 \log\log s_n^2} 2 s n 2 log log s n 2 , where s n 2 = ∑ k = 1 n E ( X k 2 ) s_n^2 = \sum_{k=1}^{n} \mathbb{E}(X_k^2) s n 2 = ∑ k = 1 n E ( X k 2 ) . For independent identically distributed (i.i.d) random variables with variance one, this becomes 2 n log log n \sqrt{2n\log\log n} 2 n log log n . The term log log n \log\log n log log n is the main feature of the theorem.
The history of the LIL goes back to the early work of Khintchine and Kolmogorov [12 ] . Kolmogorov proved an LIL for independent random variables under suitable conditions on the individual summands. Hartman and Wintner later obtained the classical finite-variance result for i.i.d random variables [9 ] . Their work showed that, in the i.i.d. setting, boundedness of the random variables is not needed. The LIL was then extended in several directions. Stout developed martingale versions of the LIL [20 ] . This was an important step. Martingales are not independent, but they still have a useful conditional mean-zero structure. Hence, martingale LILs provide a natural way to study dependent random sequences.
Another important direction is the study of weighted sums. Instead of the ordinary partial sum S n = ∑ k = 1 n X k S_n = \sum_{k=1}^{n} X_k S n = ∑ k = 1 n X k , one considers S n ( f ) = ∑ k = 1 n f ( k / n ) X k S_n(f) = \sum_{k=1}^{n} f(k/n)X_k S n ( f ) = ∑ k = 1 n f ( k / n ) X k , where f f f is a function on [ 0 , 1 ] [0, 1] [ 0 , 1 ] . The case f ≡ 1 f \equiv1 f ≡ 1 gives the ordinary partial sum. Thus, weighted sums contain the classical LIL as a special case.
Weighted LILs have a long history. Gaposhkin studied LILs connected with Cesàro and Abel summation methods [Gap65]. Tomkins studied weighted averages of independent random variables [22 ] . For example, for suitable functions f f f , Tomkins proved in [23 ] that
lim sup n → ∞ 1 2 n log log n ∑ k = 1 n f ( k / n ) X k = ( ∫ 0 1 f ( t ) 2 d t ) 1 / 2 a.s. \limsup_{n\to\infty}\frac{1}{\sqrt{2n\log\log n}}\sum_{k=1}^{n}f(k/n)X_k=\left(\int_{0}^{1}f(t)^2\,dt\right)^{1/2}\quad\text{a.s.} n → ∞ lim sup 2 n log log n 1 k = 1 ∑ n f ( k / n ) X k = ( ∫ 0 1 f ( t ) 2 d t ) 1/2 a.s.
when the random variables are centered and have variance one. He later extended this result to martingale difference sequences [24 ] . Li and Tomkins later obtained further results for weighted sums of i.i.d. random variables and also studied Banach-space versions [16 ] . Their work shows that the classical LIL also holds for a broad class of regular weights.
This leads to a natural question: What happens if the weights are allowed to be much more general? This question is important because the weights in f ( k / n ) f(k/n) f ( k / n ) have a special structure. A general triangular array does not have this property. Let A = ( a n , k ) A=(a_{n,k}) A = ( a n , k ) , 1 ≤ k ≤ n 1\leq k\leq n 1 ≤ k ≤ n , and consider S n ( A ) = ∑ k = 1 n a n , k X k S_n(A)=\sum_{k=1}^{n}a_{n,k}X_k S n ( A ) = ∑ k = 1 n a n , k X k . The coefficients may now change in an arbitrary way from one row to the next. Thus, S n ( A ) S_n(A) S n ( A ) can behave differently from ordinary partial sums. This difference is serious. The classical LIL need not hold for bounded triangular arrays. Li and Rao constructed a bounded triangular array for which
lim sup n → ∞ S n ( A ) 2 n log log n = ∞ a.s. , \limsup_{n\to\infty}\frac{S_n(A)}{\sqrt{2n\log\log n}}=\infty\quad\text{a.s.}, n → ∞ lim sup 2 n log log n S n ( A ) = ∞ a.s. ,
showing that the classical LIL fails [14 ] . Thus, even bounded weights can produce fluctuations much larger than the classical LIL scale. This shows that boundedness of the weights alone is not enough to preserve the log log n \log\log n log log n scale.
This observation changes the question. For general triangular arrays, we must first ask whether the LIL is even the correct scale. Li, Rao, and Wang first studied this problem for weighted sums with multidimensional indices [15 ] . In the one-dimensional case, they assumed E ( X 2 ) < ∞ \mathbb{E}(X^2)<\infty E ( X 2 ) < ∞ , and considered a bounded triangular array of weights K : = sup n ≥ 1 max 1 ≤ k ≤ n ∣ a n , k ∣ < ∞ K:=\sup_{n\geq1}\max_{1\leq k\leq n}|a_{n,k}|<\infty K := sup n ≥ 1 max 1 ≤ k ≤ n ∣ a n , k ∣ < ∞ . They showed that the natural scale is 2 n log n \sqrt{2n\log n} 2 n log n and proved
lim sup n → ∞ ∣ S n ( A ) ∣ 2 n log n ≤ 12 K ( E X 2 ) 1 / 2 a.s. (1) \limsup_{n\to\infty}\frac{|S_n(A)|}{\sqrt{2n\log n}}\leq12K\left(\mathbb{E}X^2\right)^{1/2}\quad\text{a.s.} \tag*{(1)} n → ∞ lim sup 2 n log n ∣ S n ( A ) ∣ ≤ 12 K ( E X 2 ) 1/2 a.s. (1)
They call (1) the law of logarithms. Later, Chen and Hao [CH11] obtained the sharper upper bound K E X 2 K\sqrt{\mathbb{E}X^2} K E X 2 under the same finite second-moment condition.
Li and Tomkins [17 ] generalize (1) . They replaced the specific 2 n log n \sqrt{2n\log n} 2 n log n normalization by more general b n = B ( n ) b_n=B(n) b n = B ( n ) , where B B B is regularly varying with index α ∈ [ 1 / 2 , 1 ) \alpha\in[1/2,1) α ∈ [ 1/2 , 1 ) , without assuming E ( X 2 ) < ∞ \mathbb{E}(X^2)<\infty E ( X 2 ) < ∞ . When α > 1 / 2 \alpha>1/2 α > 1/2 , they proved that the normalized weighted sums converge to zero, while at the critical value α = 1 / 2 \alpha=1/2 α = 1/2 they obtained a non-trivial upper bound:
lim sup n → ∞ ∣ S n ( A ) ∣ b n ≤ 6 2 ℓ sup 1 ≤ k ≤ n ∣ a n , k ∣ a.s. (2) \limsup_{n\to\infty}\frac{|S_n(A)|}{b_n}\leq6\sqrt{2}\,\ell\sup_{1\leq k\leq n}|a_{n,k}|\quad\text{a.s.} \tag*{(2)} n → ∞ lim sup b n ∣ S n ( A ) ∣ ≤ 6 2 ℓ 1 ≤ k ≤ n sup ∣ a n , k ∣ a.s. (2)
under a suitable truncated second-moment condition. Its non-commutative counterpart is formulated in Condition (3) of Theorem B.
Over the past several decades, non-commutative probability theory has developed into a rich generalization of the classical theory. In this framework, the underlying probability space is replaced by a finite von Neumann algebra ( M , τ ) (\mathcal{M},\tau) ( M , τ ) with a faithful normal tracial state, and random variables by τ \tau τ -measurable operators. The non-commutative setting is not a simple change of notation. Operators do not commute. Hence many pointwise arguments from classical probability are no longer available.
Non-commutative martingale convergence and ergodic theorems began with the original work of Cuculescu [Cuc71]. Junge and Xu [10 , 11 ] further developed these concepts by focusing on martingale and maximal inequalities. However, progress on the non-commutative LIL has been more recent. Konwerska [13 ] obtained a result similar to the Hartman-Wintner theorem for independent self-adjoint operators. Zeng proved a non-commutative analogue of Stout’s martingale LIL [Zen15]. His result gave an upper bound with constant 2. A sharp version was obtained in our companion paper [19 ] . There we proved optimal non-commutative analogues of the classical LIL for martingales and for independent random variables.
The goal of this paper is to develop the weighted versions of these results in a non-commutative framework. Here, ( M , τ ) (\mathcal{M},\tau) ( M , τ ) represents a finite von Neumann algebra with an increasing sequence ( M n ) (\mathcal{M}_n) ( M n ) of von Neumann subalgebras. For t > e t>e t > e , set L ( t ) : = max { 1 , log log t } L(t):=\max\{1,\log\log t\} L ( t ) := max { 1 , log log t } . Our first main result presents the following non-commutative version of Tomkins’ theorem [24 ] .
Theorem A. Let ( M , τ ) (\mathcal{M},\tau) ( M , τ ) be a finite von Neumann algebra and ( d n ) (d_n) ( d n ) be a sequence of self-adjoint martingale differences with τ ( d 1 ) = 0 \tau(d_1)=0 τ ( d 1 ) = 0 , E n − 1 ( d n 2 ) = 1 \mathbb{E}_{n-1}(d_n^2)=1 E n − 1 ( d n 2 ) = 1 and max m ≤ n ∥ d m ∥ ≤ M n \max_{m\le n}\|d_m\|\le M_n max m ≤ n ∥ d m ∥ ≤ M n , where ( M n ) (M_n) ( M n ) is a sequence of positive reals such that lim n → ∞ L ( n ) M n 2 n = 0 \lim_{n\to\infty}\frac{L(n)M_n^2}{n}=0 lim n → ∞ n L ( n ) M n 2 = 0 . Let p p p be a polynomial on [ 0 , 1 ] [0,1] [ 0 , 1 ] , then
lim sup n → ∞ 1 2 n L ( n ) ∑ m = 1 n p ( m / n ) d m ≤ ∥ p ∥ 2 a.u. (3) \limsup_{n\to\infty}\frac{1}{\sqrt{2nL(n)}}\sum_{m=1}^{n}p(m/n)d_m\leq\|p\|_2 \quad\text{a.u.} \tag*{(3)} n → ∞ lim sup 2 n L ( n ) 1 m = 1 ∑ n p ( m / n ) d m ≤ ∥ p ∥ 2 a.u. (3)
When p ≡ 1 p\equiv1 p ≡ 1 , this gives the non-commutative LIL from [19 ] . Thus, the present theorem extends the earlier unweighted result. The first difficulty is the triangular form of the weights. The coefficient p ( m / n ) p(m/n) p ( m / n ) depends on n n n . Therefore, the weighted sum is not an ordinary martingale sum. A direct use of the unweighted LIL is not enough. In the classical theory, Abel summation is a useful tool for weighted sums. In the non-commutative setting, however, the same argument cannot be used directly. We need estimates that are compatible with almost uniform convergence and with the lack of commutativity.
Our proof starts with monomial weights p ( t ) = t j p(t)=t^j p ( t ) = t j . The corresponding sums are ∑ m = 1 n m j d m \sum_{m=1}^{n}m^jd_m ∑ m = 1 n m j d m . We then combine these estimates for finitely many monomials to treat polynomial weights. A blocking argument is used throughout this step. We divide the indices into geometric blocks. On one block, the relevant normalizing quantities are comparable. This allows the triangular weights to be compared with fixed coefficients. We then use non-commutative maximal estimates and a non-commutative Borel-Cantelli argument. After the polynomial case, we pass to a suitable class of functions of bounded variation by approximation. We also obtain the result for the power series considered in the paper. Thus, the polynomial case is the basic step, while approximation gives the larger class of weights.
The second part of the paper asks a different question: What happens for general bounded deterministic weights? This is the natural non-commutative counterpart of the class studied by Li and Tomkins [17 ] .
Theorem B. Let ( x k ) k ≥ 1 (x_k)_{k\ge1} ( x k ) k ≥ 1 be a i.i.d sequence of mean zero self-adjoint random variables on ( M , τ ) (\mathcal{M},\tau) ( M , τ ) . Let A = ( a n , k ) ∈ D A=(a_{n,k})\in\mathcal{D} A = ( a n , k ) ∈ D with K = sup n , k ∣ a n , k ∣ K=\sup_{n,k}|a_{n,k}| K = sup n , k ∣ a n , k ∣ , and let B B B , ( b n ) (b_n) ( b n ) be as above with index α ∈ [ 1 2 , 1 ) \alpha\in[\frac{1}{2},1) α ∈ [ 2 1 , 1 ) . Assume that,
(1) τ ( B − 1 ( ∣ x 1 ∣ ) ) < ∞ \tau\left(B^{-1}(|x_1|)\right)<\infty τ ( B − 1 ( ∣ x 1 ∣ ) ) < ∞
(2) B − 1 ( t log t ) log t G ( t ) ⟶ 0 \frac{B^{-1}(t\log t)}{\log t}G(t)\longrightarrow0 l o g t B − 1 ( t l o g t ) G ( t ) ⟶ 0 as t → ∞ t\to\infty t → ∞ ,
(3) if α = 1 2 \alpha=\frac{1}{2} α = 2 1 , then there exists a constant 0 ≤ ℓ < ∞ 0\le\ell<\infty 0 ≤ ℓ < ∞ such that
lim sup n → ∞ n log n b n 2 H ( b n log n ) = ℓ 2 . \limsup_{n\to\infty}\frac{n\log n}{b_n^2}H\left(\frac{b_n}{\log n}\right)=\ell^2. n → ∞ lim sup b n 2 n log n H ( log n b n ) = ℓ 2 . Then:
(a) If 1 2 < α < 1 \frac{1}{2}<\alpha<1 2 1 < α < 1 and x 1 ∈ L 2 ( M , τ ) x_1\in L_2(\mathcal{M},\tau) x 1 ∈ L 2 ( M , τ ) ,
S n ( A ) b n ⟶ 0 , almost uniformly , where S n ( A ) : = ∑ k = 1 n a n , k x k . \frac{S_n(A)}{b_n}\longrightarrow0,\quad\text{almost uniformly},\quad\text{where }S_n(A):=\sum_{k=1}^{n}a_{n,k}x_k. b n S n ( A ) ⟶ 0 , almost uniformly , where S n ( A ) := k = 1 ∑ n a n , k x k . (b) If α = 1 2 \alpha=\frac{1}{2} α = 2 1 ,
lim sup n → ∞ S n ( A ) b n ≤ 2 K ℓ b.a.u. \limsup_{n\to\infty}\frac{S_n(A)}{b_n}\le\sqrt{2}K\ell\quad\text{b.a.u.} n → ∞ lim sup b n S n ( A ) ≤ 2 K ℓ b.a.u.
Thus, our result extends the general framework of Li and Tomkins to the non-commutative setting. It is not limited to the special case 2 n log n \sqrt{2n\log n} 2 n log n . The constant 2 \sqrt{2} 2 is the key point of this result. Under the corresponding assumptions, our bound improves the constant 6 2 6\sqrt{2} 6 2 in eq. (2) to 2 \sqrt{2} 2 . This improvement is also important in the classical setting. In the finite-variance logarithmic case, take B ( n ) = 2 n log n B(n)=\sqrt{2n\log n} B ( n ) = 2 n log n . Then the truncated second-moment parameter satisfies ℓ 2 = 1 2 E ( X 2 ) \ell^2=\frac{1}{2}\mathbb{E}(X^2) ℓ 2 = 2 1 E ( X 2 ) . Hence, part (b) of Theorem B reduces to
lim sup n → ∞ ∣ S n ( A ) ∣ 2 n log n ≤ K ( E X 2 ) 1 / 2 a.s. \limsup_{n\to\infty}\frac{|S_n(A)|}{\sqrt{2n\log n}}\le K\left(\mathbb{E}X^2\right)^{1/2}\quad\text{a.s.} n → ∞ lim sup 2 n log n ∣ S n ( A ) ∣ ≤ K ( E X 2 ) 1/2 a.s.
Thus, in the classical finite-variance case, our bound agrees with the sharp bound of Chen and Hao [CH11]. At the same time, our theorem keeps the more general normalization b n = B ( n ) b_n=B(n) b n = B ( n ) . In the critical case α = 1 / 2 \alpha=1/2 α = 1/2 , no finite second moment is required. For α ∈ ( 1 / 2 , 1 ) \alpha\in(1/2,1) α ∈ ( 1/2 , 1 ) , the finite second moment assumption can also be replaced by a weaker condition on the truncated second moment; see Remark 4.14. In this sense, the result extends the sharp classical bound to the non-commutative setting and to a much larger class of distributions.
One main difficulty in proving this result is that the usual non-commutative LIL method cannot be used here. In the earlier works of Konwerska and Zeng, and also in our companion paper, the argument is built around a martingale. This makes it possible to control the supremum over a block through the column tail probability Prob c \operatorname{Prob}_{c} Prob c , and then use non-commutative Doob and Chebyshev inequalities together with a Borel-Cantelli argument [13 Zen15, 19 ] .
In the present weighted problem, however, S n ( A ) = ∑ k = 1 n a n , k x k S_n(A) = \sum_{k=1}^{n} a_{n,k}x_k S n ( A ) = ∑ k = 1 n a n , k x k does not form a martingale, since the coefficients change with n n n . Thus, this standard maximal-inequality argument for partial sums cannot be applied directly in this setting. We overcome this problem by working directly with the definitions of almost uniform and bilaterally almost uniform boundedness. We construct suitable projections for the different parts of the sum and control their complements through trace estimates. This gives the required a.u. and b.a.u. bounds without embedding the whole weighted sequence into a single martingale.
The assumptions are motivated by the classical law of the logarithm. They involve the tail of x 1 x_1 x 1 and its truncated second moment. These conditions help us control both the large values and the main fluctuations of the sum. The critical case α = 1 / 2 \alpha= 1/2 α = 1/2 , is more delicate. Here we use a truncation of the random operators. We split the sum into a main part and tail parts. The main part is controlled by the non-commutative exponential inequality in Theorem 3.1 , as carried out in Lemma 4.17 . The large part is controlled by tail estimates and a non-commutative Borel-Cantelli argument (cf. Lemma 4.15).
A second important tool is dyadic blocking. We consider blocks 2 k − 1 < n ≤ 2 k 2^{k-1} < n \le2^k 2 k − 1 < n ≤ 2 k . Regular variation gives a comparison between b n b_n b n and B ( 2 k ) B(2^k) B ( 2 k ) on each block. This allows us to replace the original problem by a sequence of block problems.
For the intermediate part of the truncated sum, we use an argument inspired by Cuzick’s method for weighted sums [Cuz95]. This part is more delicate in the non-commutative setting. The intermediate terms are neither going to 0 in a.u. nor uniformly bounded by constant, and the truncation level of these terms changes with n n n . Moreover, the scalar centering and Borel-Cantelli argument used by Cuzick cannot be applied directly to non-commuting variables. We overcome this by working on dyadic blocks and writing the intermediate sum as a centered finite martingale part plus a deterministic mean part; see Lemma 4.20 . The centered part is controlled by a non-commutative exponential inequality and a non-commutative analogue of Cuzick’s lemma, (cf. Lemma 4.19 ).
A further difficulty appears since index α = 1 / 2 \alpha= 1/2 α = 1/2 , where the natural regularly varying integrand has index − 1 -1 − 1 and we are unable to use Karamata’s theorem [BGT89]. We avoid this critical case by a suitable comparison that changes the relevant index from − 1 -1 − 1 to − 3 / 2 -3/2 − 3/2 ; see Lemma 4.18 . This gives an almost-uniform bound for the intermediate part. By choosing the truncation parameter suitably, this bound can be made arbitrarily small; see Lemma 4.20 .
For α ∈ ( 1 / 2 , 1 ) \alpha\in(1/2,1) α ∈ ( 1/2 , 1 ) , we assume that x 1 ∈ L 2 ( M , τ ) x_1 \in L_2(\mathcal{M},\tau) x 1 ∈ L 2 ( M , τ ) . This assumption is not needed in the classical result of Li and Tomkins [17 ] . In the non-commutative setting, however, using the decomposition in eq. (42) , we can conclude bilaterally almost uniform convergence to zero without assuming the L 2 L_2 L 2 condition. On the other hand, the L 2 L_2 L 2 assumption enables us to establish almost uniform convergence.
The paper is organized as follows. In § 2 , we collect the basic tools used in the proofs. We recall almost uniform convergence and boundedness, Doob’s maximal inequality, and the non-commutative Chebyshev and Borel-Cantelli lemmas. In § 3 , we prove the weighted LIL for martingale difference sequences. We first treat monomial weights. We then pass to polynomial weights by a blocking argument. Finally, we extend the result to suitable functions of bounded variation and power series by approximation. The self-adjointness assumption is removed by the standard 2 × 2 2 \times2 2 × 2 matrix argument. In § 4 , we study general bounded deterministic triangular arrays. We use regular variation, truncation, exponential estimates, and dyadic blocking. The case α > 1 / 2 \alpha> 1/2 α > 1/2 and the critical case α = 1 / 2 \alpha= 1/2 α = 1/2 are treated separately.
Preliminaries
In this section, we briefly discuss all the objects that are necessary for this article. Throughout this paper, unless otherwise mentioned, we fix a finite von Neumann algebra M ⊆ B ( H ) \mathcal{M} \subseteq\mathcal{B}(\mathcal{H}) M ⊆ B ( H ) equipped with a faithful, normal tracial state τ \tau τ . We will denote the ∗ * ∗ -algebra of closed, densely defined operators on H \mathcal{H} H , which are affiliated to M \mathcal{M} M , by M ~ \widetilde{\mathcal{M}} M . These are considered as the non-commutative analogue of the classical measurable functions on a probability space. By the spectral theorem of the unbounded operators, the tracial state τ \tau τ can be seen to be extended to the positive part of M \mathcal{M} M . Using this for every 1 ≤ p ≤ ∞ 1 \leq p \leq\infty 1 ≤ p ≤ ∞ , one can define the non-commutative L p L_p L p -spaces equipped with the norm ∥ x ∥ p : = τ ( ∣ x ∣ p ) 1 / p \lVert x \rVert_p := \tau(|x|^p)^{1/p} ∥ x ∥ p := τ ( ∣ x ∣ p ) 1/ p , where we set L ∞ ( M ) : = M L_\infty(\mathcal{M}) := \mathcal{M} L ∞ ( M ) := M . For every a ∈ M ~ a \in\widetilde{\mathcal{M}} a ∈ M , let ∣ a ∣ = ∫ 0 ∞ λ d e λ |a| = \int_0^\infty\lambda\, d e_\lambda ∣ a ∣ = ∫ 0 ∞ λ d e λ be the spectral decomposition. Then as in the classical setup, one can define an analogue of the decreasing rearrangement, μ ( ⋅ ) : [ 0 , ∞ ) → [ 0 , ∞ ) \mu(\cdot): [0,\infty) \to[0,\infty) μ ( ⋅ ) : [ 0 , ∞ ) → [ 0 , ∞ ) by
μ t ( a ) : = inf { s ≥ 0 : τ ( e ( s , ∞ ) ( ∣ a ∣ ) ) ≤ t } . \mu_t(a) := \inf\{s \geq0 : \tau(e_{(s,\infty)}(|a|)) \leq t\}. μ t ( a ) := inf { s ≥ 0 : τ ( e ( s , ∞ ) ( ∣ a ∣ )) ≤ t } .
It is known that μ ⋅ \mu_\cdot μ ⋅ is decreasing, continuous from right and a ∈ L p ( M ) a \in L_p(\mathcal{M}) a ∈ L p ( M ) if and only if μ ( ⋅ ) ( a ) ∈ L p ( ( 0 , ∞ ) ) \mu(\cdot)(a) \in L_p((0,\infty)) μ ( ⋅ ) ( a ) ∈ L p (( 0 , ∞ )) for all a ∈ M ~ a \in\widetilde{\mathcal{M}} a ∈ M . For detailed exhibition on this topic, we refer to [18 , 21 FK86, Hia21] . We will use P ( M ) \mathcal{P}(\mathcal{M}) P ( M ) to denote the lattice of projections of the von Neumann algebra M \mathcal{M} M .
Definition 2.1. Let ( x n ) n ∈ N (x_n)_{n \in\mathbb{N}} ( x n ) n ∈ N be a sequence in M ~ \widetilde{\mathcal{M}} M . For K ≥ 0 K \geq0 K ≥ 0 , we say that ( x n ) n (x_n)_n ( x n ) n is almost uniformly (a.u) bounded by K K K if, for every ε > 0 \varepsilon> 0 ε > 0 , there exists a projection e ∈ M e \in\mathcal{M} e ∈ M with τ ( 1 − e ) < ε \tau(1-e) < \varepsilon τ ( 1 − e ) < ε such that, for every δ > 0 \delta> 0 δ > 0 ,
lim sup n → ∞ ∥ x n e ∥ ∞ ≤ K + δ . \limsup_{n \to\infty} \lVert x_n e \rVert_\infty\leq K + \delta. n → ∞ lim sup ∥ x n e ∥ ∞ ≤ K + δ . We say that ( x n ) n (x_n)_n ( x n ) n is bilaterally almost uniformly (b.a.u) bounded by K K K if lim sup n → ∞ ∥ e x n e ∥ ∞ ≤ K + δ \limsup_{n \to\infty} \lVert e x_n e \rVert_\infty\leq K + \delta lim sup n → ∞ ∥ e x n e ∥ ∞ ≤ K + δ . We write lim sup n → ∞ x n ≤ a . u . K \limsup_{n \to\infty} x_n \underset{\mathrm{a.u.}}{\leq} K lim sup n → ∞ x n a.u. ≤ K for a.u boundedness and lim sup n → ∞ x n ≤ b . a . u . K \limsup_{n \to\infty} x_n \underset{\mathrm{b.a.u.}}{\leq} K lim sup n → ∞ x n b.a.u. ≤ K for b.a.u boundedness.
Definition 2.2. Let ( x n ) n ∈ N (x_n)_{n \in\mathbb{N}} ( x n ) n ∈ N be a sequence in M ~ \widetilde{\mathcal{M}} M and x ∈ M ~ x \in\widetilde{\mathcal{M}} x ∈ M . If for any ε > 0 \varepsilon> 0 ε > 0 there exists a projection e ∈ P ( M ) e \in\mathcal{P}(\mathcal{M}) e ∈ P ( M ) with τ ( 1 − e ) < ε \tau(1-e) < \varepsilon τ ( 1 − e ) < ε such that
(a) ∥ ( x n − x ) e ∥ ∞ ⟶ 0 \lVert(x_n-x)e\rVert_\infty\longrightarrow0 ∥( x n − x ) e ∥ ∞ ⟶ 0 as n → ∞ n \to\infty n → ∞ , we say that x n x_n x n converges to x x x almost uniformly.
(b) ∥ e ( x n − x ) e ∥ ∞ ⟶ 0 \lVert e(x_n-x)e\rVert_\infty\longrightarrow0 ∥ e ( x n − x ) e ∥ ∞ ⟶ 0 as n → ∞ n \to\infty n → ∞ , we say that x n x_n x n converges to x x x bilaterally almost uniformly.
We also recall the ℓ ∞ \ell_{\infty} ℓ ∞ -column valued non-commutative L p L_p L p -space.
L p ( M ; ℓ ∞ c ) : = { x = ( x n ) ⊂ L p ( M , τ ) : there exists a ∈ L 2 p ( M , τ ) and ( y n ) ⊂ M such that x n = y n a } . L_p(\mathcal{M};\ell_{\infty}^{c}) := \{x=(x_n)\subset L_p(\mathcal{M},\tau):\text{ there exists }a\in L_{2p}(\mathcal{M},\tau)\text{ and }(y_n)\subset\mathcal{M}\text{ such that }x_n=y_na\}. L p ( M ; ℓ ∞ c ) := { x = ( x n ) ⊂ L p ( M , τ ) : there exists a ∈ L 2 p ( M , τ ) and ( y n ) ⊂ M such that x n = y n a } .
When p ≥ 2 p\ge2 p ≥ 2 , L p ( M ; ℓ ∞ c ) L_p(\mathcal{M};\ell_{\infty}^{c}) L p ( M ; ℓ ∞ c ) becomes a Banach space under the norm defined by ∥ x ∥ L p ( M ; ℓ ∞ c ) : = inf { ( sup n ∈ N ∥ y n ∥ ∞ ) ∥ a ∥ 2 p } \|x\|_{L_p(\mathcal{M};\ell_{\infty}^{c})}:=\inf\{(\sup_{n\in\mathbb{N}}\|y_n\|_{\infty})\|a\|_{2p}\} ∥ x ∥ L p ( M ; ℓ ∞ c ) := inf {( sup n ∈ N ∥ y n ∥ ∞ ) ∥ a ∥ 2 p } , x ∈ L p ( M ; ℓ ∞ c ) x\in L_p(\mathcal{M};\ell_{\infty}^{c}) x ∈ L p ( M ; ℓ ∞ c ) , where the infimum is taken over all such factorization of the element x x x . Then we have the following asymmetric Doob’s maximal inequality,
Theorem 2.3. [10 ] Let 4 ≤ p ≤ ∞ 4\le p\le\infty 4 ≤ p ≤ ∞ . For every L p ( M ) L_p(\mathcal{M}) L p ( M ) -bounded martingale x = ( x n ) n ∈ N x=(x_n)_{n\in\mathbb{N}} x = ( x n ) n ∈ N :
∥ ( x n ) n ∈ N ∥ L p ( ℓ ∞ c ) ≤ 2 2 / p sup n ∈ N ∥ x n ∥ p . \|(x_n)_{n\in\mathbb{N}}\|_{L_p(\ell_{\infty}^{c})}\le2^{2/p}\sup_{n\in\mathbb{N}}\|x_n\|_p. ∥ ( x n ) n ∈ N ∥ L p ( ℓ ∞ c ) ≤ 2 2/ p n ∈ N sup ∥ x n ∥ p .
Next, we recall the definition of the column-wise tail probability of a sequence ( x n ) (x_n) ( x n ) in M ~ \widetilde{\mathcal{M}} M and a few of its properties and the associated non-commutative analogues of Chebychev’s inequality and Borel-Cantelli lemma. See [13 ] and [Zen15]. For any t > 0 t>0 t > 0 , define
Prob c ( sup n ∈ N ∥ x n ∥ > t ) : = inf { s > 0 : ∃ projection e with τ ( 1 − e ) < s and ∥ x n e ∥ ∞ ≤ t ∀ n ∈ N } . \operatorname{Prob}_{c}\left(\sup_{n\in\mathbb{N}}\|x_n\|>t\right):=\inf\{s>0:\exists\text{ projection }e\text{ with }\tau(1-e)<s\text{ and }\|x_ne\|_{\infty}\le t\ \forall n\in\mathbb{N}\}. Prob c ( n ∈ N sup ∥ x n ∥ > t ) := inf { s > 0 : ∃ projection e with τ ( 1 − e ) < s and ∥ x n e ∥ ∞ ≤ t ∀ n ∈ N } .
This satisfies:
Prob c ( sup n ∈ N ∥ x n ∥ > t ) ≤ Prob c ( sup n ∈ N ∥ x n ∥ > r ) , \operatorname{Prob}_{c}\left(\sup_{n\in\mathbb{N}}\|x_n\|>t\right)\le\operatorname{Prob}_{c}\left(\sup_{n\in\mathbb{N}}\|x_n\|>r\right), Prob c ( n ∈ N sup ∥ x n ∥ > t ) ≤ Prob c ( n ∈ N sup ∥ x n ∥ > r ) ,
whenever t ≥ r t\ge r t ≥ r .
Moreover, if each coefficient a n ≥ 1 a_n\ge1 a n ≥ 1 for n ∈ N n\in\mathbb{N} n ∈ N , then
Prob c ( sup n ∈ N ∥ x n ∥ > t ) ≤ Prob c ( sup n ∈ N ∥ a i x n ∥ > t ) . (4) \operatorname{Prob}_{c}\left(\sup_{n\in\mathbb{N}}\|x_n\|>t\right)\le\operatorname{Prob}_{c}\left(\sup_{n\in\mathbb{N}}\|a_ix_n\|>t\right). \tag*{(4)} Prob c ( n ∈ N sup ∥ x n ∥ > t ) ≤ Prob c ( n ∈ N sup ∥ a i x n ∥ > t ) . (4)
Lemma 2.4. [13 Lemma 3.16] Let ( x n ) n ∈ N (x_n)_{n\in\mathbb{N}} ( x n ) n ∈ N be a sequence of random variables. For t > 0 t>0 t > 0 and 1 ≤ p < ∞ 1\le p<\infty 1 ≤ p < ∞ ,
Prob c ( sup n ∥ x n ∥ > t ) ≤ t − p ∥ x ∥ L p ( ℓ ∞ c ) p . \operatorname{Prob}_{c}\left(\sup_n\|x_n\|>t\right)\le t^{-p}\|x\|_{L_p(\ell_{\infty}^{c})}^{p}. Prob c ( n sup ∥ x n ∥ > t ) ≤ t − p ∥ x ∥ L p ( ℓ ∞ c ) p .
Lemma 2.5. [13 Lemma 3.17] Let ( k n ) n ≥ n 0 (k_n)_{n\ge n_0} ( k n ) n ≥ n 0 be an increasing sequence of integers and ( z n ) (z_n) ( z n ) be a sequence of random variables. Then
∀ δ > 0 , ∑ n ≥ n 0 Prob c ( sup k n ≤ m < k n + 1 ∥ z m ∥ > γ + δ ) < ∞ ⟹ lim sup n → ∞ z n ≤ γ a.u. \forall\delta>0,\qquad\sum_{n\ge n_0}\operatorname{Prob}_{c}\left(\sup_{k_n\le m<k_{n+1}}\|z_m\|>\gamma+\delta\right)<\infty\Longrightarrow\limsup_{n\to\infty}z_n\le\gamma\quad\text{a.u.} ∀ δ > 0 , n ≥ n 0 ∑ Prob c ( k n ≤ m < k n + 1 sup ∥ z m ∥ > γ + δ ) < ∞ ⟹ n → ∞ lim sup z n ≤ γ a.u.
We end this section stating the definition of independent random variables. Let M 1 , M 2 \mathcal{M}_1,\mathcal{M}_2 M 1 , M 2 be two subalgebras of M \mathcal{M} M .
Definition 2.6. We say that M 1 \mathcal{M}_1 M 1 and M 2 \mathcal{M}_2 M 2 are independent if
τ ( x y ) = τ ( x ) τ ( y ) \tau(xy)=\tau(x)\tau(y) τ ( x y ) = τ ( x ) τ ( y )
for all x ∈ M 1 x \in\mathcal{M}_1 x ∈ M 1 and y ∈ M 2 y \in\mathcal{M}_2 y ∈ M 2 . Moreover, a sequence of random variables ( x i ) i ≥ 1 (x_i)_{i\geq1} ( x i ) i ≥ 1 is said to be independent if for all i ≥ 1 i \geq1 i ≥ 1 the von Neumann algebras M i : = { 1 , x 1 , … , x i } ′ ′ \mathcal{M}_i := \{1,x_1,\ldots,x_i\}^{\prime\prime} M i := { 1 , x 1 , … , x i } ′′ and x i + 1 ′ ′ x_{i+1}^{\prime\prime} x i + 1 ′′ are independent.
Weighted LIL for martingale difference sequence
In this section, we prove Theorem A, which establishes a weighted law of the iterated logarithm for martingale difference sequences that generalizes [19 Theorem 1.2] . This theorem can be easily recovered from Lemma 3.2 by choosing B n = s n B_n=s_n B n = s n and α n = 1 \alpha_n=1 α n = 1 . As mentioned in the previous section, we fix a finite von Neumann algebra M \mathcal{M} M equipped with a faithful, normal tracial state τ \tau τ . We further assume an increasing filtration ( M n ) n ∈ N (\mathcal{M}_n)_{n\in\mathbb{N}} ( M n ) n ∈ N of von Neumann subalgebras of M \mathcal{M} M and for all n ∈ N n \in\mathbb{N} n ∈ N , E n : M → M n E_n:\mathcal{M}\to\mathcal{M}_n E n : M → M n be the associated faithful, normal, τ \tau τ -preserving conditional expectation, satisfying E m ∘ E n = E min { m , n } E_m\circ E_n=E_{\min\{m,n\}} E m ∘ E n = E m i n { m , n } for all m , n ∈ N m,n\in\mathbb{N} m , n ∈ N . We recall that a sequence ( x n ) (x_n) ( x n ) in M ~ \widetilde{\mathcal{M}} M is called a martingale sequence with respect to the filtration ( M n ) (\mathcal{M}_n) ( M n ) if we have E n ( x n + 1 ) = x n E_n(x_{n+1})=x_n E n ( x n + 1 ) = x n for all n ∈ N n\in\mathbb{N} n ∈ N . We further have the martingale differences defined by d x n : = x n − x n − 1 d x_n:=x_n-x_{n-1} d x n := x n − x n − 1 for all n ∈ N n\in\mathbb{N} n ∈ N (with x − 1 = 0 x_{-1}=0 x − 1 = 0 ). Without loss of generality, we assume that M 0 = C \mathcal{M}_0=\mathbb{C} M 0 = C and E 0 E_0 E 0 denotes the trivial conditional expectation from M \mathcal{M} M onto M 0 : = C \mathcal{M}_0:=\mathbb{C} M 0 := C . We begin by recalling the following theorem from [19 Theorem 2.10] .
Theorem 3.1. Let ( x k ) k = 0 n (x_k)_{k=0}^n ( x k ) k = 0 n be a finite self-adjoint martingale sequence and d k : = d x k d_k:=d x_k d k := d x k represent the associated martingale differences. Assume that the following conditions are true:
(i) x 0 = 0 x_0=0 x 0 = 0 ,
(ii) ∥ d k ∥ ≤ M \lVert d_k\rVert\leq M ∥ d k ∥ ≤ M for all 1 ≤ k ≤ n 1\leq k\leq n 1 ≤ k ≤ n ,
(iii) ∑ k = 1 n E k − 1 ( d k 2 ) ≤ D 2 1 \sum_{k=1}^n E_{k-1}(d_k^2)\leq D^2 1 ∑ k = 1 n E k − 1 ( d k 2 ) ≤ D 2 1 .
Then, we have the following:
(1) For all λ > 0 \lambda>0 λ > 0 ,
τ ( e λ x n ) ≤ exp [ e λ M − λ M − 1 M 2 D 2 ] . \tau\left(e^{\lambda x_n}\right)\leq\exp\left[\frac{e^{\lambda M}-\lambda M-1}{M^2}D^2\right]. τ ( e λ x n ) ≤ exp [ M 2 e λ M − λ M − 1 D 2 ] . (2) For any ϵ ∈ ( 0 , 1 ] \epsilon\in(0,1] ϵ ∈ ( 0 , 1 ] and λ ∈ [ 0 , 3 ϵ M ] \lambda\in\left[0,\frac{3\epsilon}{M}\right] λ ∈ [ 0 , M 3 ϵ ] ,
τ ( e λ x n ) ≤ exp [ 1 2 ( 1 + ϵ ) λ 2 D 2 ] . \tau\left(e^{\lambda x_n}\right)\leq\exp\left[\frac{1}{2}(1+\epsilon)\lambda^2D^2\right]. τ ( e λ x n ) ≤ exp [ 2 1 ( 1 + ϵ ) λ 2 D 2 ] .
The proof of the following result is close to [19 Theorem 3.1] . However, to study the weighted version of LIL, we have to substitute s n s_n s n and u n u_n u n in [19 Theorem 3.1] with proper coefficients. We include the proof here for the sake of completeness.
Lemma 3.2. Let ( x n ) (x_n) ( x n ) be a non-commutative self-adjoint martingale sequence in M \mathcal{M} M with associated difference sequence ( d n ) (d_n) ( d n ) . Let { α n } \{\alpha_n\} { α n } , { B n } \{B_n\} { B n } , and { c n } \{c_n\} { c n } be positive real sequences satisfying:
(1) B n ↑ ∞ B_n \uparrow\infty B n ↑ ∞ ,
(2) B n + 1 ∼ B n B_{n+1} \sim B_n B n + 1 ∼ B n ,
(3) c n 2 log log B n 2 → 0 c_n^2 \log\log B_n^2 \to0 c n 2 log log B n 2 → 0 ,
(4) for some N > 0 N > 0 N > 0 ,
∥ d m ∥ ∞ ≤ c n α n B n and ∑ m = 1 n E m − 1 ( d m 2 ) ≤ α n 2 B n 2 ⋅ 1. \lVert d_m \rVert_\infty\le c_n \alpha_n B_n \quad\text{and} \quad\sum_{m=1}^{n} E_{m-1}(d_m^2) \le\alpha_n^2 B_n^2 \cdot1. ∥ d m ∥ ∞ ≤ c n α n B n and m = 1 ∑ n E m − 1 ( d m 2 ) ≤ α n 2 B n 2 ⋅ 1.
for all n > N n > N n > N and all 1 ≤ m ≤ n 1 \le m \le n 1 ≤ m ≤ n .
Then
lim sup n → ∞ x n ( 2 B n 2 log log B n 2 ) 1 / 2 ≤ lim sup n → ∞ α n a.u. \limsup_{n \to\infty} \frac{x_n}{(2B_n^2 \log\log B_n^2)^{1/2}} \le\limsup_{n \to\infty} \alpha_n \quad\text{a.u.} n → ∞ lim sup ( 2 B n 2 log log B n 2 ) 1/2 x n ≤ n → ∞ lim sup α n a.u.
Proof. Let α = lim sup α n \alpha= \limsup\alpha_n α = lim sup α n and assume that α < ∞ \alpha< \infty α < ∞ . Choose arbitrarily δ > δ ′ > δ ′ ′ > δ ′ ′ ′ > α \delta> \delta' > \delta'' > \delta''' > \alpha δ > δ ′ > δ ′′ > δ ′′′ > α and 1 < c < δ / δ ′ 1 < c < \delta/\delta' 1 < c < δ / δ ′ .
Now choose ε 0 ∈ ( 0 , 1 ] \varepsilon_0 \in(0,1] ε 0 ∈ ( 0 , 1 ] such that 1 1 + ε 0 ⋅ ( δ ′ ′ δ ′ ′ ′ ) 2 > 1 \frac{1}{1+\varepsilon_0} \cdot\left(\frac{\delta''}{\delta'''}\right)^2 > 1 1 + ε 0 1 ⋅ ( δ ′′′ δ ′′ ) 2 > 1 . Let n k = min { m : B m ≥ c k } n_k = \min\{m : B_m \ge c^k\} n k = min { m : B m ≥ c k } . Now notice that, B n k ∼ c k B_{n_k} \sim c^k B n k ∼ c k .
Also set t n 2 = 2 log log B n 2 t_n^2 = 2 \log\log B_n^2 t n 2 = 2 log log B n 2 . Then t n k ∼ t n k + 1 t_{n_k} \sim t_{n_{k+1}} t n k ∼ t n k + 1 . Now, ∃ K 0 > 0 \exists K_0 > 0 ∃ K 0 > 0 such that, if k ≥ K 0 k \ge K_0 k ≥ K 0 then
n k > n k − 1 ≥ N n_k > n_{k-1} \ge N n k > n k − 1 ≥ N
δ B n k − 1 t n k − 1 > δ ′ B n k t n k \delta B_{n_{k-1}} t_{n_{k-1}} > \delta' B_{n_k} t_{n_k} δ B n k − 1 t n k − 1 > δ ′ B n k t n k
1 1 + ε 0 ⋅ ( δ ′ ′ δ ′ ′ ′ ) 2 t n k 2 ≥ 4 \frac{1}{1+\varepsilon_0} \cdot\left(\frac{\delta''}{\delta'''}\right)^2 t_{n_k}^2 \ge4 1 + ε 0 1 ⋅ ( δ ′′′ δ ′′ ) 2 t n k 2 ≥ 4
α n k < δ ′ ′ ′ \alpha_{n_k} < \delta''' α n k < δ ′′′
δ ′ ′ t n k c n k ≤ 3 δ ′ ′ ′ ε 0 \delta'' t_{n_k} c_{n_k} \le3\delta''' \varepsilon_0 δ ′′ t n k c n k ≤ 3 δ ′′′ ε 0
First note that, since B n t n ↑ ∞ B_n t_n \uparrow\infty B n t n ↑ ∞ and n k − 1 < n ≤ n k n_{k-1} < n \le n_k n k − 1 < n ≤ n k , we have t n B n t n k − 1 B n k − 1 ≥ 1 \frac{t_n B_n}{t_{n_{k-1}} B_{n_{k-1}}} \ge1 t n k − 1 B n k − 1 t n B n ≥ 1 . Now notice that,
Prob c ( sup n k − 1 < n ≤ n k ∥ x n t n B n ∥ > δ ) ≤ Prob c ( sup n k − 1 < n ≤ n k ∥ t n B n t n k − 1 B n k − 1 ⋅ x n t n B n ∥ > δ ) ≤ Prob c ( sup n k − 1 < n ≤ n k ∥ x n ∥ > δ B n k − 1 t n k − 1 ) ≤ Prob c ( sup n k − 1 < n ≤ n k ∥ x n ∥ > δ ′ B n k t n k ) , using (2) ≤ Prob c ( sup n k − 1 < n ≤ n k ∥ x n ∥ > δ ′ ′ t n k B n k ) , Since δ ′ ′ < δ ′ ≤ Prob c ( sup n k − 1 < n ≤ n k ∥ x n α n k B n k ∥ > δ ′ ′ t n k / δ ′ ′ ′ ) , using (4). (5) \begin{aligned}
\operatorname{Prob}_c\left(\sup_{n_{k-1}<n\le n_k}\left\lVert\frac{x_n}{t_nB_n}\right\rVert>\delta\right)
&\le\operatorname{Prob}_c\left(\sup_{n_{k-1}<n\le n_k}\left\lVert\frac{t_nB_n}{t_{n_{k-1}}B_{n_{k-1}}}\cdot\frac{x_n}{t_nB_n}\right\rVert>\delta\right)\\
&\le\operatorname{Prob}_c\left(\sup_{n_{k-1}<n\le n_k}\lVert x_n\rVert>\delta B_{n_{k-1}}t_{n_{k-1}}\right)\\
&\le\operatorname{Prob}_c\left(\sup_{n_{k-1}<n\le n_k}\lVert x_n\rVert>\delta' B_{n_k}t_{n_k}\right),\quad\text{using (2)}\\
&\le\operatorname{Prob}_c\left(\sup_{n_{k-1}<n\le n_k}\lVert x_n\rVert>\delta'' t_{n_k}B_{n_k}\right),\quad\text{Since }\delta''<\delta'\\
&\le\operatorname{Prob}_c\left(\sup_{n_{k-1}<n\le n_k}\left\lVert\frac{x_n}{\alpha_{n_k}B_{n_k}}\right\rVert>\delta''t_{n_k}/\delta'''\right),\quad\text{using (4).} \tag*{(5)}
\end{aligned} Prob c ( n k − 1 < n ≤ n k sup t n B n x n > δ ) ≤ Prob c ( n k − 1 < n ≤ n k sup t n k − 1 B n k − 1 t n B n ⋅ t n B n x n > δ ) ≤ Prob c ( n k − 1 < n ≤ n k sup ∥ x n ∥ > δ B n k − 1 t n k − 1 ) ≤ Prob c ( n k − 1 < n ≤ n k sup ∥ x n ∥ > δ ′ B n k t n k ) , using (2) ≤ Prob c ( n k − 1 < n ≤ n k sup ∥ x n ∥ > δ ′′ t n k B n k ) , Since δ ′′ < δ ′ ≤ Prob c ( n k − 1 < n ≤ n k sup α n k B n k x n > δ ′′ t n k / δ ′′′ ) , using (4). (5)
Set ε k : = δ ′ ′ t n k / δ ′ ′ ′ \varepsilon_k := \delta''t_{n_k}/\delta''' ε k := δ ′′ t n k / δ ′′′ and γ k : = α n k B n k \gamma_k := \alpha_{n_k}B_{n_k} γ k := α n k B n k . We now apply Theorem 3.1 directly to the martingale ( x n / γ k ) 1 ≤ n ≤ n k (x_n/\gamma_k)_{1\le n\le n_k} ( x n / γ k ) 1 ≤ n ≤ n k . Notice that the martingale differences of x n / γ k x_n/\gamma_k x n / γ k are d m / γ k d_m/\gamma_k d m / γ k for 1 ≤ m ≤ n k 1 \le m \le n_k 1 ≤ m ≤ n k . Now note that, ∥ d m ∥ ∞ γ k = ∥ d m ∥ ∞ α n k B n k ≤ c n k : = M k \frac{\lVert d_m\rVert_\infty}{\gamma_k} = \frac{\lVert d_m\rVert_\infty}{\alpha_{n_k}B_{n_k}} \le c_{n_k} := M_k γ k ∥ d m ∥ ∞ = α n k B n k ∥ d m ∥ ∞ ≤ c n k := M k . Also notice that, ∑ m = 1 n k E m − 1 ( d m 2 γ k 2 ) ≤ 1 \sum_{m=1}^{n_k} E_{m-1}\left(\frac{d_m^2}{\gamma_k^2}\right) \le1 ∑ m = 1 n k E m − 1 ( γ k 2 d m 2 ) ≤ 1 . Now applying Theorem 3.1 , we have
τ ( exp ( λ x n k α n k B n k ) ) ≤ exp ( ( 1 + ε 0 ) λ 2 2 ) , for λ ∈ [ 0 , 3 ε 0 / c n k ] . (6) \tau\left(\exp\left(\frac{\lambda x_{n_k}}{\alpha_{n_k}B_{n_k}}\right)\right) \le\exp\left(\frac{(1+\varepsilon_0)\lambda^2}{2}\right), \quad\text{for } \lambda\in[0,3\varepsilon_0/c_{n_k}]. \tag*{(6)} τ ( exp ( α n k B n k λ x n k ) ) ≤ exp ( 2 ( 1 + ε 0 ) λ 2 ) , for λ ∈ [ 0 , 3 ε 0 / c n k ] . (6)
Using Lemma 2.4, in continuation of (5) we have for p ≥ 4 p \ge4 p ≥ 4 and λ > 0 \lambda> 0 λ > 0
Prob c ( sup n k − 1 < n ≤ n k ∥ λ x n α n k B n k ∥ > λ ε k ) ≤ ( λ ε k ) − p ∥ ( λ x n k α n k B n k ) p ∥ L p ( ℓ c ∞ ) . \operatorname{Prob}_{c}\left(\sup_{n_{k-1}<n\le n_k}\left\|\frac{\lambda x_n}{\alpha_{n_k}B_{n_k}}\right\|>\lambda\varepsilon_k\right)\leq(\lambda\varepsilon_k)^{-p}\left\|\left(\frac{\lambda x_{n_k}}{\alpha_{n_k}B_{n_k}}\right)^p\right\|_{L^p(\ell^\infty_c)}. Prob c ( n k − 1 < n ≤ n k sup α n k B n k λ x n > λ ε k ) ≤ ( λ ε k ) − p ( α n k B n k λ x n k ) p L p ( ℓ c ∞ ) .
An application of Theorem 2.3 yields,
∥ ( λ x n α n k B n k ) ∥ L p ( ℓ c ∞ ) ≤ 2 2 / p ∥ λ x n k α n k B n k ∥ p . \left\|\left(\frac{\lambda x_n}{\alpha_{n_k}B_{n_k}}\right)\right\|_{L^p(\ell^\infty_c)}\leq2^{2/p}\left\|\frac{\lambda x_{n_k}}{\alpha_{n_k}B_{n_k}}\right\|_p. ( α n k B n k λ x n ) L p ( ℓ c ∞ ) ≤ 2 2/ p α n k B n k λ x n k p .
By the functional calculus inequality ∣ u ∣ p ≤ p p e − p ( e u + e − u ) |u|^p\leq p^p e^{-p}(e^u+e^{-u}) ∣ u ∣ p ≤ p p e − p ( e u + e − u ) , applied to x n k α n k B n k \frac{x_{n_k}}{\alpha_{n_k}B_{n_k}} α n k B n k x n k and using (6) :
∥ λ x n k α n k B n k ∥ p p ≤ p p e − p ( τ ( e λ x n k α n k B n k ) + τ ( e − λ x n k α n k B n k ) ) ≤ 4 p p e − p exp ( ( 1 + ε 0 ) λ 2 2 ) . \left\|\frac{\lambda x_{n_k}}{\alpha_{n_k}B_{n_k}}\right\|_p^p\leq p^p e^{-p}\left(\tau\left(e^{\frac{\lambda x_{n_k}}{\alpha_{n_k}B_{n_k}}}\right)+\tau\left(e^{-\frac{\lambda x_{n_k}}{\alpha_{n_k}B_{n_k}}}\right)\right)\leq4p^p e^{-p}\exp\left(\frac{(1+\varepsilon_0)\lambda^2}{2}\right). α n k B n k λ x n k p p ≤ p p e − p ( τ ( e α n k B n k λ x n k ) + τ ( e − α n k B n k λ x n k ) ) ≤ 4 p p e − p exp ( 2 ( 1 + ε 0 ) λ 2 ) .
Combining we have,
Prob c ( sup n k − 1 < n ≤ n k ∥ λ x n α n k B n k ∥ > λ ε k ) ≤ 8 ( λ ε k ) − p p p e − p exp ( ( 1 + ε 0 ) λ 2 2 ) \operatorname{Prob}_{c}\left(\sup_{n_{k-1}<n\leq n_k}\left\|\frac{\lambda x_n}{\alpha_{n_k}B_{n_k}}\right\|>\lambda\varepsilon_k\right)\leq8(\lambda\varepsilon_k)^{-p}p^p e^{-p}\exp\left(\frac{(1+\varepsilon_0)\lambda^2}{2}\right) Prob c ( n k − 1 < n ≤ n k sup α n k B n k λ x n > λ ε k ) ≤ 8 ( λ ε k ) − p p p e − p exp ( 2 ( 1 + ε 0 ) λ 2 )
≤ 8 ( p λ ε k ) p exp ( ( 1 + ε 0 ) λ 2 2 ) . \leq8\left(\frac{p}{\lambda\varepsilon_k}\right)^p\exp\left(\frac{(1+\varepsilon_0)\lambda^2}{2}\right). ≤ 8 ( λ ε k p ) p exp ( 2 ( 1 + ε 0 ) λ 2 ) .
Put p = λ ε k p=\lambda\varepsilon_k p = λ ε k . Thus aforesaid equation becomes,
Prob c ( sup n k − 1 < n ≤ n k ∥ λ x n α n k B n k ∥ > λ ε k ) ≤ 8 exp ( ( 1 + ε 0 ) λ 2 2 − λ ε k ) . \operatorname{Prob}_{c}\left(\sup_{n_{k-1}<n\leq n_k}\left\|\frac{\lambda x_n}{\alpha_{n_k}B_{n_k}}\right\|>\lambda\varepsilon_k\right)\leq8\exp\left(\frac{(1+\varepsilon_0)\lambda^2}{2}-\lambda\varepsilon_k\right). Prob c ( n k − 1 < n ≤ n k sup α n k B n k λ x n > λ ε k ) ≤ 8 exp ( 2 ( 1 + ε 0 ) λ 2 − λ ε k ) .
Now set λ = ε k 1 + ε 0 \lambda=\frac{\varepsilon_k}{1+\varepsilon_0} λ = 1 + ε 0 ε k . Using (3) and (5) it follows that p ≥ 4 p\geq4 p ≥ 4 and λ ≤ 3 ε 0 / c n k \lambda\leq3\varepsilon_0/c_{n_k} λ ≤ 3 ε 0 / c n k . Since ε k = δ ′ ′ t n k / δ ′ ′ ′ \varepsilon_k=\delta^{\prime\prime}t_{n_k}/\delta^{\prime\prime\prime} ε k = δ ′′ t n k / δ ′′′ , we have
Prob c ( sup n k − 1 < n ≤ n k ∥ λ x n α n k B n k ∥ > λ ε k ) ≤ 8 exp ( − ( 1 1 + ε 0 ⋅ ( δ ′ ′ δ ′ ′ ′ ) 2 log log B n k 2 ) ) \operatorname{Prob}_{c}\left(\sup_{n_{k-1}<n\leq n_k}\left\|\frac{\lambda x_n}{\alpha_{n_k}B_{n_k}}\right\|>\lambda\varepsilon_k\right)\leq8\exp\left(-\left(\frac{1}{1+\varepsilon_0}\cdot\left(\frac{\delta^{\prime\prime}}{\delta^{\prime\prime\prime}}\right)^2\log\log B_{n_k}^2\right)\right) Prob c ( n k − 1 < n ≤ n k sup α n k B n k λ x n > λ ε k ) ≤ 8 exp ( − ( 1 + ε 0 1 ⋅ ( δ ′′′ δ ′′ ) 2 log log B n k 2 ) )
∼ 8 ( ( 2 log c ) ⋅ k ) − ( 1 1 + ε 0 ⋅ ( δ ′ ′ δ ′ ′ ′ ) 2 ) , since B n k ∼ c k . \sim8\left((2\log c)\cdot k\right)^{-\left(\frac{1}{1+\varepsilon_0}\cdot\left(\frac{\delta^{\prime\prime}}{\delta^{\prime\prime\prime}}\right)^2\right)},\quad\text{since }B_{n_k}\sim c^k. ∼ 8 ( ( 2 log c ) ⋅ k ) − ( 1 + ε 0 1 ⋅ ( δ ′′′ δ ′′ ) 2 ) , since B n k ∼ c k .
Since 1 1 + ε 0 ⋅ ( δ ′ ′ δ ′ ′ ′ ) 2 > 1 \frac{1}{1+\varepsilon_0}\cdot\left(\frac{\delta^{\prime\prime}}{\delta^{\prime\prime\prime}}\right)^2>1 1 + ε 0 1 ⋅ ( δ ′′′ δ ′′ ) 2 > 1 by assumption, we have
∑ k Prob c ( sup n k − 1 < n ≤ n k ∥ x n t n B n ∥ > δ ) < + ∞ . \sum_k\operatorname{Prob}_{c}\left(\sup_{n_{k-1}<n\leq n_k}\left\|\frac{x_n}{t_nB_n}\right\|>\delta\right)<+\infty. k ∑ Prob c ( n k − 1 < n ≤ n k sup t n B n x n > δ ) < + ∞.
Since δ > α = lim sup α n \delta>\alpha=\limsup\alpha_n δ > α = lim sup α n is already chosen, it follows by Lemma 2.5
lim sup n → ∞ x n ( 2 B n 2 log log B n 2 ) 1 / 2 ≤ lim sup n → ∞ α n a.u. \limsup_{n\to\infty}\frac{x_n}{(2B_n^2\log\log B_n^2)^{1/2}}\leq\limsup_{n\to\infty}\alpha_n\quad\text{a.u.} n → ∞ lim sup ( 2 B n 2 log log B n 2 ) 1/2 x n ≤ n → ∞ lim sup α n a.u.
We now use this result to obtain our first non-commutative weighted law of iterated logarithm for martingale difference sequences, weighted by monomials.
Theorem 3.3. Let ( x n ) (x_n) ( x n ) and ( d n ) (d_n) ( d n ) are sequences as in Lemma 3.2. Further assume that τ ( d 1 ) = 0 \tau(d_1)=0 τ ( d 1 ) = 0 , E m − 1 ( d m 2 ) = σ m 2 ⋅ 1 < ∞ E_{m-1}(d_m^2)=\sigma_m^2\cdot1<\infty E m − 1 ( d m 2 ) = σ m 2 ⋅ 1 < ∞ and max m ≤ n ∥ d m ∥ ≤ M n \max_{m\leq n}\|d_m\|\leq M_n max m ≤ n ∥ d m ∥ ≤ M n for all m , n ∈ N m,n\in\mathbb{N} m , n ∈ N , where ( M n ) ⊂ R + (M_n)\subset\mathbb{R}^{+} ( M n ) ⊂ R + . Fix j ≥ 0 j\geq0 j ≥ 0 and define $B_n^2:=
∑ m = 1 n m 2 j σ m 2 \sum_{m=1}^{n}m^{2j}\sigma_m^2 ∑ m = 1 n m 2 j σ m 2 . Assume that for every j ≥ 0 j \ge0 j ≥ 0 , B n 2 → ∞ B_n^2 \to\infty B n 2 → ∞ and n 2 j M n 2 L ( B n 2 ) B n 2 → 0 \frac{n^{2j}M_n^2L(B_n^2)}{B_n^2} \to0 B n 2 n 2 j M n 2 L ( B n 2 ) → 0 , as n → ∞ n \to\infty n → ∞ . Then,
lim sup n → ∞ ∑ m = 1 n m j d m B n 2 log log B n 2 ≤ 2 (7) \limsup_{n\to\infty}\frac{\sum_{m=1}^{n}m^jd_m}{\sqrt{B_n^2\log\log B_n^2}}\leq\sqrt{2} \tag*{(7)} n → ∞ lim sup B n 2 log log B n 2 ∑ m = 1 n m j d m ≤ 2 (7)
Proof. Fix j ≥ 0 j \ge0 j ≥ 0 and define x n : = ∑ m = 1 n m j d m x_n := \sum_{m=1}^{n}m^jd_m x n := ∑ m = 1 n m j d m . Note that ( x n ) (x_n) ( x n ) is a sequence of self-adjoint martingales associated to the filtration ( M n ) (\mathcal{M}_n) ( M n ) . Its martingale differences are Δ m : = x m − x m − 1 = m j d m \Delta_m := x_m-x_{m-1}=m^jd_m Δ m := x m − x m − 1 = m j d m , 1 ≤ m ≤ n 1\leq m\leq n 1 ≤ m ≤ n . Set α n : = 1 \alpha_n:=1 α n := 1 and c n : = n j M n B n c_n:=\frac{n^jM_n}{B_n} c n := B n n j M n . By hypothesis n 2 j M n 2 L ( B n 2 ) B n 2 → 0 \frac{n^{2j}M_n^2L(B_n^2)}{B_n^2}\to0 B n 2 n 2 j M n 2 L ( B n 2 ) → 0 , therefore c n 2 log log B n 2 → 0 c_n^2\log\log B_n^2\to0 c n 2 log log B n 2 → 0 as n → ∞ n\to\infty n → ∞ .
Next, for every 1 ≤ m ≤ n 1\leq m\leq n 1 ≤ m ≤ n , ∥ Δ m ∥ ∞ ≤ c n B n \|\Delta_m\|_\infty\leq c_nB_n ∥ Δ m ∥ ∞ ≤ c n B n . Since for all m ∈ N m\in\mathbb{N} m ∈ N , E m − 1 ( d m 2 ) = σ m 2 ⋅ 1 E_{m-1}(d_m^2)=\sigma_m^2\cdot1 E m − 1 ( d m 2 ) = σ m 2 ⋅ 1 , we obtain
∑ m = 1 n E m − 1 ( Δ m 2 ) = B n 2 ⋅ 1. \sum_{m=1}^{n}E_{m-1}(\Delta_m^2)=B_n^2\cdot1. m = 1 ∑ n E m − 1 ( Δ m 2 ) = B n 2 ⋅ 1.
Also notice that,
0 ≤ B n + 1 2 − B n 2 ≤ ∥ Δ n + 1 ∥ 2 ≤ ( n + 1 ) 2 j M n + 1 2 . 0\leq B_{n+1}^2-B_n^2\leq\|\Delta_{n+1}\|^2\leq(n+1)^{2j}M_{n+1}^2. 0 ≤ B n + 1 2 − B n 2 ≤ ∥ Δ n + 1 ∥ 2 ≤ ( n + 1 ) 2 j M n + 1 2 .
Since by the assumption n 2 j M n 2 L ( B n 2 ) B n 2 → 0 \frac{n^{2j}M_n^2L(B_n^2)}{B_n^2}\to0 B n 2 n 2 j M n 2 L ( B n 2 ) → 0 , we have ( B n + 1 2 − B n 2 ) / B n + 1 2 → 0 (B_{n+1}^2-B_n^2)/B_{n+1}^2\to0 ( B n + 1 2 − B n 2 ) / B n + 1 2 → 0 as n → ∞ n\to\infty n → ∞ . Hence, B n + 1 ∼ B n B_{n+1}\sim B_n B n + 1 ∼ B n , as n → ∞ n\to\infty n → ∞ . This follows that, all the conditions of Lemma 3.2 are satisfied, by taking α n = 1 \alpha_n=1 α n = 1 . Therefore,
lim sup n → ∞ x n 2 B n 2 log log B n 2 ≤ lim sup n → ∞ α n = 1. \limsup_{n\to\infty}\frac{x_n}{\sqrt{2B_n^2\log\log B_n^2}}\leq\limsup_{n\to\infty}\alpha_n=1. n → ∞ lim sup 2 B n 2 log log B n 2 x n ≤ n → ∞ lim sup α n = 1.
□
Corollary 3.4. Let ( x n ) (x_n) ( x n ) and ( d n ) (d_n) ( d n ) are sequences as in Lemma 3.2. Assume that τ ( d 1 ) = 0 \tau(d_1)=0 τ ( d 1 ) = 0 , E m − 1 ( d m 2 ) = 1 E_{m-1}(d_m^2)=1 E m − 1 ( d m 2 ) = 1 and max m ≤ n ∥ d m ∥ ≤ M n \max_{m\leq n}\|d_m\|\leq M_n max m ≤ n ∥ d m ∥ ≤ M n , where ( M n ) (M_n) ( M n ) is a sequence of positive reals such that lim n → ∞ L ( n ) M n 2 / n = 0 \lim_{n\to\infty}L(n)M_n^2/n=0 lim n → ∞ L ( n ) M n 2 / n = 0 . Then, for every j ≥ 0 j\geq0 j ≥ 0 ,
lim sup n → ∞ n − j ( 2 n log log n ) − 1 / 2 ∑ m = 1 n m j d m ≤ ( 2 j + 1 ) − 1 / 2 (8) \limsup_{n\to\infty}n^{-j}(2n\log\log n)^{-1/2}\sum_{m=1}^{n}m^jd_m\leq(2j+1)^{-1/2} \tag*{(8)} n → ∞ lim sup n − j ( 2 n log log n ) − 1/2 m = 1 ∑ n m j d m ≤ ( 2 j + 1 ) − 1/2 (8)
Proof. The result follows immediately from Theorem 3.3 by taking σ m = 1 \sigma_m=1 σ m = 1 for all m ≥ 1 m\geq1 m ≥ 1 . In this case, B n 2 = ∑ m = 1 n m 2 j B_n^2=\sum_{m=1}^{n}m^{2j} B n 2 = ∑ m = 1 n m 2 j . Now use the facts B n 2 ∼ n 2 j + 1 2 j + 1 B_n^2\sim\frac{n^{2j+1}}{2j+1} B n 2 ∼ 2 j + 1 n 2 j + 1 and log log B n 2 ∼ log log n \log\log B_n^2\sim\log\log n log log B n 2 ∼ log log n in (7) , we obtain
lim sup n → ∞ n − j ( 2 n log log n ) − 1 / 2 ∑ m = 1 n m j d m ≤ ( 2 j + 1 ) − 1 / 2 . \limsup_{n\to\infty}n^{-j}(2n\log\log n)^{-1/2}\sum_{m=1}^{n}m^jd_m\leq(2j+1)^{-1/2}. n → ∞ lim sup n − j ( 2 n log log n ) − 1/2 m = 1 ∑ n m j d m ≤ ( 2 j + 1 ) − 1/2 .
□
Next, we focus on establishing the above results for more general weights, namely polynomials. We begin with the following technical lemma.
Lemma 3.5. Let c > 1 c>1 c > 1 and r n = 2 n log log n r_n=\sqrt{2n\log\log n} r n = 2 n log log n , n ∈ N n\in\mathbb{N} n ∈ N . Define a sequence n k = ⌊ c k ⌋ + 1 n_k=\lfloor c^k\rfloor+1 n k = ⌊ c k ⌋ + 1 for k ≥ 1 k\geq1 k ≥ 1 . Then there exists an integer K ≥ 1 K\geq1 K ≥ 1 such that for every k ≥ K k\geq K k ≥ K :
sup n k − 1 < n ≤ n k ∣ 1 r n − 1 r n k ∣ ≤ 6 ( c − 1 ) r n k . \sup_{n_{k-1}<n\leq n_k}\left|\frac{1}{r_n}-\frac{1}{r_{n_k}}\right|\leq\frac{6(c-1)}{r_{n_k}}. n k − 1 < n ≤ n k sup r n 1 − r n k 1 ≤ r n k 6 ( c − 1 ) .
Proof. First, observe that n k ∼ c k n_k \sim c^k n k ∼ c k . Choose K 0 ≥ 1 K_0 \ge1 K 0 ≥ 1 , such that n k − 1 < n k n_{k-1} < n_k n k − 1 < n k for all k ≥ K 0 k \ge K_0 k ≥ K 0 . In particular, for all k ≥ K 0 k \ge K_0 k ≥ K 0 :
1 ≤ n k n k − 1 ≤ c + 1 ≤ 2 c . (9) 1 \le\frac{n_k}{n_{k-1}} \le c+1 \le2c. \tag*{(9)} 1 ≤ n k − 1 n k ≤ c + 1 ≤ 2 c . (9)
Now using log log n k log log n k − 1 → 1 \frac{\log\log n_k}{\log\log n_{k-1}} \to1 l o g l o g n k − 1 l o g l o g n k → 1 as k → ∞ k \to\infty k → ∞ , there exists K 1 > K 0 K_1 > K_0 K 1 > K 0 such that for all k ≥ K 1 k \ge K_1 k ≥ K 1
log log n k ≤ 2 log log n k − 1 . (10) \log\log n_k \le2\log\log n_{k-1}. \tag*{(10)} log log n k ≤ 2 log log n k − 1 . (10)
Now notice that, for every k ≥ K 1 k \ge K_1 k ≥ K 1 , (9) and (10) imply r n k 2 ≤ 8 c r n k − 1 2 r_{n_k}^2 \le8cr_{n_{k-1}}^2 r n k 2 ≤ 8 c r n k − 1 2 . Now fix k ≥ K 1 k \ge K_1 k ≥ K 1 and let n k − 1 < n ≤ n k n_{k-1} < n \le n_k n k − 1 < n ≤ n k . Since n ≤ n k n \le n_k n ≤ n k , we have
∣ 1 r n − 1 r n k ∣ ≤ r n k 2 − r n k − 1 2 r n k − 1 2 r n k . (11) \left|\frac{1}{r_n}-\frac{1}{r_{n_k}}\right| \le\frac{r_{n_k}^2-r_{n_{k-1}}^2}{r_{n_{k-1}}^2r_{n_k}}. \tag*{(11)} r n 1 − r n k 1 ≤ r n k − 1 2 r n k r n k 2 − r n k − 1 2 . (11)
To estimate the numerator, write r n k 2 − r n k − 1 2 = 2 ( ( n k − n k − 1 ) log log n k + n k − 1 ( log log n k − log log n k − 1 ) ) r_{n_k}^2-r_{n_{k-1}}^2=2\left((n_k-n_{k-1})\log\log n_k+n_{k-1}(\log\log n_k-\log\log n_{k-1})\right) r n k 2 − r n k − 1 2 = 2 ( ( n k − n k − 1 ) log log n k + n k − 1 ( log log n k − log log n k − 1 ) ) . Since, n k = ⌊ c k ⌋ + 1 ≤ c k + 1 n_k=\lfloor c^k\rfloor+1\le c^k+1 n k = ⌊ c k ⌋ + 1 ≤ c k + 1 , n k − 1 ≥ c k − 1 n_{k-1}\ge c^{k-1} n k − 1 ≥ c k − 1 , we obtain n k − n k − 1 ≤ ( c − 1 ) n k − 1 + 1 n_k-n_{k-1}\le(c-1)n_{k-1}+1 n k − n k − 1 ≤ ( c − 1 ) n k − 1 + 1 . Now choose K 2 > K 1 K_2>K_1 K 2 > K 1 such that for all k ≥ K 2 k\ge K_2 k ≥ K 2 we have
n k − n k − 1 ≤ 2 ( c − 1 ) n k − 1 . (12) n_k-n_{k-1}\le2(c-1)n_{k-1}. \tag*{(12)} n k − n k − 1 ≤ 2 ( c − 1 ) n k − 1 . (12)
Using (10) and (12) , for all k ≥ K 2 k\ge K_2 k ≥ K 2 we obtain, ( n k − n k − 1 ) log log n k ≤ 4 ( c − 1 ) n k − 1 log log n k − 1 (n_k-n_{k-1})\log\log n_k\le4(c-1)n_{k-1}\log\log n_{k-1} ( n k − n k − 1 ) log log n k ≤ 4 ( c − 1 ) n k − 1 log log n k − 1 . Now, by the mean value theorem applied to g ( x ) = log log x g(x)=\log\log x g ( x ) = log log x , there exists ξ k ∈ ( n k − 1 , n k ) \xi_k\in(n_{k-1},n_k) ξ k ∈ ( n k − 1 , n k ) such that log log n k − log log n k − 1 = ( n k − n k − 1 ) / ξ k log ξ k \log\log n_k-\log\log n_{k-1}=(n_k-n_{k-1})/\xi_k\log\xi_k log log n k − log log n k − 1 = ( n k − n k − 1 ) / ξ k log ξ k . Since ξ k ≥ n k − 1 \xi_k\ge n_{k-1} ξ k ≥ n k − 1 , we have n k − 1 ( log log n k − log log n k − 1 ) ≤ n k − n k − 1 log n k − 1 ≤ 2 ( c − 1 ) n k − 1 log log n k − 1 n_{k-1}(\log\log n_k-\log\log n_{k-1})\le\frac{n_k-n_{k-1}}{\log n_{k-1}}\le2(c-1)n_{k-1}\log\log n_{k-1} n k − 1 ( log log n k − log log n k − 1 ) ≤ l o g n k − 1 n k − n k − 1 ≤ 2 ( c − 1 ) n k − 1 log log n k − 1 , for all k ≥ K k\ge K k ≥ K , where K > K 2 K>K_2 K > K 2 so large that 1 log n k − 1 ≤ log log n k − 1 \frac{1}{\log n_{k-1}}\le\log\log n_{k-1} l o g n k − 1 1 ≤ log log n k − 1 for all k ≥ K k\ge K k ≥ K . Combining the above estimates, we have r n k 2 − r n k − 1 2 ≤ 6 ( c − 1 ) r n k − 1 2 r_{n_k}^2-r_{n_{k-1}}^2\le6(c-1)r_{n_{k-1}}^2 r n k 2 − r n k − 1 2 ≤ 6 ( c − 1 ) r n k − 1 2 , k ≥ K k\ge K k ≥ K , and therefore, (11) yields
∣ 1 r n − 1 r n k ∣ ≤ 6 ( c − 1 ) r n k , n k − 1 < n ≤ n k , k ≥ K . \left|\frac{1}{r_n}-\frac{1}{r_{n_k}}\right|\le\frac{6(c-1)}{r_{n_k}},\quad n_{k-1}<n\le n_k,\ k\ge K. r n 1 − r n k 1 ≤ r n k 6 ( c − 1 ) , n k − 1 < n ≤ n k , k ≥ K .
For the remainder of this section, let ( r n ) n (r_n)_n ( r n ) n and ( n k ) k (n_k)_k ( n k ) k be the sequences obtained in Lemma 3.5.
Lemma 3.6. Let p ( t ) = ∑ j = 0 l a j t j p(t)=\sum_{j=0}^{l}a_jt^j p ( t ) = ∑ j = 0 l a j t j be a polynomial on [ 0 , 1 ] [0,1] [ 0 , 1 ] and ( d n ) (d_n) ( d n ) be the sequence as in Theorem 3.3. Assume that for each j = 0 , 1 , … , l j=0,1,\ldots,l j = 0 , 1 , … , l ,
lim sup n → ∞ 1 n j r n ∑ m = 1 n m j d m ≤ a . u . 1. (13) \limsup_{n\to\infty}\frac{1}{n^jr_n}\sum_{m=1}^{n}m^jd_m\underset{a.u.}{\le}1. \tag*{(13)} n → ∞ lim sup n j r n 1 m = 1 ∑ n m j d m a . u . ≤ 1. (13) Then there exists a constant C p > 0 C_p>0 C p > 0 such that
(i) lim sup k → ∞ ( 1 r n − 1 r n k ) ∑ m = 1 n p ( m n ) d m ≤ a . u . 6 ( c − 1 ) A , \limsup_{k\to\infty}\left(\frac{1}{r_n}-\frac{1}{r_{n_k}}\right)\sum_{m=1}^{n}p\left(\frac{m}{n}\right)d_m\underset{a.u.}{\le}6(c-1)A, lim sup k → ∞ ( r n 1 − r n k 1 ) ∑ m = 1 n p ( n m ) d m a . u . ≤ 6 ( c − 1 ) A ,
(ii) lim sup k → ∞ 1 r n k ∑ m = 1 n ( p ( m / n ) − p ( m / n k ) ) d m ≤ a . u . C p ( c − 1 ) , \limsup_{k\to\infty}\frac{1}{r_{n_k}}\sum_{m=1}^{n}\left(p(m/n)-p(m/n_k)\right)d_m\underset{a.u.}{\le}C_p(c-1), lim sup k → ∞ r n k 1 ∑ m = 1 n ( p ( m / n ) − p ( m / n k ) ) d m a . u . ≤ C p ( c − 1 ) ,
where A = ∑ j = 0 l ∣ a j ∣ A=\sum_{j=0}^{l}|a_j| A = ∑ j = 0 l ∣ a j ∣ .
Proof. First notice that, ∑ m = 1 n p ( m n ) d m = ∑ j = 0 l a j n j ∑ m = 1 n m j d m \sum_{m=1}^{n}p\left(\frac{m}{n}\right)d_m=\sum_{j=0}^{l}\frac{a_j}{n^j}\sum_{m=1}^{n}m^jd_m ∑ m = 1 n p ( n m ) d m = ∑ j = 0 l n j a j ∑ m = 1 n m j d m . Then using (13) , for each j j j , there exists e j ∈ P ( M ) e_j\in\mathcal{P}(M) e j ∈ P ( M ) with τ ( 1 − e j ) < ε / ( l + 1 ) \tau(1-e_j)<\varepsilon/(l+1) τ ( 1 − e j ) < ε / ( l + 1 ) and K j ′ ≥ 1 K'_j\geq1 K j ′ ≥ 1 such that ∥ 1 n j r n ( ∑ m = 1 n m j d m ) e j ∥ ∞ ≤ 2 \left\|\frac{1}{n^jr_n}\left(\sum_{m=1}^{n}m^jd_m\right)e_j\right\|_\infty\leq2 n j r n 1 ( ∑ m = 1 n m j d m ) e j ∞ ≤ 2 , for all n ≥ K j ′ n\geq K'_j n ≥ K j ′ .
Set e : = ⋀ j = 0 l e j e:=\bigwedge_{j=0}^{l}e_j e := ⋀ j = 0 l e j . Then τ ( 1 − e ) ≤ ε \tau(1-e)\leq\varepsilon τ ( 1 − e ) ≤ ε and for all n ≥ N : = max j K j ′ n\geq N:=\max_jK'_j n ≥ N := max j K j ′ :
∥ ( ∑ m = 1 n p ( m n ) d m ) e ∥ ∞ ≤ ∑ j = 0 l ∣ a j ∣ ⋅ 1 n j ∥ ( ∑ m = 1 n m j d m ) e ∥ ∞ = 2 r n A . (14) \left\|\left(\sum_{m=1}^{n}p\left(\frac{m}{n}\right)d_m\right)e\right\|_\infty
\leq\sum_{j=0}^{l}|a_j|\cdot\frac{1}{n^j}\left\|\left(\sum_{m=1}^{n}m^jd_m\right)e\right\|_\infty
=2r_nA. \tag*{(14)} ( m = 1 ∑ n p ( n m ) d m ) e ∞ ≤ j = 0 ∑ l ∣ a j ∣ ⋅ n j 1 ( m = 1 ∑ n m j d m ) e ∞ = 2 r n A . (14)
Since n k → ∞ n_k\to\infty n k → ∞ , there exists K > max { N , K 2 } K>\max\{N,K_2\} K > max { N , K 2 } such that n K 2 − 1 ≥ K n_{K_2-1}\geq K n K 2 − 1 ≥ K . Therefore, using Lemma 3.5, for all n k − 1 < n ≤ n k n_{k-1}<n\leq n_k n k − 1 < n ≤ n k and n ≥ K n\geq K n ≥ K :
∥ ( ( 1 r n − 1 r n k ) ∑ m = 1 n p ( m n ) d m ) e ∥ ∞ ≤ 6 ( c − 1 ) A , \left\|\left(\left(\frac{1}{r_n}-\frac{1}{r_{n_k}}\right)\sum_{m=1}^{n}p\left(\frac{m}{n}\right)d_m\right)e\right\|_\infty\leq6(c-1)A, ( ( r n 1 − r n k 1 ) m = 1 ∑ n p ( n m ) d m ) e ∞ ≤ 6 ( c − 1 ) A ,
This concludes (i).
Now observe that,
1 r n k ∑ m = 1 n [ p ( m n ) − p ( m n k ) ] d m = 1 r n k ∑ j = 0 l a j ( 1 n j − 1 n k j ) ∑ m = 1 n m j d m . \frac{1}{r_{n_k}}\sum_{m=1}^{n}\left[p\left(\frac{m}{n}\right)-p\left(\frac{m}{n_k}\right)\right]d_m
=\frac{1}{r_{n_k}}\sum_{j=0}^{l}a_j\left(\frac{1}{n^j}-\frac{1}{n_k^j}\right)\sum_{m=1}^{n}m^jd_m. r n k 1 m = 1 ∑ n [ p ( n m ) − p ( n k m ) ] d m = r n k 1 j = 0 ∑ l a j ( n j 1 − n k j 1 ) m = 1 ∑ n m j d m .
For n k − 1 < n ≤ n k n_{k-1}<n\leq n_k n k − 1 < n ≤ n k and each j ≥ 1 j\geq1 j ≥ 1 , by the mean value theorem applied to g ( x ) = x − j g(x)=x^{-j} g ( x ) = x − j and using (12) , we have ∣ 1 n j − 1 n k j ∣ ≤ 2 j ( c − 1 ) n k − 1 j \left|\frac{1}{n^j}-\frac{1}{n_k^j}\right|\leq\frac{2j(c-1)}{n_{k-1}^j} n j 1 − n k j 1 ≤ n k − 1 j 2 j ( c − 1 ) , for all k ≥ K k\geq K k ≥ K . For j = 0 j=0 j = 0 , we have 1 / n 0 − 1 / n k 0 = 0 1/n^0-1/n_k^0=0 1/ n 0 − 1/ n k 0 = 0 , so the j = 0 j=0 j = 0 term vanishes. Using (14) (with n n n replaced by the same n n n , and e e e as above), for each j ≥ 1 j\geq1 j ≥ 1 and n ≤ n k n\leq n_k n ≤ n k , 1 n j r n ∥ ( ∑ m = 1 n m j d m ) e ∥ ∞ ≤ 2 r n k \frac{1}{n^jr_n}\left\|\left(\sum_{m=1}^{n}m^jd_m\right)e\right\|_\infty\leq2r_{n_k} n j r n 1 ( ∑ m = 1 n m j d m ) e ∞ ≤ 2 r n k .
Hence for j ≥ 1 j\geq1 j ≥ 1 and for all k ≥ K k\geq K k ≥ K :
1 r n k ∣ 1 n j − 1 n k j ∣ ∥ ( ∑ m = 1 n m j d m ) e ∥ ∞ ≤ 1 r n k ⋅ 2 j ( c − 1 ) n k − 1 j ⋅ 2 n j r n k ≤ 4 j ( c − 1 ) n k j n k − 1 j ≤ 4 j ( c − 1 ) ( 2 c ) j , \frac{1}{r_{n_k}}\left|\frac{1}{n^j}-\frac{1}{n_k^j}\right|\left\|\left(\sum_{m=1}^{n}m^jd_m\right)e\right\|_\infty
\leq\frac{1}{r_{n_k}}\cdot\frac{2j(c-1)}{n_{k-1}^j}\cdot2n^jr_{n_k}
\leq\frac{4j(c-1)n_k^j}{n_{k-1}^j}\leq4j(c-1)(2c)^j, r n k 1 n j 1 − n k j 1 ( m = 1 ∑ n m j d m ) e ∞ ≤ r n k 1 ⋅ n k − 1 j 2 j ( c − 1 ) ⋅ 2 n j r n k ≤ n k − 1 j 4 j ( c − 1 ) n k j ≤ 4 j ( c − 1 ) ( 2 c ) j ,
where we used n k / n k − 1 ≤ 2 c n_k/n_{k-1}\leq2c n k / n k − 1 ≤ 2 c from (9) . Therefore, summing over j j j , we have
∥ ( 1 r n k ∑ m = 1 n [ p ( m n ) − p ( m n k ) ] d m ) e ∥ ∞ ≤ ∑ j = 1 l ∣ a j ∣ ⋅ 4 j ( c − 1 ) ( 2 c ) j = ( c − 1 ) C p , (15) \left\|\left(\frac{1}{r_{n_k}}\sum_{m=1}^{n}\left[p\left(\frac{m}{n}\right)-p\left(\frac{m}{n_k}\right)\right]d_m\right)e\right\|_\infty
\leq\sum_{j=1}^{l}|a_j|\cdot4j(c-1)(2c)^j
=(c-1)C_p, \tag*{(15)} ( r n k 1 m = 1 ∑ n [ p ( n m ) − p ( n k m ) ] d m ) e ∞ ≤ j = 1 ∑ l ∣ a j ∣ ⋅ 4 j ( c − 1 ) ( 2 c ) j = ( c − 1 ) C p , (15)
for a finite constant C p = ∑ j = 1 l 4 j ( 2 c ) j ∣ a j ∣ C_p=\sum_{j=1}^{l}4j(2c)^j|a_j| C p = ∑ j = 1 l 4 j ( 2 c ) j ∣ a j ∣ . This proves (ii). □ \square □
Lemma 3.7. Let p ( t ) = ∑ j = 0 l a j t j p(t)=\sum_{j=0}^{l}a_jt^j p ( t ) = ∑ j = 0 l a j t j be a polynomial on [ 0 , 1 ] [0,1] [ 0 , 1 ] such that ∥ p ∥ 2 : = ( ∫ 0 1 p ( t ) 2 d t ) 1 / 2 = 1 \|p\|_2:=\left(\int_0^1p(t)^2dt\right)^{1/2}=1 ∥ p ∥ 2 := ( ∫ 0 1 p ( t ) 2 d t ) 1/2 = 1 and ( d n ) (d_n) ( d n ) be the sequence as in Corollary 3.4. Then
lim sup k → ∞ 1 r n k ∑ m = 1 n p ( m / n k ) d m ≤ 1 a.u. (16) \limsup_{k\to\infty}\frac{1}{r_{n_k}}\sum_{m=1}^{n}p(m/n_k)d_m\leq1\quad\text{a.u.} \tag*{(16)} k → ∞ lim sup r n k 1 m = 1 ∑ n p ( m / n k ) d m ≤ 1 a.u. (16)
Proof. Let us first denote A k ( n ) = 1 r n k ∑ m = 1 n p ( m / n k ) d m A_k(n)=\frac{1}{r_{n_k}}\sum_{m=1}^{n}p(m/n_k)d_m A k ( n ) = r n k 1 ∑ m = 1 n p ( m / n k ) d m , for n k − 1 < n ≤ n k n_{k-1}<n\leq n_k n k − 1 < n ≤ n k . For fixed k k k , the weights p ( m / n k ) p(m/n_k) p ( m / n k ) are deterministic scalars, so ( A k ( n ) ) 1 < n ≤ n k (A_k(n))_{1<n\leq n_k} ( A k ( n ) ) 1 < n ≤ n k forms a self-adjoint martingale with respect to the filtration ( M n ) (\mathcal{M}_n) ( M n ) .
We will show that, for any δ > 0 \delta>0 δ > 0 , ∑ k Prob c ( sup n k − 1 < n ≤ n k ∥ A k ( n ) ∥ > 1 + δ ) < ∞ \sum_k\operatorname{Prob}_c\left(\sup_{n_{k-1}<n\leq n_k}\|A_k(n)\|>1+\delta\right)<\infty ∑ k Prob c ( sup n k − 1 < n ≤ n k ∥ A k ( n ) ∥ > 1 + δ ) < ∞ .
Choose ε 0 ∈ ( 0 , 1 ) \varepsilon_0 \in(0,1) ε 0 ∈ ( 0 , 1 ) so that ( 1 + δ ) 2 1 + ε 0 > 1 \frac{(1+\delta)^2}{1+\varepsilon_0}>1 1 + ε 0 ( 1 + δ ) 2 > 1 . Consider the martingale differences Δ m = 1 r n k p ( m / n k ) d m \Delta_m=\frac{1}{r_{n_k}}p(m/n_k)d_m Δ m = r n k 1 p ( m / n k ) d m . Now notice that, ∥ Δ m ∥ ∞ ≤ ( ∥ p ∥ ∞ / r n k ) ⋅ M n k = M k \|\Delta_m\|_\infty\leq(\|p\|_\infty/r_{n_k})\cdot M_{n_k}=M_k ∥ Δ m ∥ ∞ ≤ ( ∥ p ∥ ∞ / r n k ) ⋅ M n k = M k , for all 1 ≤ m ≤ n k 1\leq m\leq n_k 1 ≤ m ≤ n k . Also notice that, ∑ m = 1 n k E m − 1 ( Δ m 2 ) = D k 2 \sum_{m=1}^{n_k}E_{m-1}(\Delta_m^2)=D_k^2 ∑ m = 1 n k E m − 1 ( Δ m 2 ) = D k 2 . Since lim k → ∞ 1 n k ∑ m = 1 n k p ( m / n k ) 2 = ∥ p ∥ 2 \lim_{k\to\infty}\frac{1}{n_k}\sum_{m=1}^{n_k}p(m/n_k)^2=\|p\|_2 lim k → ∞ n k 1 ∑ m = 1 n k p ( m / n k ) 2 = ∥ p ∥ 2 , we have
D k 2 ∼ 1 2 log log n k , as k → ∞ . (17) D_k^2\sim\frac{1}{2\log\log n_k},\qquad\text{as }k\to\infty. \tag*{(17)} D k 2 ∼ 2 log log n k 1 , as k → ∞. (17)
Therefore, ( A k ( n ) ) 1 < n ≤ n k (A_k(n))_{1<n\leq n_k} ( A k ( n ) ) 1 < n ≤ n k satisfies all the conditions of Theorem 3.1 , hence for any λ ∈ [ 0 , 3 ε 0 / M k ] \lambda\in[0,3\varepsilon_0/M_k] λ ∈ [ 0 , 3 ε 0 / M k ] ,
τ ( exp ( λ A k ( n k ) ) ) ≤ exp ( ( 1 + ε 0 ) λ 2 D k 2 2 ) . (18) \tau\left(\exp(\lambda A_k(n_k))\right)\leq\exp\left(\frac{(1+\varepsilon_0)\lambda^2D_k^2}{2}\right). \tag*{(18)} τ ( exp ( λ A k ( n k )) ) ≤ exp ( 2 ( 1 + ε 0 ) λ 2 D k 2 ) . (18)
Let us now estimate the column-wise tail probability. Let q ≥ 4 q\geq4 q ≥ 4 (to be chosen). Take λ > 0 \lambda>0 λ > 0 .
Prob c ( sup n k − 1 < n ≤ n k ∥ A k ( n ) ∥ > 1 + δ ) = Prob c ( sup n k − 1 < n ≤ n k ∥ λ A k ( n ) ∥ > λ ( 1 + δ ) ) ≤ ( λ ( 1 + δ ) ) − q ∥ ( λ A k ( n ) ) n k − 1 < n ≤ n k ∥ L q ( ℓ c ∞ ) q ≤ ( λ ( 1 + δ ) ) − q ( 2 2 q ) q ∥ λ A k ( n k ) ∥ q q . \begin{aligned}
\operatorname{Prob}_{c}\left(\sup_{n_k-1<n\leq n_k}\|A_k(n)\|>1+\delta\right)
&=\operatorname{Prob}_{c}\left(\sup_{n_k-1<n\leq n_k}\|\lambda A_k(n)\|>\lambda(1+\delta)\right)\\
&\leq(\lambda(1+\delta))^{-q}\left\|(\lambda A_k(n))_{n_k-1<n\leq n_k}\right\|_{L^q(\ell^\infty_c)}^q\\
&\leq(\lambda(1+\delta))^{-q}\left(\frac{2^2}{q}\right)^q\|\lambda A_k(n_k)\|_q^q.
\end{aligned} Prob c ( n k − 1 < n ≤ n k sup ∥ A k ( n ) ∥ > 1 + δ ) = Prob c ( n k − 1 < n ≤ n k sup ∥ λ A k ( n ) ∥ > λ ( 1 + δ ) ) ≤ ( λ ( 1 + δ ) ) − q ∥ ( λ A k ( n ) ) n k − 1 < n ≤ n k ∥ L q ( ℓ c ∞ ) q ≤ ( λ ( 1 + δ ) ) − q ( q 2 2 ) q ∥ λ A k ( n k ) ∥ q q .
Since A k ( n k ) A_k(n_k) A k ( n k ) is self-adjoint, applying functional calculus for the function ∣ u ∣ q ≤ q q e − q ( e u + e − u ) |u|^q\leq q^q e^{-q}(e^u+e^{-u}) ∣ u ∣ q ≤ q q e − q ( e u + e − u ) , u ≥ 0 u\geq0 u ≥ 0 we obtain
τ ( λ ∣ A k ( n k ) ∣ q ) ≤ q q e − q [ τ ( e λ A k ( n k ) ) + τ ( e − λ A k ( n k ) ) ] . (19) \tau(\lambda|A_k(n_k)|^q)\leq q^q e^{-q}\left[\tau(e^{\lambda A_k(n_k)})+\tau(e^{-\lambda A_k(n_k)})\right]. \tag*{(19)} τ ( λ ∣ A k ( n k ) ∣ q ) ≤ q q e − q [ τ ( e λ A k ( n k ) ) + τ ( e − λ A k ( n k ) ) ] . (19)
Combining all we have,
Prob c ( sup n k − 1 < n ≤ n k ∥ λ A k ( n ) ∥ > λ ( 1 + δ ) ) ≤ ( λ ( 1 + δ ) ) − q ⋅ ( 2 2 q ) q ⋅ q q e − q ⋅ 2 exp ( ( 1 + ε 0 ) λ 2 D k 2 2 ) ≤ 8 ( q λ ( 1 + δ ) e ) q exp ( ( 1 + ε 0 ) λ 2 D k 2 2 ) . (20) \begin{aligned}
\operatorname{Prob}_{c}\left(\sup_{n_k-1<n\leq n_k}\|\lambda A_k(n)\|>\lambda(1+\delta)\right)
&\leq(\lambda(1+\delta))^{-q}\cdot\left(\frac{2^2}{q}\right)^q\cdot q^q e^{-q}\cdot2\exp\left(\frac{(1+\varepsilon_0)\lambda^2D_k^2}{2}\right)\\
&\leq8\left(\frac{q}{\lambda(1+\delta)e}\right)^q\exp\left(\frac{(1+\varepsilon_0)\lambda^2D_k^2}{2}\right). \tag*{(20)}
\end{aligned} Prob c ( n k − 1 < n ≤ n k sup ∥ λ A k ( n ) ∥ > λ ( 1 + δ ) ) ≤ ( λ ( 1 + δ ) ) − q ⋅ ( q 2 2 ) q ⋅ q q e − q ⋅ 2 exp ( 2 ( 1 + ε 0 ) λ 2 D k 2 ) ≤ 8 ( λ ( 1 + δ ) e q ) q exp ( 2 ( 1 + ε 0 ) λ 2 D k 2 ) . (20)
Now we optimize by setting q = λ ( 1 + δ ) q=\lambda(1+\delta) q = λ ( 1 + δ ) , we get
Prob c ( sup n k − 1 < n ≤ n k ∥ A k ( n ) ∥ > 1 + δ ) ≤ 8 exp ( ( 1 + ε 0 ) λ 2 D k 2 2 − λ ( 1 + δ ) ) . \operatorname{Prob}_{c}\left(\sup_{n_k-1<n\leq n_k}\|A_k(n)\|>1+\delta\right)\leq8\exp\left(\frac{(1+\varepsilon_0)\lambda^2D_k^2}{2}-\lambda(1+\delta)\right). Prob c ( n k − 1 < n ≤ n k sup ∥ A k ( n ) ∥ > 1 + δ ) ≤ 8 exp ( 2 ( 1 + ε 0 ) λ 2 D k 2 − λ ( 1 + δ ) ) .
Since D k → 0 D_k\to0 D k → 0 , choose K 0 ≥ 1 K_0\geq1 K 0 ≥ 1 such that ( 1 + δ ) 2 ( 1 + ε 0 ) D k 2 ≥ 4 \frac{(1+\delta)^2}{(1+\varepsilon_0)D_k^2}\geq4 ( 1 + ε 0 ) D k 2 ( 1 + δ ) 2 ≥ 4 for all k ≥ K 0 k\geq K_0 k ≥ K 0 . Also recall from the assumption that lim n → ∞ L ( n ) M n 2 n = 0 \lim_{n\to\infty}\frac{L(n)M_n^2}{n}=0 lim n → ∞ n L ( n ) M n 2 = 0 . Therefore, M k D k 2 → 0 \frac{M_k}{D_k^2}\to0 D k 2 M k → 0 as k → ∞ k\to\infty k → ∞ . Now choose K ≥ K 0 K\geq K_0 K ≥ K 0 such that M k D k 2 ≤ 3 ε 0 ( 1 + ε 0 ) 1 + δ \frac{M_k}{D_k^2}\leq\frac{3\varepsilon_0(1+\varepsilon_0)}{1+\delta} D k 2 M k ≤ 1 + δ 3 ε 0 ( 1 + ε 0 ) , ∀ k ≥ K \forall k\geq K ∀ k ≥ K .
Now set λ = 1 + δ ( 1 + ε 0 ) D k 2 \lambda=\frac{1+\delta}{(1+\varepsilon_0)D_k^2} λ = ( 1 + ε 0 ) D k 2 1 + δ . Then it follows that q ≥ 4 q\geq4 q ≥ 4 and 0 ≤ λ ≤ 3 ε 0 M k 0\leq\lambda\leq\frac{3\varepsilon_0}{M_k} 0 ≤ λ ≤ M k 3 ε 0 . Now putting the value of λ \lambda λ we have
Prob c ( sup n k − 1 < n ≤ n k ∥ A k ( n ) ∥ > 1 + δ ) ≤ 8 exp ( − ( 1 + δ ) 2 2 ( 1 + ε 0 ) D k 2 ) \operatorname{Prob}_{c}\left(\sup_{n_k-1<n\leq n_k}\|A_k(n)\|>1+\delta\right)\leq8\exp\left(-\frac{(1+\delta)^2}{2(1+\varepsilon_0)D_k^2}\right) Prob c ( n k − 1 < n ≤ n k sup ∥ A k ( n ) ∥ > 1 + δ ) ≤ 8 exp ( − 2 ( 1 + ε 0 ) D k 2 ( 1 + δ ) 2 )
∼ 8 exp ( − ( 1 + δ ) 2 1 + ε 0 log log n k ) , using x r e f ‘ ( 3.12 ) ‘ . \sim8\exp\left(-\frac{(1+\delta)^2}{1+\varepsilon_0}\log\log n_k\right),\quad\text{using }{xref}`(3.12)`. ∼ 8 exp ( − 1 + ε 0 ( 1 + δ ) 2 log log n k ) , using x r e f ‘ ( 3.12 ) ‘.
∼ 8 ( k log c ) − ( 1 + δ ) 2 1 + ε 0 . \sim8(k\log c)^{-\frac{(1+\delta)^2}{1+\varepsilon_0}}. ∼ 8 ( k log c ) − 1 + ε 0 ( 1 + δ ) 2 .
By the choice of ε 0 \varepsilon_0 ε 0 , it follows that
∑ k Prob c ( sup n k − 1 < n ≤ n k ∥ A k ( n ) ∥ > 1 + δ ) < ∞ . \sum_k \operatorname{Prob}_c\left(\sup_{n_{k-1}<n\le n_k}\|A_k(n)\|>1+\delta\right)<\infty. k ∑ Prob c ( n k − 1 < n ≤ n k sup ∥ A k ( n ) ∥ > 1 + δ ) < ∞.
Hence using Lemma 2.5
lim sup n → ∞ A k ( n ) ≤ 1. \limsup_{n\to\infty} A_k(n)\le1. n → ∞ lim sup A k ( n ) ≤ 1.
a.u.
Proof of Theorem A . For any function g g g in [ 0 , 1 ] [0,1] [ 0 , 1 ] we define
Z n ( g ) : = 1 r n ∑ m = 1 n g ( m n ) d m . (21) Z_n(g):=\frac{1}{r_n}\sum_{m=1}^n g\left(\frac{m}{n}\right)d_m. \tag*{(21)} Z n ( g ) := r n 1 m = 1 ∑ n g ( n m ) d m . (21)
Now if ∥ p ∥ 2 = 0 \|p\|_2=0 ∥ p ∥ 2 = 0 , we must have p ≡ 0 p\equiv0 p ≡ 0 . Then (3) is trivial. So, without loss of generality, we may assume that ∥ p ∥ 2 = 1 \|p\|_2=1 ∥ p ∥ 2 = 1 . Also assume that, p ( t ) = ∑ j = 0 l a j t j p(t)=\sum_{j=0}^l a_jt^j p ( t ) = ∑ j = 0 l a j t j and choose ε > 0 \varepsilon>0 ε > 0 . Now, let us define A = ∑ j = 0 l ∣ a j ∣ A=\sum_{j=0}^l|a_j| A = ∑ j = 0 l ∣ a j ∣ . Let c > 1 c>1 c > 1 . Define a sequence n k = [ c k ] + 1 n_k=[c^k]+1 n k = [ c k ] + 1 , k ≥ 1 k\ge1 k ≥ 1 . Now write,
Z n ( p ) = 1 r n k ∑ m = 1 n p ( m n k ) d m ⏟ = : A k ( n ) + ( 1 r n − 1 r n k ) ∑ m = 1 n p ( m n ) d m ⏟ = : B k ( n ) + 1 r n k ∑ m = 1 n [ p ( m n ) − p ( m n k ) ] d m ⏟ = : C k ( n ) . (22) Z_n(p)=\underbrace{\frac{1}{r_{n_k}}\sum_{m=1}^n p\left(\frac{m}{n_k}\right)d_m}_{=:A_k(n)}
+\underbrace{\left(\frac{1}{r_n}-\frac{1}{r_{n_k}}\right)\sum_{m=1}^n p\left(\frac{m}{n}\right)d_m}_{=:B_k(n)}
+\underbrace{\frac{1}{r_{n_k}}\sum_{m=1}^n\left[p\left(\frac{m}{n}\right)-p\left(\frac{m}{n_k}\right)\right]d_m}_{=:C_k(n)}. \tag*{(22)} Z n ( p ) = =: A k ( n ) r n k 1 m = 1 ∑ n p ( n k m ) d m + =: B k ( n ) ( r n 1 − r n k 1 ) m = 1 ∑ n p ( n m ) d m + =: C k ( n ) r n k 1 m = 1 ∑ n [ p ( n m ) − p ( n k m ) ] d m . (22)
Using Corollary 3.4, for each j ∈ { 0 , 1 , … , l } j\in\{0,1,\ldots,l\} j ∈ { 0 , 1 , … , l } we have,
lim sup n → ∞ 1 n j r n ∑ m = 1 n m j d m ≤ ( 2 j + 1 ) − 1 / 2 ≤ 1. \limsup_{n\to\infty}\frac{1}{n^jr_n}\sum_{m=1}^n m^jd_m\le(2j+1)^{-1/2}\le1. n → ∞ lim sup n j r n 1 m = 1 ∑ n m j d m ≤ ( 2 j + 1 ) − 1/2 ≤ 1.
a.u. {#eq:3.18}
Now using Lemma 3.6 it follows that, lim sup k → ∞ ( B k ( n ) + C k ( n ) ) ≤ ( C + 6 A ) ( c − 1 ) \limsup_{k\to\infty}(B_k(n)+C_k(n))\le(C+6A)(c-1) lim sup k → ∞ ( B k ( n ) + C k ( n )) ≤ ( C + 6 A ) ( c − 1 ) , for some constant C > 0 C>0 C > 0 . Now choose c > 1 c>1 c > 1 close enough to 1 that ( c − 1 ) ( 6 A + C ) < ε (c-1)(6A+C)<\varepsilon ( c − 1 ) ( 6 A + C ) < ε . Combining the above steps with Lemma 3.7 it follows that, lim sup n → ∞ ( A k ( n ) + B k ( n ) + C k ( n ) ) ≤ 1 + ε \limsup_{n\to\infty}(A_k(n)+B_k(n)+C_k(n))\le1+\varepsilon lim sup n → ∞ ( A k ( n ) + B k ( n ) + C k ( n )) ≤ 1 + ε . Since ε > 0 \varepsilon>0 ε > 0 chosen arbitrarily, it follows from (3.17) lim sup n → ∞ Z n ( p ) ≤ 1 \limsup_{n\to\infty}Z_n(p)\le1 lim sup n → ∞ Z n ( p ) ≤ 1 . a.u.
Theorem 3.8. Let ( x n ) (x_n) ( x n ) be a non-commutative martingale sequence in a finite von Neumann algebra M \mathcal{M} M with associated difference sequence ( d n ) (d_n) ( d n ) satisfying τ ( d 1 ) = 0 \tau(d_1)=0 τ ( d 1 ) = 0 , E n − 1 ( d n ∗ d n ) = E n − 1 ( d n d n ∗ ) = 1 E_{n-1}(d_n^*d_n)=E_{n-1}(d_nd_n^*)=1 E n − 1 ( d n ∗ d n ) = E n − 1 ( d n d n ∗ ) = 1 and max m ≤ n ∥ d m ∥ ≤ M n \max_{m\le n}\|d_m\|\le M_n max m ≤ n ∥ d m ∥ ≤ M n , where ( M n ) (M_n) ( M n ) is a sequence of positive reals such that lim n → ∞ L ( n ) M n 2 / n = 0 \lim_{n\to\infty}L(n)M_n^2/n=0 lim n → ∞ L ( n ) M n 2 / n = 0 . Let p p p be a polynomial on [ 0 , 1 ] [0,1] [ 0 , 1 ] , then
lim sup n → ∞ 1 2 n L ( n ) ∑ m = 1 n p ( m / n ) d m ≤ ∥ p ∥ 2 . \limsup_{n\to\infty}\frac{1}{\sqrt{2nL(n)}}\sum_{m=1}^n p(m/n)d_m\le\|p\|_2. n → ∞ lim sup 2 n L ( n ) 1 m = 1 ∑ n p ( m / n ) d m ≤ ∥ p ∥ 2 . a.u. {#eq:3.19}
Proof. We employ the standard 2 × 2 2\times2 2 × 2 matrix trick, as in [19 Theorem 3.3] , to reduce the non-self-adjoint case to the self-adjoint setting. Let M ~ = M ⊗ ‾ M 2 ( C ) \widetilde{\mathcal{M}}=\mathcal{M}\overline{\otimes}M_2(\mathbb{C}) M = M ⊗ M 2 ( C ) , M ~ k = M k ⊗ ‾ M 2 ( C ) \widetilde{\mathcal{M}}_k=\mathcal{M}_k\overline{\otimes}M_2(\mathbb{C}) M k = M k ⊗ M 2 ( C ) , k ≥ 1 k\ge1 k ≥ 1 . For each n ≥ 0 n\ge0 n ≥ 0 , define
d ~ n = ( 0 d n d n ∗ 0 ) ∈ M ~ . \widetilde{d}_n=
\begin{pmatrix}
0 & d_n\\
d_n^* & 0
\end{pmatrix}\in\widetilde{\mathcal{M}}. d n = ( 0 d n ∗ d n 0 ) ∈ M .
@page 15/33 starts_mid=1 ends_mid=1
∼ 8 ( k log c ) − ( 1 + δ ) 2 1 + ε 0 . \sim8(k\log c)^{-\frac{(1+\delta)^2}{1+\varepsilon_0}}. ∼ 8 ( k log c ) − 1 + ε 0 ( 1 + δ ) 2 .
By the choice of ε 0 \varepsilon_0 ε 0 , it follows that
∑ k Prob c ( sup n k − 1 < n ≤ n k ∥ A k ( n ) ∥ > 1 + δ ) < ∞ . \sum_k \operatorname{Prob}_c\left(\sup_{n_{k-1}<n\le n_k}\|A_k(n)\|>1+\delta\right)<\infty. k ∑ Prob c ( n k − 1 < n ≤ n k sup ∥ A k ( n ) ∥ > 1 + δ ) < ∞.
Hence using Lemma 2.5
lim sup n → ∞ A k ( n ) ≤ 1. \limsup_{n\to\infty} A_k(n)\le1. n → ∞ lim sup A k ( n ) ≤ 1.
a.u.
Proof of Theorem A . For any function g g g in [ 0 , 1 ] [0,1] [ 0 , 1 ] we define
Z n ( g ) : = 1 r n ∑ m = 1 n g ( m n ) d m . (23) Z_n(g):=\frac{1}{r_n}\sum_{m=1}^n g\left(\frac{m}{n}\right)d_m. \tag*{(23)} Z n ( g ) := r n 1 m = 1 ∑ n g ( n m ) d m . (23)
Now if ∥ p ∥ 2 = 0 \|p\|_2=0 ∥ p ∥ 2 = 0 , we must have p ≡ 0 p\equiv0 p ≡ 0 . Then (3) is trivial. So, without loss of generality, we may assume that ∥ p ∥ 2 = 1 \|p\|_2=1 ∥ p ∥ 2 = 1 . Also assume that, p ( t ) = ∑ j = 0 l a j t j p(t)=\sum_{j=0}^l a_jt^j p ( t ) = ∑ j = 0 l a j t j and choose ε > 0 \varepsilon>0 ε > 0 . Now, let us define A = ∑ j = 0 l ∣ a j ∣ A=\sum_{j=0}^l|a_j| A = ∑ j = 0 l ∣ a j ∣ . Let c > 1 c>1 c > 1 . Define a sequence n k = [ c k ] + 1 n_k=[c^k]+1 n k = [ c k ] + 1 , k ≥ 1 k\ge1 k ≥ 1 . Now write,
Z n ( p ) = 1 r n k ∑ m = 1 n p ( m n k ) d m ⏟ = : A k ( n ) + ( 1 r n − 1 r n k ) ∑ m = 1 n p ( m n ) d m ⏟ = : B k ( n ) + 1 r n k ∑ m = 1 n [ p ( m n ) − p ( m n k ) ] d m ⏟ = : C k ( n ) . (24) Z_n(p)=\underbrace{\frac{1}{r_{n_k}}\sum_{m=1}^n p\left(\frac{m}{n_k}\right)d_m}_{=:A_k(n)}
+\underbrace{\left(\frac{1}{r_n}-\frac{1}{r_{n_k}}\right)\sum_{m=1}^n p\left(\frac{m}{n}\right)d_m}_{=:B_k(n)}
+\underbrace{\frac{1}{r_{n_k}}\sum_{m=1}^n\left[p\left(\frac{m}{n}\right)-p\left(\frac{m}{n_k}\right)\right]d_m}_{=:C_k(n)}. \tag*{(24)} Z n ( p ) = =: A k ( n ) r n k 1 m = 1 ∑ n p ( n k m ) d m + =: B k ( n ) ( r n 1 − r n k 1 ) m = 1 ∑ n p ( n m ) d m + =: C k ( n ) r n k 1 m = 1 ∑ n [ p ( n m ) − p ( n k m ) ] d m . (24)
Using Corollary 3.4, for each j ∈ { 0 , 1 , … , l } j\in\{0,1,\ldots,l\} j ∈ { 0 , 1 , … , l } we have,
lim sup n → ∞ 1 n j r n ∑ m = 1 n m j d m ≤ ( 2 j + 1 ) − 1 / 2 ≤ 1. \limsup_{n\to\infty}\frac{1}{n^jr_n}\sum_{m=1}^n m^jd_m\le(2j+1)^{-1/2}\le1. n → ∞ lim sup n j r n 1 m = 1 ∑ n m j d m ≤ ( 2 j + 1 ) − 1/2 ≤ 1.
a.u. {#eq:3.18}
Now using Lemma 3.6 it follows that, lim sup k → ∞ ( B k ( n ) + C k ( n ) ) ≤ ( C + 6 A ) ( c − 1 ) \limsup_{k\to\infty}(B_k(n)+C_k(n))\le(C+6A)(c-1) lim sup k → ∞ ( B k ( n ) + C k ( n )) ≤ ( C + 6 A ) ( c − 1 ) , for some constant C > 0 C>0 C > 0 . Now choose c > 1 c>1 c > 1 close enough to 1 that ( c − 1 ) ( 6 A + C ) < ε (c-1)(6A+C)<\varepsilon ( c − 1 ) ( 6 A + C ) < ε . Combining the above steps with Lemma 3.7 it follows that, lim sup n → ∞ ( A k ( n ) + B k ( n ) + C k ( n ) ) ≤ 1 + ε \limsup_{n\to\infty}(A_k(n)+B_k(n)+C_k(n))\le1+\varepsilon lim sup n → ∞ ( A k ( n ) + B k ( n ) + C k ( n )) ≤ 1 + ε . Since ε > 0 \varepsilon>0 ε > 0 chosen arbitrarily, it follows from (3.17) lim sup n → ∞ Z n ( p ) ≤ 1 \limsup_{n\to\infty}Z_n(p)\le1 lim sup n → ∞ Z n ( p ) ≤ 1 . a.u.
Theorem 3.8. Let ( x n ) (x_n) ( x n ) be a non-commutative martingale sequence in a finite von Neumann algebra M \mathcal{M} M with associated difference sequence ( d n ) (d_n) ( d n ) satisfying τ ( d 1 ) = 0 \tau(d_1)=0 τ ( d 1 ) = 0 , E n − 1 ( d n ∗ d n ) = E n − 1 ( d n d n ∗ ) = 1 E_{n-1}(d_n^*d_n)=E_{n-1}(d_nd_n^*)=1 E n − 1 ( d n ∗ d n ) = E n − 1 ( d n d n ∗ ) = 1 and max m ≤ n ∥ d m ∥ ≤ M n \max_{m\le n}\|d_m\|\le M_n max m ≤ n ∥ d m ∥ ≤ M n , where ( M n ) (M_n) ( M n ) is a sequence of positive reals such that lim n → ∞ L ( n ) M n 2 / n = 0 \lim_{n\to\infty}L(n)M_n^2/n=0 lim n → ∞ L ( n ) M n 2 / n = 0 . Let p p p be a polynomial on [ 0 , 1 ] [0,1] [ 0 , 1 ] , then
lim sup n → ∞ 1 2 n L ( n ) ∑ m = 1 n p ( m / n ) d m ≤ ∥ p ∥ 2 . \limsup_{n\to\infty}\frac{1}{\sqrt{2nL(n)}}\sum_{m=1}^n p(m/n)d_m\le\|p\|_2. n → ∞ lim sup 2 n L ( n ) 1 m = 1 ∑ n p ( m / n ) d m ≤ ∥ p ∥ 2 . a.u. {#eq:3.19}
Proof. We employ the standard 2 × 2 2\times2 2 × 2 matrix trick, as in [19 Theorem 3.3] , to reduce the non-self-adjoint case to the self-adjoint setting. Let M ~ = M ⊗ ‾ M 2 ( C ) \widetilde{\mathcal{M}}=\mathcal{M}\overline{\otimes}M_2(\mathbb{C}) M = M ⊗ M 2 ( C ) , M ~ k = M k ⊗ ‾ M 2 ( C ) \widetilde{\mathcal{M}}_k=\mathcal{M}_k\overline{\otimes}M_2(\mathbb{C}) M k = M k ⊗ M 2 ( C ) , k ≥ 1 k\ge1 k ≥ 1 . For each n ≥ 0 n\ge0 n ≥ 0 , define
d ~ n = ( 0 d n d n ∗ 0 ) ∈ M ~ . \widetilde{d}_n=
\begin{pmatrix}
0 & d_n\\
d_n^* & 0
\end{pmatrix}\in\widetilde{\mathcal{M}}. d n = ( 0 d n ∗ d n 0 ) ∈ M .
Then d ~ n \widetilde{d}_n d n is self-adjoint. Since ( d n ) (d_n) ( d n ) is a martingale difference sequence with respect to ( M n , E n ) (\mathcal{M}_n,\mathcal{E}_n) ( M n , E n ) , it follows immediately that ( d ~ n ) (\widetilde{d}_n) ( d n ) is a self-adjoint martingale difference sequence with respect to ( M ~ n , E ~ n ) (\widetilde{\mathcal{M}}_n,\widetilde{\mathcal{E}}_n) ( M n , E n ) , where for all n ≥ 0 n \ge0 n ≥ 0
E ~ n ( d ~ n + 1 ) = ( 0 E n ( d n + 1 ) E n ( d n + 1 ∗ ) 0 ) . \widetilde{\mathcal{E}}_n(\widetilde{d}_{n+1}) =
\begin{pmatrix}
0 & E_n(d_{n+1}) \\
E_n(d_{n+1}^{*}) & 0
\end{pmatrix}. E n ( d n + 1 ) = ( 0 E n ( d n + 1 ∗ ) E n ( d n + 1 ) 0 ) .
Also, for n ≥ 1 n \ge1 n ≥ 1 , ∥ d ~ n ∥ ∞ = ∥ d n ∥ ∞ ≤ M n \lVert\widetilde{d}_n\rVert_\infty= \lVert d_n\rVert_\infty\le M_n ∥ d n ∥ ∞ = ∥ d n ∥ ∞ ≤ M n . Furthermore, notice that
d ~ n 2 = ( d n d n ∗ 0 0 d n ∗ d n ) , so E ~ n − 1 ( d ~ n 2 ) = ( 1 0 0 1 ) = 1 M ~ . \widetilde{d}_n^2 =
\begin{pmatrix}
d_nd_n^{*} & 0 \\
0 & d_n^{*}d_n
\end{pmatrix},
\quad\text{so}\quad
\widetilde{\mathcal{E}}_{n-1}(\widetilde{d}_n^2) =
\begin{pmatrix}
1 & 0 \\
0 & 1
\end{pmatrix}
=1_{\widetilde{\mathcal{M}}}. d n 2 = ( d n d n ∗ 0 0 d n ∗ d n ) , so E n − 1 ( d n 2 ) = ( 1 0 0 1 ) = 1 M .
Therefore, all assumptions of Theorem A hold for ( d ~ n ) (\widetilde{d}_n) ( d n ) . Hence,
lim sup n → ∞ 1 2 n L ( n ) ∑ m = 1 n p ( m n ) d ~ m ≤ a . u . ∥ p ∥ 2 in M ~ . \limsup_{n\to\infty}\frac{1}{\sqrt{2nL(n)}}\sum_{m=1}^{n}p\left(\frac{m}{n}\right)\widetilde{d}_m
\leq_{\mathrm{a.u.}}\lVert p\rVert_2
\quad\text{in }\widetilde{\mathcal{M}}. n → ∞ lim sup 2 n L ( n ) 1 m = 1 ∑ n p ( n m ) d m ≤ a.u. ∥ p ∥ 2 in M .
Observe that
∑ m = 1 n p ( m n ) d ~ m = ( 0 ∑ m = 1 n p ( m n ) d m ∑ m = 1 n p ( m n ) d m ∗ 0 ) . \sum_{m=1}^{n}p\left(\frac{m}{n}\right)\widetilde{d}_m
=
\begin{pmatrix}
0 & \sum_{m=1}^{n}p\left(\frac{m}{n}\right)d_m \\
\sum_{m=1}^{n}p\left(\frac{m}{n}\right)d_m^{*} & 0
\end{pmatrix}. m = 1 ∑ n p ( n m ) d m = ( 0 ∑ m = 1 n p ( n m ) d m ∗ ∑ m = 1 n p ( n m ) d m 0 ) .
Let r ~ = ( 0 0 1 0 ) ∈ M ~ \widetilde{r}=\begin{pmatrix}0&0\\1&0\end{pmatrix}\in\widetilde{\mathcal{M}} r = ( 0 1 0 0 ) ∈ M . Then
r ~ ( ∑ m = 1 n p ( m n ) d ~ m ) = ( 0 0 0 ∑ m = 1 n p ( m n ) d m ) . \widetilde{r}\left(\sum_{m=1}^{n}p\left(\frac{m}{n}\right)\widetilde{d}_m\right)
=
\begin{pmatrix}
0 & 0 \\
0 & \sum_{m=1}^{n}p\left(\frac{m}{n}\right)d_m
\end{pmatrix}. r ( m = 1 ∑ n p ( n m ) d m ) = ( 0 0 0 ∑ m = 1 n p ( n m ) d m ) .
Consequently,
lim sup n → ∞ 1 2 n L ( n ) r ~ ( ∑ m = 1 n p ( m n ) d ~ m ) ≤ a . u . ∥ p ∥ 2 . \limsup_{n\to\infty}\frac{1}{\sqrt{2nL(n)}}\widetilde{r}\left(\sum_{m=1}^{n}p\left(\frac{m}{n}\right)\widetilde{d}_m\right)
\leq_{\mathrm{a.u.}}\lVert p\rVert_2. n → ∞ lim sup 2 n L ( n ) 1 r ( m = 1 ∑ n p ( n m ) d m ) ≤ a.u. ∥ p ∥ 2 .
Since almost uniform boundedness does not depend on the ambient von Neumann algebra (see [19 Remark 2.3] ) we obtain,
lim sup n → ∞ 1 2 n L ( n ) ∑ m = 1 n p ( m / n ) d m ≤ a . u . ∥ p ∥ 2 . \limsup_{n\to\infty}\frac{1}{\sqrt{2nL(n)}}\sum_{m=1}^{n}p(m/n)d_m
\leq_{\mathrm{a.u.}}\lVert p\rVert_2. n → ∞ lim sup 2 n L ( n ) 1 m = 1 ∑ n p ( m / n ) d m ≤ a.u. ∥ p ∥ 2 .
Lemma 3.9. Let ( d n ) n ≥ 1 (d_n)_{n\ge1} ( d n ) n ≥ 1 be a sequence of self-adjoint random variables in ( M , τ ) (\mathcal{M},\tau) ( M , τ ) . Let { b n } \{b_n\} { b n } be a non-decreasing sequence of positive real numbers such that
lim sup n → ∞ 1 b n ∑ m = 1 n d m ≤ a . u . 1. \limsup_{n\to\infty}\frac{1}{b_n}\sum_{m=1}^{n}d_m\leq_{\mathrm{a.u.}}1. n → ∞ lim sup b n 1 m = 1 ∑ n d m ≤ a.u. 1.
Let { a n m } m , n ≥ 1 \{a_{nm}\}_{m,n\geq1} { a nm } m , n ≥ 1 be a double sequence of real numbers such that a n m = 0 a_{nm}=0 a nm = 0 , whenever m > n m>n m > n . Assume moreover that, there exists L ∈ R L\in\mathbb{R} L ∈ R such that lim n → ∞ a n m = L \lim_{n\to\infty}a_{nm}=L lim n → ∞ a nm = L for every fixed m ≥ 1 m\geq1 m ≥ 1 . Then for any positive sequence { β n } \{\beta_n\} { β n } with lim inf n → ∞ β n > 0 \liminf_{n\to\infty}\beta_n>0 lim inf n → ∞ β n > 0 ,
lim sup n → ∞ 1 β n ∑ m = 1 n a n m d m ≤ lim sup n → ∞ a . u . b n β n ∑ m = 1 n ∣ a n m − a n , m + 1 ∣ . \limsup_{n\to\infty}\frac{1}{\beta_n}\sum_{m=1}^{n}a_{nm}d_m
\leq
\underset{\mathrm{a.u.}}{\limsup_{n\to\infty}}\frac{b_n}{\beta_n}\sum_{m=1}^{n}|a_{nm}-a_{n,m+1}|. n → ∞ lim sup β n 1 m = 1 ∑ n a nm d m ≤ a.u. n → ∞ lim sup β n b n m = 1 ∑ n ∣ a nm − a n , m + 1 ∣.
Proof. Define x n : = ∑ m = 1 n d m x_n:=\sum_{m=1}^{n}d_m x n := ∑ m = 1 n d m and let ε , δ > 0 \varepsilon,\delta>0 ε , δ > 0 be arbitrary. By (3.20), there exists a projection e ∈ M e\in\mathcal{M} e ∈ M with τ ( 1 − e ) < ε \tau(1-e)<\varepsilon τ ( 1 − e ) < ε and N = N ( e ) ∈ N N=N(e)\in\mathbb{N} N = N ( e ) ∈ N such that
∥ x m e ∥ ≤ ( 1 + δ ) b m for all m ≥ N . (25) \|x_me\|\leq(1+\delta)b_m\quad\text{for all }m\geq N. \tag*{(25)} ∥ x m e ∥ ≤ ( 1 + δ ) b m for all m ≥ N . (25)
Since a n m a_{nm} a nm are real scalars, we have ∑ m = 1 n a n m d m = ∑ m = 1 n ( a n m − a n , m + 1 ) x m \sum_{m=1}^{n}a_{nm}d_m=\sum_{m=1}^{n}(a_{nm}-a_{n,m+1})x_m ∑ m = 1 n a nm d m = ∑ m = 1 n ( a nm − a n , m + 1 ) x m , for n ≥ 1 n\geq1 n ≥ 1 , where we used x 0 = 0 x_0=0 x 0 = 0 and a n , n + 1 = 0 a_{n,n+1}=0 a n , n + 1 = 0 . Now notice that, since b n b_n b n is non-decreasing
∥ ( ∑ m = 1 n a n m d m ) e ∥ ∞ ≤ ∑ m = 1 n ∣ a n m − a n , m + 1 ∣ ⋅ ∥ x m e ∥ ∞ ≤ R N + ( 1 + δ ) ∑ m = N n ∣ a n m − a n , m + 1 ∣ b m , using xref‘(3.21)‘ ≤ R N + ( 1 + δ ) b n ∑ m = N n ∣ a n m − a n , m + 1 ∣ , (26) \begin{aligned}
\left\|\left(\sum_{m=1}^{n}a_{nm}d_m\right)e\right\|_{\infty}
\leq\sum_{m=1}^{n}|a_{nm}-a_{n,m+1}|\cdot\|x_me\|_{\infty} \\
&\leq R_N+(1+\delta)\sum_{m=N}^{n}|a_{nm}-a_{n,m+1}|b_m,\quad\text{using {xref}`(3.21)`} \\
&\leq R_N+(1+\delta)b_n\sum_{m=N}^{n}|a_{nm}-a_{n,m+1}|, \tag*{(26)}
\end{aligned} ( m = 1 ∑ n a nm d m ) e ∞ ≤ m = 1 ∑ n ∣ a nm − a n , m + 1 ∣ ⋅ ∥ x m e ∥ ∞ ≤ R N + ( 1 + δ ) m = N ∑ n ∣ a nm − a n , m + 1 ∣ b m , using xref ‘(3.21)‘ ≤ R N + ( 1 + δ ) b n m = N ∑ n ∣ a nm − a n , m + 1 ∣ , (26)
where R N = ∑ m = 1 N − 1 ∣ a n m − a n , m + 1 ∣ ⋅ ∥ x m e ∥ ∞ R_N=\sum_{m=1}^{N-1}|a_{nm}-a_{n,m+1}|\cdot\|x_me\|_{\infty} R N = ∑ m = 1 N − 1 ∣ a nm − a n , m + 1 ∣ ⋅ ∥ x m e ∥ ∞ . Now observe that, since lim n → ∞ a n m = L \lim_{n\to\infty}a_{nm}=L lim n → ∞ a nm = L for every fixed m ≥ 1 m\geq1 m ≥ 1 and lim inf n → ∞ β n > 0 \liminf_{n\to\infty}\beta_n>0 lim inf n → ∞ β n > 0 , we have lim n → ∞ ∣ a n m − a n , m + 1 ∣ / β n = 0 \lim_{n\to\infty}|a_{nm}-a_{n,m+1}|/\beta_n=0 lim n → ∞ ∣ a nm − a n , m + 1 ∣/ β n = 0 , for every fixed m ≥ 1 m\geq1 m ≥ 1 implies lim n → ∞ R N / β n = 0 \lim_{n\to\infty}R_N/\beta_n=0 lim n → ∞ R N / β n = 0 . It follows from (26) ,
lim sup n → ∞ β n − 1 ∥ ( ∑ m = 1 n a n m d m ) e ∥ ∞ ≤ ( 1 + δ ) lim sup n → ∞ β n − 1 b n ∑ m = N n ∣ a n m − a n , m + 1 ∣ . \underset{n\to\infty}{\limsup}\,\beta_n^{-1}\left\|\left(\sum_{m=1}^{n}a_{nm}d_m\right)e\right\|_{\infty}
\leq(1+\delta)\underset{n\to\infty}{\limsup}\,\beta_n^{-1}b_n\sum_{m=N}^{n}|a_{nm}-a_{n,m+1}|. n → ∞ lim sup β n − 1 ( m = 1 ∑ n a nm d m ) e ∞ ≤ ( 1 + δ ) n → ∞ lim sup β n − 1 b n m = N ∑ n ∣ a nm − a n , m + 1 ∣.
This completes the proof of the lemma. □
Lemma 3.10
Lemma 3.10. Let ( M , τ ) (\mathcal{M},\tau) ( M , τ ) and ( d n ) (d_n) ( d n ) be same as in Lemma 3.9 and lim sup a . u . 1 ( 2 n log log n ) 1 / 2 ∑ m = 1 n d m ≤ 1 \underset{\mathrm{a.u.}}{\limsup}\frac{1}{(2n\log\log n)^{1/2}}\sum_{m=1}^{n}d_m\leq1 a.u. lim sup ( 2 n l o g l o g n ) 1/2 1 ∑ m = 1 n d m ≤ 1 . Then for any function g : [ 0 , 1 ] → R g:[0,1]\to\mathbb{R} g : [ 0 , 1 ] → R of bounded variation with total variation V g V_g V g ,
lim sup n → ∞ a . u . 1 ( 2 n log log n ) 1 / 2 ∑ m = 1 n g ( m / n ) d m ≤ V g + ∣ g ( 1 ) ∣ . (27) \underset{\mathrm{a.u.}}{\limsup_{n\to\infty}}\frac{1}{(2n\log\log n)^{1/2}}\sum_{m=1}^{n}g(m/n)d_m
\leq V_g+|g(1)|. \tag*{(27)} a.u. n → ∞ lim sup ( 2 n log log n ) 1/2 1 m = 1 ∑ n g ( m / n ) d m ≤ V g + ∣ g ( 1 ) ∣. (27)
Proof. Define the double sequence of real numbers { a n m } m , n ≥ 1 \{a_{nm}\}_{m,n\geq1} { a nm } m , n ≥ 1 by
a n m : = { g ( m / n ) if 1 ≤ m ≤ n , 0 if m > n . (28) a_{nm}:=
\begin{cases}
g(m/n) & \text{if }1\leq m\leq n,\\
0 & \text{if }m>n.
\end{cases} \tag*{(28)} a nm := { g ( m / n ) 0 if 1 ≤ m ≤ n , if m > n . (28)
Set b n = β n : = ( 2 n log log n ) 1 / 2 b_n=\beta_n := (2n\log\log n)^{1/2} b n = β n := ( 2 n log log n ) 1/2 . For each fixed m ≥ 1 m\ge1 m ≥ 1 , as n → ∞ n\to\infty n → ∞ , we have a n m = g ( m / n ) → g ( 0 ) a_{nm}=g(m/n)\to g(0) a nm = g ( m / n ) → g ( 0 ) . So condition of Lemma 3.9 holds with L = g ( 0 ) L=g(0) L = g ( 0 ) . Since a n m = 0 a_{nm}=0 a nm = 0 for all m > n m>n m > n , and b n = β n = ( 2 n log log n ) 1 / 2 b_n=\beta_n=(2n\log\log n)^{1/2} b n = β n = ( 2 n log log n ) 1/2 , the hypotheses of Lemma 3.9 are satisfied. Therefore, we have
lim sup n → ∞ 1 ( 2 n log log n ) 1 / 2 ∑ m = 1 n g ( m / n ) d m ≤ a . u . lim sup n → ∞ ∑ m = 1 n ∣ a n m − a n , m + 1 ∣ . (29) \limsup_{n\to\infty}\frac{1}{(2n\log\log n)^{1/2}}\sum_{m=1}^{n}g(m/n)d_m\underset{\mathrm{a.u.}}{\leq}\limsup_{n\to\infty}\sum_{m=1}^{n}|a_{nm}-a_{n,m+1}|. \tag*{(29)} n → ∞ lim sup ( 2 n log log n ) 1/2 1 m = 1 ∑ n g ( m / n ) d m a.u. ≤ n → ∞ lim sup m = 1 ∑ n ∣ a nm − a n , m + 1 ∣. (29)
Therefore, for every n ≥ 1 n\geq1 n ≥ 1 , ∑ m = 1 n ∣ a n m − a n , m + 1 ∣ ≤ V g + ∣ g ( 1 ) ∣ \sum_{m=1}^{n}|a_{nm}-a_{n,m+1}|\leq V_g+|g(1)| ∑ m = 1 n ∣ a nm − a n , m + 1 ∣ ≤ V g + ∣ g ( 1 ) ∣ . Since this bound holds for every n ≥ 1 n\geq1 n ≥ 1 , then by (29) we have
lim sup n → ∞ 1 ( 2 n log log n ) 1 / 2 ∑ m = 1 n g ( m / n ) d m ≤ a . u . V g + ∣ g ( 1 ) ∣ . \limsup_{n\to\infty}\frac{1}{(2n\log\log n)^{1/2}}\sum_{m=1}^{n}g(m/n)d_m\underset{\mathrm{a.u.}}{\leq}V_g+|g(1)|. n → ∞ lim sup ( 2 n log log n ) 1/2 1 m = 1 ∑ n g ( m / n ) d m a.u. ≤ V g + ∣ g ( 1 ) ∣.
Corollary 3.11. Let ( M , τ ) (\mathcal{M},\tau) ( M , τ ) and ( d n ) (d_n) ( d n ) be same as Theorem 3.8. Let f f f be a function on [ 0 , 1 ] [0,1] [ 0 , 1 ] such that a sequence ( p m ) (p_m) ( p m ) of polynomials exists such that p m ( 1 ) → f ( 1 ) p_m(1)\to f(1) p m ( 1 ) → f ( 1 ) and V p m − f → 0 V_{p_m-f}\to0 V p m − f → 0 as m → ∞ m\to\infty m → ∞ . Then
lim sup n → ∞ 1 2 n L ( n ) ∑ m = 1 n f ( m / n ) d m ≤ a . u . ∥ f ∥ 2 . (30) \limsup_{n\to\infty}\frac{1}{\sqrt{2nL(n)}}\sum_{m=1}^{n}f(m/n)d_m\underset{\mathrm{a.u.}}{\leq}\|f\|_2. \tag*{(30)} n → ∞ lim sup 2 n L ( n ) 1 m = 1 ∑ n f ( m / n ) d m a.u. ≤ ∥ f ∥ 2 . (30) In particular, (30) holds if f ( t ) = ∑ j = 0 ∞ a j t j f(t)=\sum_{j=0}^{\infty}a_jt^j f ( t ) = ∑ j = 0 ∞ a j t j is a power series with either
(i) radius of convergence greater than 1 1 1 , or
(ii) lim sup n → ∞ ∣ a n ∣ 1 / n = 1 \limsup_{n\to\infty}|a_n|^{1/n}=1 lim sup n → ∞ ∣ a n ∣ 1/ n = 1 , ∑ j = 0 ∞ ∣ a j ∣ < ∞ \sum_{j=0}^{\infty}|a_j|<\infty ∑ j = 0 ∞ ∣ a j ∣ < ∞ .
Proof. First notice that, for any t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and m ≥ 1 m\geq1 m ≥ 1 , ∣ p m ( t ) − f ( t ) ∣ ≤ V f − p m + ∣ f ( 1 ) − p m ( 1 ) ∣ |p_m(t)-f(t)|\leq V_{f-p_m}+|f(1)-p_m(1)| ∣ p m ( t ) − f ( t ) ∣ ≤ V f − p m + ∣ f ( 1 ) − p m ( 1 ) ∣ , see [Apo57]. By the hypothesis, V f − p m ⟶ m → ∞ 0 V_{f-p_m}\mathop{\longrightarrow}_{m\to\infty}0 V f − p m ⟶ m → ∞ 0 and ∣ f ( 1 ) − p m ( 1 ) ∣ ⟶ m → ∞ 0 |f(1)-p_m(1)|\mathop{\longrightarrow}_{m\to\infty}0 ∣ f ( 1 ) − p m ( 1 ) ∣ ⟶ m → ∞ 0 . Hence we have, ∣ p m ( t ) − f ( t ) ∣ → 0 |p_m(t)-f(t)|\to0 ∣ p m ( t ) − f ( t ) ∣ → 0 , as m → ∞ m\to\infty m → ∞ , for each t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . Therefore p m → f p_m\to f p m → f uniformly, so ∥ p m ∥ 2 → ∥ f ∥ 2 \|p_m\|_2\to\|f\|_2 ∥ p m ∥ 2 → ∥ f ∥ 2 as m → ∞ m\to\infty m → ∞ .
Fix m ≥ 1 m\geq1 m ≥ 1 . Using (3.16) we decompose,
Z n ( f ) = Z n ( f − p m ) + Z n ( p m ) . (31) Z_n(f)=Z_n(f-p_m)+Z_n(p_m). \tag*{(31)} Z n ( f ) = Z n ( f − p m ) + Z n ( p m ) . (31)
Now recall from Theorem 3.8,
lim sup n → ∞ Z n ( p m ) ≤ a . u . ∥ p m ∥ 2 . (32) \limsup_{n\to\infty}Z_n(p_m)\underset{\mathrm{a.u.}}{\leq}\|p_m\|_2. \tag*{(32)} n → ∞ lim sup Z n ( p m ) a.u. ≤ ∥ p m ∥ 2 . (32)
Since f f f is a BV function and p m p_m p m is a polynomial (hence also BV), applying Lemma 3.10 to f − p m f-p_m f − p m we have
lim sup n → ∞ Z n ( f − p m ) ≤ a . u . V f − p m + ∣ f ( 1 ) − p m ( 1 ) ∣ . (33) \limsup_{n\to\infty}Z_n(f-p_m)\underset{\mathrm{a.u.}}{\leq}V_{f-p_m}+|f(1)-p_m(1)|. \tag*{(33)} n → ∞ lim sup Z n ( f − p m ) a.u. ≤ V f − p m + ∣ f ( 1 ) − p m ( 1 ) ∣. (33)
Now using (32) and (33) to (31) we get lim sup n → ∞ Z n ( f ) ≤ a . u . V f − p m + ∣ f ( 1 ) − p m ( 1 ) ∣ + ∥ p m ∥ 2 \limsup_{n\to\infty}Z_n(f)\underset{\mathrm{a.u.}}{\leq}V_{f-p_m}+|f(1)-p_m(1)|+\|p_m\|_2 lim sup n → ∞ Z n ( f ) a.u. ≤ V f − p m + ∣ f ( 1 ) − p m ( 1 ) ∣ + ∥ p m ∥ 2 . Letting m ↑ ∞ m\uparrow\infty m ↑ ∞ , (30) obtains.
Now, let f ( t ) = ∑ j = 0 ∞ a j t j f(t)=\sum_{j=0}^{\infty}a_jt^j f ( t ) = ∑ j = 0 ∞ a j t j and suppose that either (i) or (ii) holds. For each m ≥ 1 m\geq1 m ≥ 1 , define p m ( t ) = ∑ j = 0 m a j t j p_m(t)=\sum_{j=0}^{m}a_jt^j p m ( t ) = ∑ j = 0 m a j t j . Then p m ( 1 ) → f ( 1 ) p_m(1)\to f(1) p m ( 1 ) → f ( 1 ) and V p m − f → 0 V_{p_m-f}\to0 V p m − f → 0 as m → ∞ m\to\infty m → ∞ [24 ] . Consequently, (30) holds. □
non-commutative Law of Logarithm for Deterministic Weighted Sums
In this section, we prove Theorem B, which establishes a weighted analogue of the non-commutative law of iterated logarithm. We aim to investigate the almost sure asymptotic behavior of weighted sums of non-commutative random variables under some tail conditions on their distributions. Let ( x k ) k ≥ 1 (x_k)_{k \ge1} ( x k ) k ≥ 1 be a i.i.d. sequence of mean zero self-adjoint random variables of ( M , τ ) (\mathcal{M},\tau) ( M , τ ) . Let A = { a n , k : 1 ≤ k ≤ n , n ≥ 1 } A = \{a_{n,k}: 1 \le k \le n,\ n \ge1\} A = { a n , k : 1 ≤ k ≤ n , n ≥ 1 } be a triangular array of real numbers and set
D : = { A = ( a n , k ) : K ( A ) : = sup n , k ∣ a n , k ∣ < ∞ } . (34) \mathcal{D} := \left\{A = (a_{n,k}): K(A) := \sup_{n,k} |a_{n,k}| < \infty\right\}. \tag*{(34)} D := { A = ( a n , k ) : K ( A ) := n , k sup ∣ a n , k ∣ < ∞ } . (34)
Throughout, for A ∈ D A \in\mathcal{D} A ∈ D we write K : = K ( A ) = sup n , k ∣ a n , k ∣ < ∞ K := K(A) = \sup_{n,k} |a_{n,k}| < \infty K := K ( A ) = sup n , k ∣ a n , k ∣ < ∞ .
To generalize the asymptotic estimate more than the law of iterated logarithm, we normalise the averages by a real sequence obtained from a regularly varying function. In the following we state the required definition and few of the properties which are useful in our case.
Definition 4.1. A positive measurable function B : ( 0 , ∞ ) → ( 0 , ∞ ) B : (0,\infty) \to(0,\infty) B : ( 0 , ∞ ) → ( 0 , ∞ ) is said to be regularly varying at infinity with index α ∈ R \alpha\in\mathbb{R} α ∈ R , written B ∈ R V α B \in\mathrm{RV}_{\alpha} B ∈ RV α , if, for every λ > 0 \lambda> 0 λ > 0 , lim t → ∞ B ( λ t ) B ( t ) = λ α \lim_{t \to\infty} \frac{B(\lambda t)}{B(t)} = \lambda^\alpha lim t → ∞ B ( t ) B ( λ t ) = λ α . Equivalently, B B B can be represented in the form B ( t ) = t α L ( t ) B(t) = t^\alpha L(t) B ( t ) = t α L ( t ) , where L L L is a slowly varying function at infinity, that is, lim t → ∞ L ( λ t ) L ( t ) = 1 \lim_{t \to\infty} \frac{L(\lambda t)}{L(t)} = 1 lim t → ∞ L ( t ) L ( λ t ) = 1 , λ > 0 \lambda> 0 λ > 0 .
Let B : ( 0 , ∞ ) → ( 0 , ∞ ) B : (0,\infty) \to(0,\infty) B : ( 0 , ∞ ) → ( 0 , ∞ ) be positive, strictly increasing, and regularly varying at infinity with index α \alpha α , where 1 2 ≤ α < 1 \frac{1}{2} \le\alpha< 1 2 1 ≤ α < 1 . Set b n : = B ( n ) b_n := B(n) b n := B ( n ) , n ≥ 1 n \ge1 n ≥ 1 . Since B B B is regularly varying it admits a generalized inverse B − 1 B^{-1} B − 1 on its range, that is, B − 1 ( s ) : = inf { u > 0 : B ( u ) > s } B^{-1}(s) := \inf\{u > 0 : B(u) > s\} B − 1 ( s ) := inf { u > 0 : B ( u ) > s } , s > 0 s > 0 s > 0 . We also use the following standard properties of regularly varying functions. For the proofs and more information, we refer to [BGT89].
(RV1) if φ ∈ R V α \varphi\in\mathrm{RV}_{\alpha} φ ∈ RV α then log φ ( t ) ∼ α log t \log\varphi(t) \sim\alpha\log t log φ ( t ) ∼ α log t and if α > 0 \alpha> 0 α > 0 then φ ( t ) → ∞ \varphi(t) \to\infty φ ( t ) → ∞ .
(RV2) if B B B is positive, strictly increasing, and B ∈ R V α B \in\mathrm{RV}_{\alpha} B ∈ RV α with α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , its generalized inverse B − 1 B^{-1} B − 1 satisfies B − 1 ∈ R V 1 / α B^{-1} \in\mathrm{RV}_{1/\alpha} B − 1 ∈ RV 1/ α and B − 1 ( B ( t ) ) ∼ t B^{-1}(B(t)) \sim t B − 1 ( B ( t )) ∼ t .
(RV3) if θ \theta θ is slowly varying and λ t → λ ∈ ( 0 , ∞ ) \lambda_t \to\lambda\in(0,\infty) λ t → λ ∈ ( 0 , ∞ ) as t → ∞ t \to\infty t → ∞ , then θ ( λ t t ) / θ ( t ) → 1 \theta(\lambda_t t)/\theta(t) \to1 θ ( λ t t ) / θ ( t ) → 1 .
(RV4) If f ∈ R V α f \in\mathrm{RV}_{\alpha} f ∈ RV α and g g g is eventually positive, measurable, g ( t ) → ∞ g(t) \to\infty g ( t ) → ∞ , and g ∈ R V σ g \in\mathrm{RV}_{\sigma} g ∈ RV σ with σ > 0 \sigma> 0 σ > 0 , then f ∘ g ∈ R V α σ f \circ g \in\mathrm{RV}_{\alpha\sigma} f ∘ g ∈ RV α σ .
(RV5) [Karamata’s tail theorem] If h h h is positive, measurable, locally integrable on some [ t 0 , ∞ ) [t_0,\infty) [ t 0 , ∞ ) , and h ∈ R V α h \in\mathrm{RV}_{\alpha} h ∈ RV α with α < − 1 \alpha< -1 α < − 1 , then for every ε > 0 \varepsilon> 0 ε > 0 there is c ε c_{\varepsilon} c ε such that
∫ c ∞ h ( t ) d t ≤ ( 1 + ε ) c h ( c ) − α − 1 \int_c^{\infty} h(t)\,dt \le(1+\varepsilon)\frac{ch(c)}{-\alpha-1} ∫ c ∞ h ( t ) d t ≤ ( 1 + ε ) − α − 1 c h ( c )
for all c ≥ c ε c \ge c_{\varepsilon} c ≥ c ε .
(RV6) If f ∈ R V ρ f \in\mathrm{RV}_{\rho} f ∈ RV ρ be a positive regularly varying function with index ρ ∈ R \rho\in\mathbb{R} ρ ∈ R . Then, for every η > 0 \eta> 0 η > 0 and A > 1 A > 1 A > 1 , there exists x 0 = x 0 ( η , A ) > 0 x_0=x_0(\eta,A)>0 x 0 = x 0 ( η , A ) > 0 such that
f ( y ) f ( x ) ≤ A max { ( y x ) ρ + η , ( y x ) ρ − η } , x , y ≥ x 0 . \frac{f(y)}{f(x)} \le A\max\left\{\left(\frac{y}{x}\right)^{\rho+\eta},\left(\frac{y}{x}\right)^{\rho-\eta}\right\},\quad x,y\ge x_0. f ( x ) f ( y ) ≤ A max { ( x y ) ρ + η , ( x y ) ρ − η } , x , y ≥ x 0 .
Fix a j ≥ 1 j \ge1 j ≥ 1 and for t > 0 t>0 t > 0 let us define:
G ( t ) : = τ ( 1 [ t , ∞ ) ( ∣ x j ∣ ) ) = μ ( [ t , ∞ ) ) , (35) G(t):=\tau\left(1_{[t,\infty)}(|x_j|)\right)=\mu([t,\infty)), \tag*{(35)} G ( t ) := τ ( 1 [ t , ∞ ) ( ∣ x j ∣ ) ) = μ ([ t , ∞ )) , (35)
H ( t ) : = τ ( x j 2 1 [ 0 , t ) ( ∣ x j ∣ ) ) = ∫ [ 0 , t ) s 2 d μ ( s ) , (36) H(t):=\tau\left(x_j^2 1_{[0,t)}(|x_j|)\right)=\int_{[0,t)}s^2d\mu(s), \tag*{(36)} H ( t ) := τ ( x j 2 1 [ 0 , t ) ( ∣ x j ∣ ) ) = ∫ [ 0 , t ) s 2 d μ ( s ) , (36)
where μ \mu μ is the spectral distribution measure of ∣ x 1 ∣ |x_1| ∣ x 1 ∣ with respect to τ \tau τ , defined by μ ( A ) : = τ ( 1 A ( ∣ x 1 ∣ ) ) \mu(A):=\tau(1_A(|x_1|)) μ ( A ) := τ ( 1 A ( ∣ x 1 ∣ )) , for every Borel subset A A A of [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . By identical distribution, G G G and H H H do not depend on the choice of j j j . For j ≥ 1 j\ge1 j ≥ 1 and c > 0 c>0 c > 0 we write the spectral truncation x j ( c ) : = x j 1 [ 0 , c ) ( ∣ x j ∣ ) x_j^{(c)}:=x_j1_{[0,c)}(|x_j|) x j ( c ) := x j 1 [ 0 , c ) ( ∣ x j ∣ ) , a bounded self-adjoint element of W ∗ ( x j ) W^*(x_j) W ∗ ( x j ) with ∥ x j ( c ) ∥ ∞ ≤ c \left\|x_j^{(c)}\right\|_{\infty}\le c x j ( c ) ∞ ≤ c .
Example 4.2. Take B ( t ) = t log t B(t)=\sqrt{t}\log t B ( t ) = t log t , for t ≥ e t\ge e t ≥ e . Then, observe that B ∈ R V 1 / 2 B\in\mathrm{RV}_{1/2} B ∈ RV 1/2 . Let ( x n ) (x_n) ( x n ) be mean-zero, identically distributed, self-adjoint such that
G ( t ) = { 1 , 0 ≤ t ≤ 1 , t − 2 , t > 1. G(t)=
\begin{cases}
1, & 0\le t\le1,\\
t^{-2}, & t>1.
\end{cases} G ( t ) = { 1 , t − 2 , 0 ≤ t ≤ 1 , t > 1. Then, τ ( ∣ x 1 ∣ ) < ∞ \tau(|x_1|)<\infty τ ( ∣ x 1 ∣ ) < ∞ and x 1 ∉ L 2 ( M , τ ) x_1\notin L_2(\mathcal{M},\tau) x 1 ∈ / L 2 ( M , τ ) . Also note that, H ( t ) = 2 log t H(t)=2\log t H ( t ) = 2 log t , t ≥ 1 t\ge1 t ≥ 1 . Hence, τ ( B − 1 ( ∣ x 1 ∣ ) ) < ∞ \tau(B^{-1}(|x_1|))<\infty τ ( B − 1 ( ∣ x 1 ∣ )) < ∞ , and B − 1 ( t log t ) log t G ( t ) ⟶ 0 \frac{B^{-1}(t\log t)}{\log t}G(t)\longrightarrow0 l o g t B − 1 ( t l o g t ) G ( t ) ⟶ 0 . Further, n log n b n 2 H ( b n log n ) = 1 \frac{n\log n}{b_n^2}H\left(\frac{b_n}{\log n}\right)=1 b n 2 n l o g n H ( l o g n b n ) = 1 . Thus, ℓ = 1 \ell=1 ℓ = 1 . Therefore, all the assumptions of Theorem B(b) holds. In particular, the theorem applies even though x 1 ∉ L 2 ( M , τ ) x_1\notin L_2(\mathcal{M},\tau) x 1 ∈ / L 2 ( M , τ ) . It gives lim sup n → ∞ S n ( A ) n log n ≤ 2 K \limsup_{n\to\infty}\frac{S_n(A)}{\sqrt{n\log n}}\le\sqrt{2}\,K lim sup n → ∞ n l o g n S n ( A ) ≤ 2 K b.a.u.
Lemma 4.3. Let G G G and H H H be as in eqs. (35) and (36) . Then the following identities hold:
(i) For every c ≥ 0 c\ge0 c ≥ 0 ,
H ( c ) = ∫ 0 c 2 t G ( t ) d t − c 2 G ( c ) . H(c)=\int_0^c 2tG(t)\,dt-c^2G(c). H ( c ) = ∫ 0 c 2 tG ( t ) d t − c 2 G ( c ) . (ii) If ∫ [ c , ∞ ) s d μ ( s ) < ∞ \int_{[c,\infty)}s\,d\mu(s)<\infty ∫ [ c , ∞ ) s d μ ( s ) < ∞ , then
∫ [ c , ∞ ) s d μ ( s ) = c G ( c ) + ∫ c ∞ G ( t ) d t . \int_{[c,\infty)}s\,d\mu(s)=cG(c)+\int_c^{\infty}G(t)\,dt. ∫ [ c , ∞ ) s d μ ( s ) = c G ( c ) + ∫ c ∞ G ( t ) d t .
Proof. (i) Using Tonelli’s theorem, we have
∫ 0 c 2 t G ( t ) d t = ∫ 0 c 2 t ∫ [ 0 , ∞ ) 1 [ 0 , s ] ( t ) d μ ( s ) d t = ∫ [ 0 , ∞ ) ( ∫ 0 c 2 t ⋅ 1 [ 0 , s ] ( t ) d t ) d μ ( s ) . \int_0^c 2tG(t)\,dt=\int_0^c 2t\int_{[0,\infty)}\mathbf{1}_{[0,s]}(t)\,d\mu(s)\,dt=\int_{[0,\infty)}\left(\int_0^c 2t\cdot\mathbf{1}_{[0,s]}(t)\,dt\right)d\mu(s). ∫ 0 c 2 tG ( t ) d t = ∫ 0 c 2 t ∫ [ 0 , ∞ ) 1 [ 0 , s ] ( t ) d μ ( s ) d t = ∫ [ 0 , ∞ ) ( ∫ 0 c 2 t ⋅ 1 [ 0 , s ] ( t ) d t ) d μ ( s ) .
Now notice that, ∫ 0 c 2 t ⋅ 1 [ 0 , s ] ( t ) d t = min ( s , c ) 2 \int_0^c 2t\cdot\mathbf{1}_{[0,s]}(t)\,dt=\min(s,c)^2 ∫ 0 c 2 t ⋅ 1 [ 0 , s ] ( t ) d t = min ( s , c ) 2 . Hence
∫ 0 c 2 t G ( t ) d t = ∫ [ 0 , c ) s 2 d μ ( s ) + ∫ [ c , ∞ ) c 2 d μ ( s ) = H ( c ) + c 2 G ( c ) . \int_0^c 2tG(t)\,dt=\int_{[0,c)}s^2d\mu(s)+\int_{[c,\infty)}c^2d\mu(s)=H(c)+c^2G(c). ∫ 0 c 2 tG ( t ) d t = ∫ [ 0 , c ) s 2 d μ ( s ) + ∫ [ c , ∞ ) c 2 d μ ( s ) = H ( c ) + c 2 G ( c ) .
(ii) Using Tonelli’s theorem, we have
∫ c ∞ G ( t ) d t = ∫ c ∞ ∫ [ 0 , ∞ ) 1 [ 0 , s ] ( t ) d μ ( s ) d t = ∫ [ 0 , ∞ ) ( ∫ c ∞ 1 [ 0 , s ] ( t ) d t ) d μ ( s ) . \int_c^\infty G(t)\,dt=\int_c^\infty\int_{[0,\infty)}\mathbf{1}_{[0,s]}(t)\,d\mu(s)\,dt=\int_{[0,\infty)}\left(\int_c^\infty\mathbf{1}_{[0,s]}(t)\,dt\right)d\mu(s). ∫ c ∞ G ( t ) d t = ∫ c ∞ ∫ [ 0 , ∞ ) 1 [ 0 , s ] ( t ) d μ ( s ) d t = ∫ [ 0 , ∞ ) ( ∫ c ∞ 1 [ 0 , s ] ( t ) d t ) d μ ( s ) .
Since ∫ c ∞ 1 [ 0 , s ] ( t ) d t = max { s − c , 0 } \int_c^\infty\mathbf{1}_{[0,s]}(t)\,dt=\max\{s-c,0\} ∫ c ∞ 1 [ 0 , s ] ( t ) d t = max { s − c , 0 } , we have
∫ c ∞ G ( t ) d t = ∫ [ c , ∞ ) ( s − c ) d μ ( s ) = ∫ [ c , ∞ ) s d μ ( s ) − c G ( c ) . \int_c^\infty G(t)\,dt=\int_{[c,\infty)}(s-c)d\mu(s)=\int_{[c,\infty)}sd\mu(s)-cG(c). ∫ c ∞ G ( t ) d t = ∫ [ c , ∞ ) ( s − c ) d μ ( s ) = ∫ [ c , ∞ ) s d μ ( s ) − c G ( c ) .
□
Lemma 4.4. Let φ \varphi φ be regularly varying of index α ∈ R \alpha\in\mathbb{R} α ∈ R . If ( λ t ) t > 0 (\lambda_t)_{t>0} ( λ t ) t > 0 is a family of positive reals with λ t → λ ∈ ( 0 , ∞ ) \lambda_t\to\lambda\in(0,\infty) λ t → λ ∈ ( 0 , ∞ ) as t → ∞ t\to\infty t → ∞ , then φ ( λ t t ) φ ( t ) ⟶ λ α \frac{\varphi(\lambda_t t)}{\varphi(t)}\longrightarrow\lambda^\alpha φ ( t ) φ ( λ t t ) ⟶ λ α .
Proof. By definition of regular variation, write φ ( t ) = t α θ ( t ) \varphi(t)=t^\alpha\theta(t) φ ( t ) = t α θ ( t ) with θ \theta θ slowly varying. Then, we get
φ ( λ t t ) φ ( t ) = λ t α ⋅ θ ( λ t t ) θ ( t ) . \frac{\varphi(\lambda_t t)}{\varphi(t)}=\lambda_t^\alpha\cdot\frac{\theta(\lambda_t t)}{\theta(t)}. φ ( t ) φ ( λ t t ) = λ t α ⋅ θ ( t ) θ ( λ t t ) . Since x ↦ x α x\mapsto x^\alpha x ↦ x α is continuous at λ > 0 \lambda>0 λ > 0 and λ t → λ \lambda_t\to\lambda λ t → λ , we have λ t α → λ α \lambda_t^\alpha\to\lambda^\alpha λ t α → λ α . Since λ t → λ ∈ ( 0 , ∞ ) \lambda_t\to\lambda\in(0,\infty) λ t → λ ∈ ( 0 , ∞ ) , (RV3) gives θ ( λ t t ) / θ ( t ) → 1 \theta(\lambda_t t)/\theta(t)\to1 θ ( λ t t ) / θ ( t ) → 1 , as t → ∞ t\to\infty t → ∞ .
Lemma 4.5. Let b n : = B ( n ) b_n:=B(n) b n := B ( n ) , n ≥ 1 n\ge1 n ≥ 1 , where B : ( 0 , ∞ ) → ( 0 , ∞ ) B:(0,\infty)\to(0,\infty) B : ( 0 , ∞ ) → ( 0 , ∞ ) be positive, strictly increasing, and regularly varying at infinity with index α ∈ [ 1 2 , 1 ) \alpha\in[\frac{1}{2},1) α ∈ [ 2 1 , 1 ) satisfying condition (1) of Theorem B. Fix δ > 0 \delta>0 δ > 0 and, for n ≥ 2 n\ge2 n ≥ 2 , set c n : = δ b n / log n c_n:=\delta b_n/\log n c n := δ b n / log n , n ≥ 2 n\ge2 n ≥ 2 . Then c n → ∞ c_n\to\infty c n → ∞ as n → ∞ n\to\infty n → ∞ , and
(a) log c n ∼ α log n \log c_n\sim\alpha\log n log c n ∼ α log n ;
(b) c n log c n ∼ α δ b n c_n\log c_n\sim\alpha\delta b_n c n log c n ∼ α δ b n ;
(c) B − 1 ( c n log c n ) ∼ ( α δ ) 1 / α n B^{-1}(c_n\log c_n)\sim(\alpha\delta)^{1/\alpha}n B − 1 ( c n log c n ) ∼ ( α δ ) 1/ α n .
Proof. The map t ↦ B ( t ) / log t t\mapsto B(t)/\log t t ↦ B ( t ) / log t is regularly varying of index α > 0 \alpha>0 α > 0 , hence by (RV1), we have log ( B ( t ) / log t ) ∼ α log t → ∞ \log(B(t)/\log t)\sim\alpha\log t\to\infty log ( B ( t ) / log t ) ∼ α log t → ∞ , so B ( t ) / log t → ∞ B(t)/\log t\to\infty B ( t ) / log t → ∞ . Observe that, b n / log n → ∞ b_n/\log n\to\infty b n / log n → ∞ , and multiplying by the fixed constant δ \delta δ , we get c n = δ b n / log n → ∞ c_n=\delta b_n/\log n\to\infty c n = δ b n / log n → ∞ as n → ∞ n\to\infty n → ∞ .
(a) By (RV1), log B ( n ) ∼ α log n \log B(n)\sim\alpha\log n log B ( n ) ∼ α log n . Since log c n = log δ + log b n − log log n \log c_n=\log\delta+\log b_n-\log\log n log c n = log δ + log b n − log log n , and log δ \log\delta log δ is a fixed constant while log log n = o ( log n ) \log\log n=o(\log n) log log n = o ( log n ) , substituting log b n = log B ( n ) ∼ α log n \log b_n=\log B(n)\sim\alpha\log n log b n = log B ( n ) ∼ α log n gives $\log c_n\sim\alpha\log n.
(b) Multiply c n = δ b n / log n c_n=\delta b_n/\log n c n = δ b n / log n by log c n \log c_n log c n and use (a), we obtain
c n log c n = δ b n log n ⋅ log c n ∼ δ b n log n ⋅ α log n = α δ b n . c_n\log c_n=\delta\frac{b_n}{\log n}\cdot\log c_n\sim\delta\frac{b_n}{\log n}\cdot\alpha\log n=\alpha\delta b_n. c n log c n = δ log n b n ⋅ log c n ∼ δ log n b n ⋅ α log n = α δ b n .
(c) Set λ n : = c n log c n / b n \lambda_n:=c_n\log c_n/b_n λ n := c n log c n / b n . Now by (b), λ n → λ : = α δ ∈ ( 0 , ∞ ) \lambda_n\to\lambda:=\alpha\delta\in(0,\infty) λ n → λ := α δ ∈ ( 0 , ∞ ) . Since B − 1 ∈ RV 1 / α B^{-1}\in\operatorname{RV}_{1/\alpha} B − 1 ∈ RV 1/ α , we may write B − 1 ( s ) = s 1 / α L 2 ( s ) B^{-1}(s)=s^{1/\alpha}L_2(s) B − 1 ( s ) = s 1/ α L 2 ( s ) , where L 2 L_2 L 2 is slowly varying. Now notice that, λ n → λ ∈ ( 0 , ∞ ) \lambda_n\to\lambda\in(0,\infty) λ n → λ ∈ ( 0 , ∞ ) and (RV3) gives
B − 1 ( λ n b n ) B − 1 ( b n ) = λ n 1 / α L 2 ( λ n b n ) L 2 ( b n ) ⟶ λ 1 / α = ( α δ ) 1 / α . \frac{B^{-1}(\lambda_n b_n)}{B^{-1}(b_n)}=\lambda_n^{1/\alpha}\frac{L_2(\lambda_n b_n)}{L_2(b_n)}\longrightarrow\lambda^{1/\alpha}=(\alpha\delta)^{1/\alpha}. B − 1 ( b n ) B − 1 ( λ n b n ) = λ n 1/ α L 2 ( b n ) L 2 ( λ n b n ) ⟶ λ 1/ α = ( α δ ) 1/ α .
Now using (RV2), we have B − 1 ( b n ) = B − 1 ( B ( n ) ) ∼ n B^{-1}(b_n)=B^{-1}(B(n))\sim n B − 1 ( b n ) = B − 1 ( B ( n )) ∼ n . Since λ n b n = c n log c n \lambda_n b_n=c_n\log c_n λ n b n = c n log c n by definition of λ n \lambda_n λ n , combining above we get B − 1 ( c n log c n ) ∼ ( α δ ) 1 / α n B^{-1}(c_n\log c_n)\sim(\alpha\delta)^{1/\alpha}n B − 1 ( c n log c n ) ∼ ( α δ ) 1/ α n .
Lemma 4.6. Let B B B be as in Lemma 4.5. Define h ( t ) : = log t / B − 1 ( t log t ) h(t):=\log t/B^{-1}(t\log t) h ( t ) := log t / B − 1 ( t log t ) for t > 1 t>1 t > 1 . Then
(i) h h h is positive and measurable on some interval [ t 0 , ∞ ) [t_0,\infty) [ t 0 , ∞ ) ;
(ii) h ∈ RV − 1 / α h\in\operatorname{RV}_{-1/\alpha} h ∈ RV − 1/ α , and − 1 / α < − 1 -1/\alpha<-1 − 1/ α < − 1 since α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) ;
(iii) h h h is locally integrable on [ t 0 , ∞ ) [t_0,\infty) [ t 0 , ∞ ) ;
(iv) with c n , δ c_n,\delta c n , δ as in Lemma 4.5, c n h ( c n ) ∼ K ( α , δ ) b n / n c_nh(c_n)\sim K(\alpha,\delta)b_n/n c n h ( c n ) ∼ K ( α , δ ) b n / n as n → ∞ n\to\infty n → ∞ , where K ( α , δ ) : = α δ ( α δ ) − 1 / α K(\alpha,\delta):=\alpha\delta(\alpha\delta)^{-1/\alpha} K ( α , δ ) := α δ ( α δ ) − 1/ α .
Proof. (ii) Notice that, the map g ( t ) : = t log t g(t):=t\log t g ( t ) := t log t is measurable. Observe that g ( t ) → ∞ g(t)\to\infty g ( t ) → ∞ , and belongs to RV 1 \operatorname{RV}_1 RV 1 . By (RV4) applied with f = B − 1 ∈ RV 1 / α f=B^{-1}\in\operatorname{RV}_{1/\alpha} f = B − 1 ∈ RV 1/ α and g ∈ RV 1 g\in\operatorname{RV}_1 g ∈ RV 1 , notice that t ↦ B − 1 ( t log t ) t\mapsto B^{-1}(t\log t) t ↦ B − 1 ( t log t ) is regularly varying with index 1 / α 1/\alpha 1/ α . Since log t ∈ RV 0 \log t\in\operatorname{RV}_0 log t ∈ RV 0 , the quotient h h h is regularly varying with index ρ h : = 0 − 1 / α = − 1 / α \rho_h:=0-1/\alpha=-1/\alpha ρ h := 0 − 1/ α = − 1/ α , and ρ h < − 1 \rho_h<-1 ρ h < − 1 because α < 1 \alpha<1 α < 1 .
(i), (iii) Note that B − 1 B^{-1} B − 1 is Borel measurable. Then, there exists t 0 > 0 t_0>0 t 0 > 0 such that B − 1 ( t log t ) > 0 B^{-1}(t\log t)>0 B − 1 ( t log t ) > 0 for all t ≥ t 0 t\ge t_0 t ≥ t 0 . It follows that h h h is positive and measurable on [ t 0 , ∞ ) [t_0,\infty) [ t 0 , ∞ ) , thus proving (i). Now suppose that [ t 1 , t 2 ] ⊂ [ t 0 , ∞ ) [t_1,t_2]\subset[t_0,\infty) [ t 1 , t 2 ] ⊂ [ t 0 , ∞ ) be compact. Since both B − 1 B^{-1} B − 1 and the function t ↦ t log t t\mapsto t\log t t ↦ t log t are non-decreasing, we have
0 < B − 1 ( t 1 log t 1 ) ≤ B − 1 ( t log t ) ≤ B − 1 ( t 2 log t 2 ) < ∞ , t ∈ [ t 1 , t 2 ] . 0<B^{-1}(t_1\log t_1)\le B^{-1}(t\log t)\le B^{-1}(t_2\log t_2)<\infty,\qquad t\in[t_1,t_2]. 0 < B − 1 ( t 1 log t 1 ) ≤ B − 1 ( t log t ) ≤ B − 1 ( t 2 log t 2 ) < ∞ , t ∈ [ t 1 , t 2 ] .
Therefore, 0 ≤ h ( t ) ≤ log t 2 B − 1 ( t 1 log t 1 ) 0\le h(t)\le\frac{\log t_2}{B^{-1}(t_1\log t_1)} 0 ≤ h ( t ) ≤ B − 1 ( t 1 l o g t 1 ) l o g t 2 , t ∈ [ t 1 , t 2 ] t\in[t_1,t_2] t ∈ [ t 1 , t 2 ] . Thus h h h is bounded on every compact subinterval of [ t 0 , ∞ ) [t_0,\infty) [ t 0 , ∞ ) , and hence h h h is locally integrable on [ t 0 , ∞ ) [t_0,\infty) [ t 0 , ∞ ) , which proves (iii).
(iv) By definition of h h h and Lemma 4.5 (b),(c), we get
c n h ( c n ) = c n ⋅ log c n B − 1 ( c n log c n ) ∼ α δ b n ( α δ ) 1 / α n = b n n ⋅ α δ ( α δ ) − 1 / α = K ( α , δ ) b n n . c_nh(c_n)=c_n\cdot\frac{\log c_n}{B^{-1}(c_n\log c_n)}\sim\frac{\alpha\delta b_n}{(\alpha\delta)^{1/\alpha}n}=\frac{b_n}{n}\cdot\alpha\delta(\alpha\delta)^{-1/\alpha}=K(\alpha,\delta)\frac{b_n}{n}. c n h ( c n ) = c n ⋅ B − 1 ( c n log c n ) log c n ∼ ( α δ ) 1/ α n α δ b n = n b n ⋅ α δ ( α δ ) − 1/ α = K ( α , δ ) n b n .
Proposition 4.7. Let b n b_n b n be same as in Lemma 4.5 and μ \mu μ be as in (35) . Under Conditions (1) and (2) of Theorem B, for every δ > 0 \delta> 0 δ > 0 ,
n b n ∫ [ δ b n / log n , ∞ ) s d μ ( s ) ⟶ 0 ( n → ∞ ) . \frac{n}{b_n}\int_{[\delta b_n/\log n,\ \infty)}s\,d\mu(s)\longrightarrow0 \qquad(n\to\infty). b n n ∫ [ δ b n / l o g n , ∞ ) s d μ ( s ) ⟶ 0 ( n → ∞ ) . Proof. Fix δ > 0 \delta> 0 δ > 0 and set c n : = δ b n log n c_n := \frac{\delta b_n}{\log n} c n := l o g n δ b n , n ≥ 2 n \ge2 n ≥ 2 . Then by Lemma 4.5, there exists N ≥ 1 N \ge1 N ≥ 1 such that c n c_n c n satisfies the hypothesis of Condition (2) of Theorem B for all n ≥ N n \ge N n ≥ N . By Lemma 4.3 , the identity applies at c = c n c=c_n c = c n gives
n b n ∫ [ c n , ∞ ) s d μ ( s ) = n b n c n G ( c n ) + n b n ∫ c n ∞ G ( t ) d t . \frac{n}{b_n}\int_{[c_n,\infty)}s\,d\mu(s)=\frac{n}{b_n}c_nG(c_n)+\frac{n}{b_n}\int_{c_n}^{\infty}G(t)\,dt. b n n ∫ [ c n , ∞ ) s d μ ( s ) = b n n c n G ( c n ) + b n n ∫ c n ∞ G ( t ) d t . where the first term is = : T 1 ( n ) =:T_1(n) =: T 1 ( n ) and the second is = : T 2 ( n ) =:T_2(n) =: T 2 ( n ) .
We prove that T 1 ( n ) → 0 T_1(n)\to0 T 1 ( n ) → 0 and T 2 ( n ) → 0 T_2(n)\to0 T 2 ( n ) → 0 as n → ∞ n\to\infty n → ∞ . By definition of T 1 T_1 T 1 ,
T 1 ( n ) = δ n log n G ( c n ) = δ ⋅ ( α δ ) 1 / α 1 ⋅ B − 1 ( c n log c n ) log c n G ( c n ) ⋅ ( 1 + o ( 1 ) ) , T_1(n)=\delta\frac{n}{\log n}G(c_n)=\delta\cdot\frac{(\alpha\delta)^{1/\alpha}}{1}\cdot\frac{B^{-1}(c_n\log c_n)}{\log c_n}G(c_n)\cdot(1+o(1)), T 1 ( n ) = δ log n n G ( c n ) = δ ⋅ 1 ( α δ ) 1/ α ⋅ log c n B − 1 ( c n log c n ) G ( c n ) ⋅ ( 1 + o ( 1 )) , using n ∼ ( α δ ) − 1 / α B − 1 ( c n log c n ) n\sim(\alpha\delta)^{-1/\alpha}B^{-1}(c_n\log c_n) n ∼ ( α δ ) − 1/ α B − 1 ( c n log c n ) and log n ∼ log c n / α \log n\sim\log c_n/\alpha log n ∼ log c n / α , both from Lemma 4.5(a),(c). Since c n → ∞ c_n\to\infty c n → ∞ , Condition (2) of Theorem B gives B − 1 ( c n log c n ) G ( c n ) / log c n → 0 B^{-1}(c_n\log c_n)G(c_n)/\log c_n\to0 B − 1 ( c n log c n ) G ( c n ) / log c n → 0 . So, T 1 ( n ) → 0 T_1(n)\to0 T 1 ( n ) → 0 as n → ∞ n\to\infty n → ∞ .
Let h h h be as in Lemma 4.6 . Now observe that, Condition (2) of Theorem B gives G ( t ) / h ( t ) → 0 G(t)/h(t)\to0 G ( t ) / h ( t ) → 0 as t → ∞ t\to\infty t → ∞ .
Set η ( c ) : = sup t ≥ c G ( t ) h ( t ) \eta(c):=\sup_{t\ge c}\frac{G(t)}{h(t)} η ( c ) := sup t ≥ c h ( t ) G ( t ) . Note that η ( c ) → 0 \eta(c)\to0 η ( c ) → 0 as c → ∞ c\to\infty c → ∞ , and ∫ c ∞ G ( t ) d t ≤ η ( c ) ∫ c ∞ h ( t ) d t \int_c^\infty G(t)\,dt\le\eta(c)\int_c^\infty h(t)\,dt ∫ c ∞ G ( t ) d t ≤ η ( c ) ∫ c ∞ h ( t ) d t for all c > 1 c>1 c > 1 . By Lemma 4.6 (ii),(iii), h h h satisfies the hypotheses of (RV5) with ρ = ρ h = − 1 / α \rho=\rho_h=-1/\alpha ρ = ρ h = − 1/ α , so for every ε > 0 \varepsilon>0 ε > 0 there is c ε c_\varepsilon c ε such that
∫ c ∞ h ( t ) d t ≤ ( 1 + ε ) α 1 − α c h ( c ) , c ≥ c ε . (37) \int_c^\infty h(t)\,dt\le(1+\varepsilon)\frac{\alpha}{1-\alpha}ch(c),\qquad c\ge c_\varepsilon. \tag*{(37)} ∫ c ∞ h ( t ) d t ≤ ( 1 + ε ) 1 − α α c h ( c ) , c ≥ c ε . (37) using − ρ h − 1 = 1 / α − 1 = ( 1 − α ) / α -\rho_h-1=1/\alpha-1=(1-\alpha)/\alpha − ρ h − 1 = 1/ α − 1 = ( 1 − α ) / α . Since c n → ∞ c_n\to\infty c n → ∞ , there is N ε N_\varepsilon N ε with c n ≥ c ε c_n\ge c_\varepsilon c n ≥ c ε for n ≥ N ε n\ge N_\varepsilon n ≥ N ε . Now combining with (37) at c = c n c=c_n c = c n , we have
∫ c n ∞ G ( t ) d t ≤ ( 1 + ε ) α 1 − α η ( c n ) c n h ( c n ) , n ≥ N ε . \int_{c_n}^\infty G(t)\,dt\le(1+\varepsilon)\frac{\alpha}{1-\alpha}\eta(c_n)c_nh(c_n),\qquad n\ge N_\varepsilon. ∫ c n ∞ G ( t ) d t ≤ ( 1 + ε ) 1 − α α η ( c n ) c n h ( c n ) , n ≥ N ε . By Lemma 4.6 (iv), we get c n h ( c n ) ∼ K ( α , δ ) b n / n c_nh(c_n)\sim K(\alpha,\delta)b_n/n c n h ( c n ) ∼ K ( α , δ ) b n / n . Therefore,
T 2 ( n ) = n b n ∫ c n ∞ G ( t ) d t ≤ ( 1 + ε ) α 1 − α K ( α , δ ) η ( c n ) ( 1 + o ( 1 ) ) , n ≥ N ε . T_2(n)=\frac{n}{b_n}\int_{c_n}^\infty G(t)\,dt\le(1+\varepsilon)\frac{\alpha}{1-\alpha}K(\alpha,\delta)\eta(c_n)(1+o(1)),\qquad n\ge N_\varepsilon. T 2 ( n ) = b n n ∫ c n ∞ G ( t ) d t ≤ ( 1 + ε ) 1 − α α K ( α , δ ) η ( c n ) ( 1 + o ( 1 )) , n ≥ N ε . Using η ( c n ) → 0 \eta(c_n)\to0 η ( c n ) → 0 , we conclude that T 2 ( n ) → 0 T_2(n)\to0 T 2 ( n ) → 0 as n → ∞ n\to\infty n → ∞ .
Lemma 4.8. Assume that τ ( B − 1 ( ∣ x 1 ∣ ) ) < ∞ \tau\left(B^{-1}(|x_1|)\right)<\infty τ ( B − 1 ( ∣ x 1 ∣ ) ) < ∞ and recall G G G from (35) . Then, for every Δ > 0 \Delta>0 Δ > 0 , ∑ k = 1 ∞ G ( Δ b k ) < ∞ \sum_{k=1}^{\infty}G(\Delta b_k)<\infty ∑ k = 1 ∞ G ( Δ b k ) < ∞ .
Proof. First take Δ = 1 \Delta=1 Δ = 1 . Let s > 0 s>0 s > 0 and define N ( s ) : = # { k ≥ 1 : B ( k ) ≤ s } N(s):=\#\{k\ge1:B(k)\le s\} N ( s ) := # { k ≥ 1 : B ( k ) ≤ s } . Now notice that, N ( s ) = ∑ k ≥ 1 1 [ B ( k ) , ∞ ) ( s ) N(s)=\sum_{k\ge1}\mathbf{1}_{[B(k),\infty)}(s) N ( s ) = ∑ k ≥ 1 1 [ B ( k ) , ∞ ) ( s ) . Since B B B is increasing, we have N ( s ) ≤ B − 1 ( s ) + 1 N(s)\le B^{-1}(s)+1 N ( s ) ≤ B − 1 ( s ) + 1 . Using normality of τ \tau τ and functional calculus on N ( s ) N(s) N ( s ) , we have ∑ k = 1 ∞ G ( b k ) ≤ τ ( B − 1 ( ∣ x 1 ∣ ) ) + τ ( 1 ) < ∞ \sum_{k=1}^{\infty}G(b_k)\le\tau\left(B^{-1}(|x_1|)\right)+\tau(1)<\infty ∑ k = 1 ∞ G ( b k ) ≤ τ ( B − 1 ( ∣ x 1 ∣ ) ) + τ ( 1 ) < ∞ , by hypothesis.
Now take Δ > 0 \Delta> 0 Δ > 0 in general. Define, B Δ : = Δ B B_{\Delta} := \Delta B B Δ := Δ B . Observe that, B Δ B_{\Delta} B Δ is positive, strictly increasing, and regularly varying with index α \alpha α , with B Δ − 1 ( s ) = B − 1 ( s / Δ ) B_{\Delta}^{-1}(s) = B^{-1}(s/\Delta) B Δ − 1 ( s ) = B − 1 ( s /Δ ) . Therefore, ∑ k G ( Δ b k ) ≤ τ ( B − 1 ( ∣ x 1 ∣ / Δ ) ) + τ ( 1 ) \sum_k G(\Delta b_k) \le\tau\left(B^{-1}(|x_1|/\Delta)\right) + \tau(1) ∑ k G ( Δ b k ) ≤ τ ( B − 1 ( ∣ x 1 ∣/Δ ) ) + τ ( 1 ) .
It remains to prove that τ ( B − 1 ( ∣ x 1 ∣ Δ ) ) < ∞ \tau\left(B^{-1}\left(\frac{|x_1|}{\Delta}\right)\right) < \infty τ ( B − 1 ( Δ ∣ x 1 ∣ ) ) < ∞ . Since B − 1 ∈ RV 1 / α B^{-1} \in\operatorname{RV}_{1/\alpha} B − 1 ∈ RV 1/ α , using Lemma 4.4 we have B − 1 ( s / Δ ) B − 1 ( s ) ⟶ Δ − 1 / α \frac{B^{-1}(s/\Delta)}{B^{-1}(s)} \longrightarrow\Delta^{-1/\alpha} B − 1 ( s ) B − 1 ( s /Δ ) ⟶ Δ − 1/ α , as s → ∞ s \to\infty s → ∞ . Hence, there exists s 0 ∈ R s_0 \in\mathbb{R} s 0 ∈ R and C < ∞ C < \infty C < ∞ such that, B − 1 ( s / Δ ) ≤ C B − 1 ( s ) B^{-1}(s/\Delta) \le C B^{-1}(s) B − 1 ( s /Δ ) ≤ C B − 1 ( s ) , ∀ s ≥ s 0 \forall s \ge s_0 ∀ s ≥ s 0 . Also, notice that, for 0 ≤ s < s 0 0 \le s < s_0 0 ≤ s < s 0 we have B − 1 ( s / Δ ) ≤ B − 1 ( s 0 / Δ ) B^{-1}(s/\Delta) \le B^{-1}(s_0/\Delta) B − 1 ( s /Δ ) ≤ B − 1 ( s 0 /Δ ) . Therefore, we can write
τ ( B − 1 ( ∣ x 1 ∣ Δ ) ) = τ ( B − 1 ( ∣ x 1 ∣ Δ ) 1 [ 0 , s 0 ) ( ∣ x 1 ∣ ) ) + τ ( B − 1 ( ∣ x 1 ∣ Δ ) 1 [ s 0 , ∞ ) ( ∣ x 1 ∣ ) ) ≤ B − 1 ( s 0 / Δ ) τ ( 1 [ 0 , s 0 ) ( ∣ x 1 ∣ ) ) + C τ ( B − 1 ( ∣ x 1 ∣ ) 1 [ s 0 , ∞ ) ( ∣ x 1 ∣ ) ) < ∞ . \begin{aligned}
\tau\left(B^{-1}\left(\frac{|x_1|}{\Delta}\right)\right)
= \tau\left(B^{-1}\left(\frac{|x_1|}{\Delta}\right)1_{[0,s_0)}(|x_1|)\right) + \tau\left(B^{-1}\left(\frac{|x_1|}{\Delta}\right)1_{[s_0,\infty)}(|x_1|)\right) \\
&\le B^{-1}(s_0/\Delta)\tau\left(1_{[0,s_0)}(|x_1|)\right) + C\tau\left(B^{-1}(|x_1|)1_{[s_0,\infty)}(|x_1|)\right) < \infty.
\end{aligned} τ ( B − 1 ( Δ ∣ x 1 ∣ ) ) = τ ( B − 1 ( Δ ∣ x 1 ∣ ) 1 [ 0 , s 0 ) ( ∣ x 1 ∣ ) ) + τ ( B − 1 ( Δ ∣ x 1 ∣ ) 1 [ s 0 , ∞ ) ( ∣ x 1 ∣ ) ) ≤ B − 1 ( s 0 /Δ ) τ ( 1 [ 0 , s 0 ) ( ∣ x 1 ∣ ) ) + C τ ( B − 1 ( ∣ x 1 ∣ ) 1 [ s 0 , ∞ ) ( ∣ x 1 ∣ ) ) < ∞.
This completes the proof. □
For k ≥ 0 k \ge0 k ≥ 0 set M k : = W ∗ ( 1 , x 1 , … , x k ) ⊆ M \mathcal{M}_k := W^*(1,x_1,\ldots,x_k) \subseteq\mathcal{M} M k := W ∗ ( 1 , x 1 , … , x k ) ⊆ M , M 0 : = C ⋅ 1 \mathcal{M}_0 := \mathbb{C}\cdot1 M 0 := C ⋅ 1 . Since τ \tau τ is finite, normal, and faithful, each subalgebra M k \mathcal{M}_k M k admits a (unique, trace-preserving) normal conditional expectation E k : M ⟶ M k E_k:\mathcal{M}\longrightarrow\mathcal{M}_k E k : M ⟶ M k .
Lemma 4.9. Let ( x k ) (x_k) ( x k ) and ( a n , k ) (a_{n,k}) ( a n , k ) are same as in Theorem B. Fix n ≥ 1 n \ge1 n ≥ 1 and c > 0 c > 0 c > 0 , and set, for 1 ≤ k ≤ n 1 \le k \le n 1 ≤ k ≤ n ,
x k ( c ) : = x k 1 [ 0 , c ) ( ∣ x k ∣ ) , d n , k : = a n , k ( x k ( c ) − τ ( x k ( c ) ) ) , y k : = ∑ j = 1 k d n , j , y 0 : = 0. x_k^{(c)} := x_k1_{[0,c)}(|x_k|), \qquad d_{n,k} := a_{n,k}\left(x_k^{(c)}-\tau\left(x_k^{(c)}\right)\right), \qquad y_k := \sum_{j=1}^{k}d_{n,j}, \quad y_0 := 0. x k ( c ) := x k 1 [ 0 , c ) ( ∣ x k ∣ ) , d n , k := a n , k ( x k ( c ) − τ ( x k ( c ) ) ) , y k := j = 1 ∑ k d n , j , y 0 := 0. Then
(a) E k − 1 ( d n , k ) = 0 E_{k-1}(d_{n,k}) = 0 E k − 1 ( d n , k ) = 0 for every 1 ≤ k ≤ n 1 \le k \le n 1 ≤ k ≤ n ;
(b) ( y k ) k = 0 n (y_k)_{k=0}^{n} ( y k ) k = 0 n is a self-adjoint martingale with respect to the filtration ( M k , E k ) k = 0 n (\mathcal{M}_k,E_k)_{k=0}^{n} ( M k , E k ) k = 0 n .
Proof. (a) Write f ( t ) : = a n , k ( t 1 ( − c , c ) ( t ) − τ ( x k ( c ) ) ) f(t) := a_{n,k}\left(t1_{(-c,c)}(t)-\tau\left(x_k^{(c)}\right)\right) f ( t ) := a n , k ( t 1 ( − c , c ) ( t ) − τ ( x k ( c ) ) ) . Note that d n , k = f ( x k ) ∈ W ∗ ( x k ) d_{n,k}=f(x_k)\in W^*(x_k) d n , k = f ( x k ) ∈ W ∗ ( x k ) . Since, M k − 1 = W ∗ ( 1 , x 1 , … , x k − 1 ) \mathcal{M}_{k-1}=W^*(1,x_1,\ldots,x_{k-1}) M k − 1 = W ∗ ( 1 , x 1 , … , x k − 1 ) is independent of W ∗ ( x k ) W^*(x_k) W ∗ ( x k ) , we have τ ( z x k ( c ) ) = τ ( z ) τ ( x k ( c ) ) , ∀ z ∈ M k − 1 \tau(zx_k^{(c)})=\tau(z)\tau(x_k^{(c)}),\ \forall z\in\mathcal{M}_{k-1} τ ( z x k ( c ) ) = τ ( z ) τ ( x k ( c ) ) , ∀ z ∈ M k − 1 . The scalar operator τ ( x k ( c ) ) ⋅ 1 ∈ M k − 1 \tau(x_k^{(c)})\cdot1\in\mathcal{M}_{k-1} τ ( x k ( c ) ) ⋅ 1 ∈ M k − 1 and satisfies τ ( z ⋅ τ ( x k ( c ) ) ⋅ 1 ) = τ ( z x k ( c ) ) , ∀ z ∈ M k − 1 \tau\left(z\cdot\tau(x_k^{(c)})\cdot1\right)=\tau(zx_k^{(c)}),\ \forall z\in\mathcal{M}_{k-1} τ ( z ⋅ τ ( x k ( c ) ) ⋅ 1 ) = τ ( z x k ( c ) ) , ∀ z ∈ M k − 1 . By uniqueness of the conditional expectation and faithfulness of τ \tau τ , we have E k − 1 ( x k ( c ) ) = τ ( x k ( c ) ) ⋅ 1 E_{k-1}(x_k^{(c)})=\tau(x_k^{(c)})\cdot1 E k − 1 ( x k ( c ) ) = τ ( x k ( c ) ) ⋅ 1 . It follows that, E k − 1 ( d n , k ) = 0 E_{k-1}(d_{n,k})=0 E k − 1 ( d n , k ) = 0 .
(b) Notice that, y k y_k y k are self-adjoint for every k k k . For each j ≤ k j\le k j ≤ k , d n , j ∈ W ∗ ( x j ) d_{n,j}\in W^*(x_j) d n , j ∈ W ∗ ( x j ) . Now d n , j ∈ M k d_{n,j}\in\mathcal{M}_k d n , j ∈ M k for every j ≤ k j\le k j ≤ k . Now y k ∈ M k y_k\in\mathcal{M}_k y k ∈ M k . Write y k = y k − 1 + d n , k y_k=y_{k-1}+d_{n,k} y k = y k − 1 + d n , k . By linearity of E k − 1 E_{k-1} E k − 1 , we have E k − 1 ( y k ) = E k − 1 ( y k − 1 ) + E k − 1 ( d n , k ) E_{k-1}(y_k)=E_{k-1}(y_{k-1})+E_{k-1}(d_{n,k}) E k − 1 ( y k ) = E k − 1 ( y k − 1 ) + E k − 1 ( d n , k ) . Since y k − 1 ∈ F k − 1 y_{k-1}\in\mathcal{F}_{k-1} y k − 1 ∈ F k − 1 , we have E k − 1 ( y k − 1 ) = y k − 1 E_{k-1}(y_{k-1})=y_{k-1} E k − 1 ( y k − 1 ) = y k − 1 . By part (a), E k − 1 ( d n , k ) = 0 E_{k-1}(d_{n,k})=0 E k − 1 ( d n , k ) = 0 . So E k − 1 ( y k ) = y k − 1 + 0 = y k − 1 E_{k-1}(y_k)=y_{k-1}+0=y_{k-1} E k − 1 ( y k ) = y k − 1 + 0 = y k − 1 .
Lemma 4.10. Let y k y_k y k and d n , k d_{n,k} d n , k be as in Lemma 4.9 . For a positive sequence ( δ n ) (\delta_n) ( δ n ) , set c = δ n b n / log n c=\delta_n b_n/\log n c = δ n b n / log n , with ( δ n ) (\delta_n) ( δ n ) and A ∈ D A\in\mathcal{D} A ∈ D with K = sup n , k ∣ a n , k ∣ K=\sup_{n,k}|a_{n,k}| K = sup n , k ∣ a n , k ∣ , we obtain
(a) ∥ d n , k ∥ ∞ ≤ M n : = 2 K c \|d_{n,k}\|_{\infty}\le M_n:=2Kc ∥ d n , k ∥ ∞ ≤ M n := 2 K c for every 1 ≤ k ≤ n 1\le k\le n 1 ≤ k ≤ n ;
(b) E k − 1 ( d n , k 2 ) = τ ( d n , k 2 ) ⋅ 1 E_{k-1}(d_{n,k}^2)=\tau(d_{n,k}^2)\cdot1 E k − 1 ( d n , k 2 ) = τ ( d n , k 2 ) ⋅ 1 for every 1 ≤ k ≤ n 1\le k\le n 1 ≤ k ≤ n ;
(c) ∑ k = 1 n E k − 1 ( d n , k 2 ) ≤ D n 2 ⋅ 1 \sum_{k=1}^{n}E_{k-1}(d_{n,k}^2)\le D_n^2\cdot1 ∑ k = 1 n E k − 1 ( d n , k 2 ) ≤ D n 2 ⋅ 1 , where D n 2 : = n K 2 H ( δ n b n log n ) D_n^2:=nK^2H\left(\frac{\delta_n b_n}{\log n}\right) D n 2 := n K 2 H ( l o g n δ n b n ) .
Proof. (a) By construction ∥ x k ( c ) ∥ ∞ ≤ c = δ n b n / log n \lVert x_k^{(c)}\rVert_\infty\le c = \delta_n b_n/\log n ∥ x k ( c ) ∥ ∞ ≤ c = δ n b n / log n . So, ∥ x k ( c ) − τ ( x k ( c ) ) ∥ ∞ ≤ 2 c \lVert x_k^{(c)}-\tau(x_k^{(c)})\rVert_\infty\le2c ∥ x k ( c ) − τ ( x k ( c ) ) ∥ ∞ ≤ 2 c , and ∥ d n , k ∥ ∞ = ∣ a n , k ∣ ⋅ ∥ x k ( c ) − τ ( x k ( c ) ) ∥ ∞ ≤ K ⋅ 2 c = M n \lVert d_{n,k}\rVert_\infty= |a_{n,k}|\cdot\lVert x_k^{(c)}-\tau(x_k^{(c)})\rVert_\infty\le K\cdot2c = M_n ∥ d n , k ∥ ∞ = ∣ a n , k ∣ ⋅ ∥ x k ( c ) − τ ( x k ( c ) ) ∥ ∞ ≤ K ⋅ 2 c = M n .
(b) Let g ( t ) : = f ( t ) 2 g(t):=f(t)^2 g ( t ) := f ( t ) 2 , with f f f as in the proof of Lemma 4.9 , a bounded Borel function, so d n , k 2 = g ( x k ) ∈ W ∗ ( x k ) d_{n,k}^2=g(x_k)\in W^*(x_k) d n , k 2 = g ( x k ) ∈ W ∗ ( x k ) . Thus, uniqueness of the conditional expectation forces E k − 1 ( d n , k 2 ) = τ ( d n , k 2 ) ⋅ 1 E_{k-1}(d_{n,k}^2)=\tau(d_{n,k}^2)\cdot1 E k − 1 ( d n , k 2 ) = τ ( d n , k 2 ) ⋅ 1 .
(c) By definition, τ ( d n , k 2 ) = a n , k 2 ( τ ( ( x k ( c ) ) 2 ) − τ ( x k ( c ) ) 2 ) ≤ a n , k 2 τ ( ( x k ( c ) ) 2 ) = a n , k 2 H ( c ) \tau(d_{n,k}^2)=a_{n,k}^2\left(\tau\left((x_k^{(c)})^2\right)-\tau(x_k^{(c)})^2\right)\le a_{n,k}^2\tau\left((x_k^{(c)})^2\right)=a_{n,k}^2H(c) τ ( d n , k 2 ) = a n , k 2 ( τ ( ( x k ( c ) ) 2 ) − τ ( x k ( c ) ) 2 ) ≤ a n , k 2 τ ( ( x k ( c ) ) 2 ) = a n , k 2 H ( c ) , using (36) and identical distribution of the x k x_k x k . Summing over k = 1 , … , n k=1,\ldots,n k = 1 , … , n and using A ∈ D A\in\mathcal{D} A ∈ D (so a n , k 2 ≤ K 2 a_{n,k}^2\le K^2 a n , k 2 ≤ K 2 ), we have
∑ k = 1 n E k − 1 ( d n , k 2 ) = ∑ k = 1 n τ ( d n , k 2 ) ⋅ 1 ≤ ( ∑ k = 1 n a n , k 2 ) H ( c ) ⋅ 1 ≤ n K 2 H ( c ) ⋅ 1 = D n 2 ⋅ 1. \sum_{k=1}^{n}E_{k-1}(d_{n,k}^2)=\sum_{k=1}^{n}\tau(d_{n,k}^2)\cdot1\le\left(\sum_{k=1}^{n}a_{n,k}^2\right)H(c)\cdot1\le nK^2H(c)\cdot1=D_n^2\cdot1. k = 1 ∑ n E k − 1 ( d n , k 2 ) = k = 1 ∑ n τ ( d n , k 2 ) ⋅ 1 ≤ ( k = 1 ∑ n a n , k 2 ) H ( c ) ⋅ 1 ≤ n K 2 H ( c ) ⋅ 1 = D n 2 ⋅ 1.
Proof of Theorem B(a)
Fix Δ > 0 \Delta>0 Δ > 0 and n ≥ 1 n\ge1 n ≥ 1 and let 1 ≤ k ≤ n 1\le k\le n 1 ≤ k ≤ n . Using same notation as Lemma 4.9 , define x k ( c n ) : = x k 1 [ 0 , c n ) ( ∣ x k ∣ ) x_k^{(c_n)}:=x_k1_{[0,c_n)}(|x_k|) x k ( c n ) := x k 1 [ 0 , c n ) ( ∣ x k ∣ ) , z k ( c n ) : = x k 1 [ c n , ∞ ) ( ∣ x k ∣ ) z_k^{(c_n)}:=x_k1_{[c_n,\infty)}(|x_k|) z k ( c n ) := x k 1 [ c n , ∞ ) ( ∣ x k ∣ ) , so that x k = x k ( c n ) + z k ( c n ) x_k=x_k^{(c_n)}+z_k^{(c_n)} x k = x k ( c n ) + z k ( c n ) . Here, we take c n : = Δ b n c_n:=\Delta b_n c n := Δ b n . Define, for 0 ≤ l ≤ n 0\le l\le n 0 ≤ l ≤ n ,
y l : = ∑ k = 1 l a n , k ( x k ( c n ) − τ ( x k ( c n ) ) ) , m l : = ∑ k = 1 l a n , k τ ( x k ( c n ) ) , v l : = ∑ k = 1 l a n , k z k ( c n ) . (38) y_l:=\sum_{k=1}^{l}a_{n,k}\left(x_k^{(c_n)}-\tau(x_k^{(c_n)})\right),\qquad m_l:=\sum_{k=1}^{l}a_{n,k}\tau(x_k^{(c_n)}),\qquad v_l:=\sum_{k=1}^{l}a_{n,k}z_k^{(c_n)}. \tag*{(38)} y l := k = 1 ∑ l a n , k ( x k ( c n ) − τ ( x k ( c n ) ) ) , m l := k = 1 ∑ l a n , k τ ( x k ( c n ) ) , v l := k = 1 ∑ l a n , k z k ( c n ) . (38)
Then, exactly,
S n ( A ) = y n + m n + v n . (39) S_n(A)=y_n+m_n+v_n. \tag*{(39)} S n ( A ) = y n + m n + v n . (39)
By Lemma 4.9 applied with c = Δ b n > 0 c=\Delta b_n>0 c = Δ b n > 0 , the difference d n , k : = a n , k ( x k ( c n ) − τ ( x k ( c n ) ) ) d_{n,k}:=a_{n,k}\left(x_k^{(c_n)}-\tau(x_k^{(c_n)})\right) d n , k := a n , k ( x k ( c n ) − τ ( x k ( c n ) ) ) satisfies E k − 1 ( d n , k ) = 0 E_{k-1}(d_{n,k})=0 E k − 1 ( d n , k ) = 0 for 1 ≤ k ≤ n 1\le k\le n 1 ≤ k ≤ n , and ( y l : = ∑ k = 1 l d n , k ) l = 0 n \left(y_l:=\sum_{k=1}^{l}d_{n,k}\right)_{l=0}^{n} ( y l := ∑ k = 1 l d n , k ) l = 0 n is a self-adjoint martingale. To prove Theorem B(a), we will set σ 2 : = τ ( x 1 2 ) < ∞ \sigma^2:=\tau(x_1^2)<\infty σ 2 := τ ( x 1 2 ) < ∞ .
Lemma 4.11. For every fixed Δ > 0 \Delta>0 Δ > 0 , m n / b n ⟶ 0 m_n/b_n\longrightarrow0 m n / b n ⟶ 0 as n → ∞ n\to\infty n → ∞ .
Proof. Since τ ( x k ) = 0 \tau(x_k)=0 τ ( x k ) = 0 for every k k k , τ ( x k ( c n ) ) = − τ ( z k ( c n ) ) \tau(x_k^{(c_n)})=-\tau(z_k^{(c_n)}) τ ( x k ( c n ) ) = − τ ( z k ( c n ) ) . Using ∣ τ ( y ) ∣ ≤ τ ( ∣ y ∣ ) |\tau(y)|\le\tau(|y|) ∣ τ ( y ) ∣ ≤ τ ( ∣ y ∣ ) and identical distribution,
∣ m n ∣ / b n ≤ K ⋅ n b n τ ( ∣ x 1 ∣ 1 [ Δ b n , ∞ ) ( ∣ x 1 ∣ ) ) = K n b n ∫ [ Δ b n , ∞ ) s d μ ( s ) . |m_n|/b_n\le K\cdot\frac{n}{b_n}\tau\left(|x_1|1_{[\Delta b_n,\infty)}(|x_1|)\right)=K\frac{n}{b_n}\int_{[\Delta b_n,\infty)}s\,d\mu(s). ∣ m n ∣/ b n ≤ K ⋅ b n n τ ( ∣ x 1 ∣ 1 [ Δ b n , ∞ ) ( ∣ x 1 ∣ ) ) = K b n n ∫ [ Δ b n , ∞ ) s d μ ( s ) .
By Proposition 4.7 , it follows that lim n → ∞ m n / b n = 0 \lim_{n\to\infty}m_n/b_n=0 lim n → ∞ m n / b n = 0 .
Lemma 4.12. For every fixed Δ > 0 \Delta>0 Δ > 0 , v n b n → a . u . 0 \frac{v_n}{b_n}\xrightarrow{\mathrm{a.u.}}0 b n v n a.u. 0 as n → ∞ n\to\infty n → ∞ .
Proof. Set q k : = 1 [ Δ b k , ∞ ) ( ∣ x k ∣ ) q_k:=1_{[\Delta b_k,\infty)}(|x_k|) q k := 1 [ Δ b k , ∞ ) ( ∣ x k ∣ ) . Now by Lemma 4.8 , we have ∑ k τ ( q k ) < ∞ \sum_k\tau(q_k)<\infty ∑ k τ ( q k ) < ∞ . Choose N N N with ∑ k ≥ N τ ( q k ) < ε 1 / 2 \sum_{k\ge N}\tau(q_k)<\varepsilon_1/2 ∑ k ≥ N τ ( q k ) < ε 1 /2 . Let p v ( 1 ) : = ⋀ k ≥ N ( 1 − q k ) p_v^{(1)}:=\bigwedge_{k\ge N}(1-q_k) p v ( 1 ) := ⋀ k ≥ N ( 1 − q k ) . This gives τ ( 1 − p v ( 1 ) ) < ε 1 / 2 \tau(1-p_v^{(1)})<\varepsilon_1/2 τ ( 1 − p v ( 1 ) ) < ε 1 /2 . Since B B B is increasing, we have 1 [ Δ b n , ∞ ) ( ∣ x k ∣ ) ≤ q k 1_{[\Delta b_n,\infty)}(|x_k|)\le q_k 1 [ Δ b n , ∞ ) ( ∣ x k ∣ ) ≤ q k , for n ≥ k ≥ N n\ge k\ge N n ≥ k ≥ N . It follows that x k 1 [ Δ b n , ∞ ) ( ∣ x k ∣ ) p v ( 1 ) = 0 x_k1_{[\Delta b_n,\infty)}(|x_k|)p_v^{(1)}=0 x k 1 [ Δ b n , ∞ ) ( ∣ x k ∣ ) p v ( 1 ) = 0 .
Now choose R < ∞ R<\infty R < ∞ , such that ∑ k < N τ ( 1 [ R , ∞ ) ( ∣ x k ∣ ) ) < ε 1 / 2 \sum_{k<N}\tau(1_{[R,\infty)}(|x_k|))<\varepsilon_1/2 ∑ k < N τ ( 1 [ R , ∞ ) ( ∣ x k ∣ )) < ε 1 /2 . Set p v ( 2 ) : = ⋀ k < N 1 [ 0 , R ) ( ∣ x k ∣ ) p_v^{(2)}:=\bigwedge_{k<N}1_{[0,R)}(|x_k|) p v ( 2 ) := ⋀ k < N 1 [ 0 , R ) ( ∣ x k ∣ ) . Since b n → ∞ b_n\to\infty b n → ∞ , there exists n 0 ∈ N n_0\in\mathbb{N} n 0 ∈ N , such that Δ b n ≥ R \Delta b_n\ge R Δ b n ≥ R , ∀ n ≥ n 0 \forall n\ge n_0 ∀ n ≥ n 0 . Then for all n ≥ n 0 n\ge n_0 n ≥ n 0 , we have x k 1 [ Δ b n , ∞ ) ( ∣ x k ∣ ) p v ( 2 ) = 0 x_k1_{[\Delta b_n,\infty)}(|x_k|)p_v^{(2)}=0 x k 1 [ Δ b n , ∞ ) ( ∣ x k ∣ ) p v ( 2 ) = 0 . Now take, p v = p v ( 1 ) ∧ p v ( 2 ) p_v=p_v^{(1)}\wedge p_v^{(2)} p v = p v ( 1 ) ∧ p v ( 2 ) . Then, τ ( 1 − p v ) < ε 1 \tau(1-p_v)<\varepsilon_1 τ ( 1 − p v ) < ε 1 and v n p v = 0 v_np_v=0 v n p v = 0 for all n ≥ n 0 n\ge n_0 n ≥ n 0 . In particular ∥ v n b n p v ∥ ∞ = 0 \left\lVert\frac{v_n}{b_n}p_v\right\rVert_\infty=0 b n v n p v ∞ = 0 for all n ≥ n 0 n\ge n_0 n ≥ n 0 .
This completes the proof. □
Lemma 4.13. Fix θ ∈ ( 0 , 2 α − 1 ) \theta\in(0,2\alpha-1) θ ∈ ( 0 , 2 α − 1 ) . For every ε > 0 \varepsilon>0 ε > 0 there exists Δ > 0 \Delta>0 Δ > 0 such that
(1) ∑ n = 1 ∞ τ ( 1 ( ε b n , ∞ ) ( ∣ y n ∣ ) ) < ∞ \sum_{n=1}^{\infty}\tau\left(1_{(\varepsilon b_n,\infty)}(|y_n|)\right)<\infty ∑ n = 1 ∞ τ ( 1 ( ε b n , ∞ ) ( ∣ y n ∣ ) ) < ∞ .
(2) lim sup n → ∞ a . u . y n b n ≤ ε \limsup_{n\to\infty}^{\mathrm{a.u.}}\frac{y_n}{b_n}\leq\varepsilon lim sup n → ∞ a.u. b n y n ≤ ε .
Proof. (1): Fix Δ > 0 \Delta>0 Δ > 0 and n ∈ N n\in\mathbb{N} n ∈ N . Now recall from (38) , ( y l ) l = 0 n (y_l)_{l=0}^{n} ( y l ) l = 0 n is a martingale. By Theorem 3.1 with M n = 2 K Δ b n M_n=2K\Delta b_n M n = 2 K Δ b n , D n 2 = K 2 σ 2 n D_n^2=K^2\sigma^2n D n 2 = K 2 σ 2 n as in Lemma 4.10 , for every λ > 0 \lambda>0 λ > 0 , we have τ ( e λ y n ) ≤ exp ( e λ M n − λ M n − 1 M n 2 D n 2 ) \tau\left(e^{\lambda y_n}\right)\leq\exp\left(\frac{e^{\lambda M_n}-\lambda M_n-1}{M_n^2}D_n^2\right) τ ( e λ y n ) ≤ exp ( M n 2 e λ M n − λ M n − 1 D n 2 ) .
Set t n : = ε b n t_n:=\varepsilon b_n t n := ε b n and
u n : = t n M n D n 2 = ε b n ⋅ 2 K Δ b n K 2 σ 2 n = 2 ε Δ K σ 2 b n 2 n . (40) u_n:=\frac{t_nM_n}{D_n^2}=\frac{\varepsilon b_n\cdot2K\Delta b_n}{K^2\sigma^2n}=\frac{2\varepsilon\Delta}{K\sigma^2}\frac{b_n^2}{n}. \tag*{(40)} u n := D n 2 t n M n = K 2 σ 2 n ε b n ⋅ 2 K Δ b n = K σ 2 2 ε Δ n b n 2 . (40)
Choose λ n : = log ( 1 + u n ) / M n \lambda_n:=\log(1+u_n)/M_n λ n := log ( 1 + u n ) / M n , so e λ n M n = 1 + u n e^{\lambda_nM_n}=1+u_n e λ n M n = 1 + u n . By functional calculus, we have
τ ( 1 ( t n , ∞ ) ( y n ) ) ≤ exp ( − λ n t n + D n 2 M n 2 ( u n − log ( 1 + u n ) ) ) . \tau\left(1_{(t_n,\infty)}(y_n)\right)\leq\exp\left(-\lambda_nt_n+\frac{D_n^2}{M_n^2}\left(u_n-\log(1+u_n)\right)\right). τ ( 1 ( t n , ∞ ) ( y n ) ) ≤ exp ( − λ n t n + M n 2 D n 2 ( u n − log ( 1 + u n ) ) ) .
Now write a n : = t n / M n = ε / ( 2 K Δ ) a_n:=t_n/M_n=\varepsilon/(2K\Delta) a n := t n / M n = ε / ( 2 K Δ ) and s n : = D n 2 / M n 2 s_n:=D_n^2/M_n^2 s n := D n 2 / M n 2 . Note that s n u n = a n s_nu_n=a_n s n u n = a n . Then notice that, − λ n t n + D n 2 M n 2 ( u n − log ( 1 + u n ) ) ≤ − a n ( log ( 1 + u n ) − 1 ) -\lambda_nt_n+\frac{D_n^2}{M_n^2}\left(u_n-\log(1+u_n)\right)\leq-a_n\left(\log(1+u_n)-1\right) − λ n t n + M n 2 D n 2 ( u n − log ( 1 + u n ) ) ≤ − a n ( log ( 1 + u n ) − 1 ) . Hence
τ ( 1 ( ε b n , ∞ ) ( y n ) ) ≤ exp ( − ε 2 K Δ ( log ( 1 + u n ) − 1 ) ) . (41) \tau\left(1_{(\varepsilon b_n,\infty)}(y_n)\right)\leq\exp\left(-\frac{\varepsilon}{2K\Delta}\left(\log(1+u_n)-1\right)\right). \tag*{(41)} τ ( 1 ( ε b n , ∞ ) ( y n ) ) ≤ exp ( − 2 K Δ ε ( log ( 1 + u n ) − 1 ) ) . (41)
Since b n 2 / n ∈ R V 2 α − 1 b_n^2/n\in\mathrm{RV}_{2\alpha-1} b n 2 / n ∈ RV 2 α − 1 and 2 α − 1 > θ > 0 2\alpha-1>\theta>0 2 α − 1 > θ > 0 , we have ( b n 2 / n ) / n θ ∈ R V 2 α − 1 − θ (b_n^2/n)/n^\theta\in\mathrm{RV}_{2\alpha-1-\theta} ( b n 2 / n ) / n θ ∈ RV 2 α − 1 − θ . Hence, by (RV1) there are n 1 n_1 n 1 and c θ > 0 c_\theta>0 c θ > 0 with b n 2 / n ≥ c θ n θ b_n^2/n\geq c_\theta n^\theta b n 2 / n ≥ c θ n θ for n ≥ n 1 n\geq n_1 n ≥ n 1 . Now, using eq. (40) , we have u n ≥ 2 ε Δ c θ K σ 2 n θ = : C ε , Δ , θ n θ u_n\geq\frac{2\varepsilon\Delta c_\theta}{K\sigma^2}n^\theta=:C_{\varepsilon,\Delta,\theta}n^\theta u n ≥ K σ 2 2 ε Δ c θ n θ =: C ε , Δ , θ n θ , n ≥ n 1 n\geq n_1 n ≥ n 1 , so log ( 1 + u n ) ≥ θ log n + log C ε , Δ , θ \log(1+u_n)\geq\theta\log n+\log C_{\varepsilon,\Delta,\theta} log ( 1 + u n ) ≥ θ log n + log C ε , Δ , θ for n ≥ n 1 n\geq n_1 n ≥ n 1 . Substituting into eq. (41) , for n ≥ n 1 n\geq n_1 n ≥ n 1 ,
τ ( 1 ( ε b n , ∞ ) ( y n ) ) ≤ exp ( − ε 2 K Δ ( θ log n + log C ε , Δ , θ ) ) ≤ C ε , Δ , θ ′ n − ε θ / ( 2 K Δ ) . \begin{aligned}
\tau\left(1_{(\varepsilon b_n,\infty)}(y_n)\right)\leq\exp\left(-\frac{\varepsilon}{2K\Delta}\left(\theta\log n+\log C_{\varepsilon,\Delta,\theta}\right)\right) \\
&\leq C'_{\varepsilon,\Delta,\theta}n^{-\varepsilon\theta/(2K\Delta)}.
\end{aligned} τ ( 1 ( ε b n , ∞ ) ( y n ) ) ≤ exp ( − 2 K Δ ε ( θ log n + log C ε , Δ , θ ) ) ≤ C ε , Δ , θ ′ n − εθ / ( 2 K Δ ) .
for a constant C ε , Δ , θ ′ = exp ( − ε 2 K Δ ( log C ε , Δ , θ − 1 ) ) C'_{\varepsilon,\Delta,\theta}=\exp\left(-\frac{\varepsilon}{2K\Delta}\left(\log C_{\varepsilon,\Delta,\theta}-1\right)\right) C ε , Δ , θ ′ = exp ( − 2 K Δ ε ( log C ε , Δ , θ − 1 ) ) not depending on n n n . Now choose Δ : = ε θ 4 K > 0 \Delta:=\frac{\varepsilon\theta}{4K}>0 Δ := 4 K εθ > 0 , so that ε θ / ( 2 K Δ ) = 2 > 1 \varepsilon\theta/(2K\Delta)=2>1 εθ / ( 2 K Δ ) = 2 > 1 . Then ∑ n τ ( 1 ( ε b n , ∞ ) ( y n ) ) < ∞ \sum_n\tau\left(1_{(\varepsilon b_n,\infty)}(y_n)\right)<\infty ∑ n τ ( 1 ( ε b n , ∞ ) ( y n ) ) < ∞ . The same argument applied to − y n -y_n − y n (with the same M n , D n 2 M_n,D_n^2 M n , D n 2 ) gives ∑ n τ ( 1 ( ε b n , ∞ ) ( − y n ) ) < ∞ \sum_n\tau\left(1_{(\varepsilon b_n,\infty)}(-y_n)\right)<\infty ∑ n τ ( 1 ( ε b n , ∞ ) ( − y n ) ) < ∞ .
(2): For each n n n , let e n : = 1 [ 0 , ε b n ] ( ∣ y n ∣ ) ∈ P ( M ) e_n:=1_{[0,\varepsilon b_n]}(|y_n|)\in\mathcal{P}(\mathcal{M}) e n := 1 [ 0 , ε b n ] ( ∣ y n ∣ ) ∈ P ( M ) . Then ∥ y n e n ∥ ∞ ≤ ε b n \|y_ne_n\|_\infty\leq\varepsilon b_n ∥ y n e n ∥ ∞ ≤ ε b n , and τ ( 1 − e n ) = τ ( 1 ( ε b n , ∞ ) ( ∣ y n ∣ ) ) \tau(1-e_n)=\tau\left(1_{(\varepsilon b_n,\infty)}(|y_n|)\right) τ ( 1 − e n ) = τ ( 1 ( ε b n , ∞ ) ( ∣ y n ∣ ) ) . By (1), ∑ n τ ( 1 − e n ) < ∞ \sum_n\tau(1-e_n)<\infty ∑ n τ ( 1 − e n ) < ∞ . Given δ > 0 \delta>0 δ > 0 , choose N ∈ N N\in\mathbb{N} N ∈ N with ∑ n ≥ N τ ( 1 − e n ) < δ \sum_{n\geq N}\tau(1-e_n)<\delta ∑ n ≥ N τ ( 1 − e n ) < δ . Now set p N : = ⋁ n ≥ N e n ∈ P ( M ) p_N:=\bigvee_{n\geq N}e_n\in\mathcal{P}(\mathcal{M}) p N := ⋁ n ≥ N e n ∈ P ( M ) . Then, τ ( 1 − p N ) ≤ ∑ n ≥ N τ ( 1 − e n ) < δ \tau(1-p_N)\leq\sum_{n\geq N}\tau(1-e_n)<\delta τ ( 1 − p N ) ≤ ∑ n ≥ N τ ( 1 − e n ) < δ . Since p N ≤ e n p_N\leq e_n p N ≤ e n , ∥ y n p N ∥ ∞ ≤ ∥ y n e n ∥ ∞ ≤ ε b n \|y_np_N\|_\infty\leq\|y_ne_n\|_\infty\leq\varepsilon b_n ∥ y n p N ∥ ∞ ≤ ∥ y n e n ∥ ∞ ≤ ε b n , for every n ≥ N n\geq N n ≥ N . So lim sup n ∥ ( y n / b n ) p N ∥ ∞ ≤ ε \limsup_n\|(y_n/b_n)p_N\|_\infty\leq\varepsilon lim sup n ∥ ( y n / b n ) p N ∥ ∞ ≤ ε . As δ > 0 \delta>0 δ > 0 was arbitrary, this proves the claim. □
Remark 4.14. The assumption x 1 ∈ L 2 ( M , τ ) x_1\in L_2(\mathcal{M},\tau) x 1 ∈ L 2 ( M , τ ) in Theorem B(a) is stronger than what is actually needed. It may be replaced by the following condition: there exists θ 0 > 0 \theta_0>0 θ 0 > 0 such that, for every Δ > 0 \Delta>0 Δ > 0 ,
lim sup n → ∞ n 1 + θ 0 b n 2 H ( Δ b n ) < ∞ . \limsup_{n\to\infty}\frac{n^{1+\theta_0}}{b_n^2}H(\Delta b_n)<\infty. n → ∞ lim sup b n 2 n 1 + θ 0 H ( Δ b n ) < ∞.
To see this, recall that the finite second moment assumption enters the proof of Lemma 4.13 only through the estimate D n 2 ≤ K 2 σ 2 n D_n^2 \le K^2\sigma^2 n D n 2 ≤ K 2 σ 2 n . If the condition above is assumed instead, one can use D n 2 ≤ K 2 n H ( Δ b n ) D_n^2 \le K^2 n H(\Delta b_n) D n 2 ≤ K 2 n H ( Δ b n ) , and the rest of the argument goes through in the same way. Consequently, the conclusion of Theorem B(a) still holds under this condition, even when x 1 x_1 x 1 does not have a finite second moment.
Now we will proof the part (a) of Theorem B.
Proof of Theorem B(a). Fix θ ∈ ( 0 , 2 α − 1 ) \theta\in(0,2\alpha-1) θ ∈ ( 0 , 2 α − 1 ) . For j ≥ 1 j \ge1 j ≥ 1 set ε j : = 2 − j \varepsilon_j := 2^{-j} ε j := 2 − j and Δ j : = Δ \Delta_j := \Delta Δ j := Δ as constructed in Lemma 4.13 .
Fix j ≥ 1 j \ge1 j ≥ 1 . By Lemma 4.11 applied with Δ = Δ j \Delta= \Delta_j Δ = Δ j , there is n j ′ ∈ N n'_j \in\mathbb{N} n j ′ ∈ N such that ∣ m n ∣ / b n < ε j \lvert m_n\rvert/b_n < \varepsilon_j ∣ m n ∣ / b n < ε j for all n ≥ n j ′ n \ge n'_j n ≥ n j ′ .
By Lemma 4.12 and Lemma 4.13 applied with Δ = Δ j \Delta= \Delta_j Δ = Δ j , there is a projection p j p_j p j and n j ′ ′ ∈ N n''_j \in\mathbb{N} n j ′′ ∈ N such that τ ( 1 − p j ) ≤ ε j \tau(1-p_j) \le\varepsilon_j τ ( 1 − p j ) ≤ ε j , and v n p j = 0 v_n p_j = 0 v n p j = 0 for all n ≥ n j ′ ′ n \ge n''_j n ≥ n j ′′ and
lim sup n → ∞ ∥ ( y n / b n ) p j ∥ ∞ ≤ ε j \limsup_{n \to\infty} \lVert(y_n/b_n)p_j\rVert_\infty\le\varepsilon_j n → ∞ lim sup ∥( y n / b n ) p j ∥ ∞ ≤ ε j
Now using (39) and the triangle inequality, for n ≥ max ( n j ′ , n j ′ ′ ) n \ge\max(n'_j,n''_j) n ≥ max ( n j ′ , n j ′′ ) we have,
∥ S n ( A ) b n p j ∥ ∞ ≤ ∥ y n b n p j ∥ ∞ + ε j + 0. \left\lVert\frac{S_n(A)}{b_n}p_j \right\rVert_\infty\le\left\lVert\frac{y_n}{b_n}p_j \right\rVert_\infty+\varepsilon_j+0. b n S n ( A ) p j ∞ ≤ b n y n p j ∞ + ε j + 0.
Therefore, we have lim sup n → ∞ ∥ S n ( A ) b n p j ∥ ∞ ≤ 2 ε j \limsup_{n\to\infty}\left\lVert\frac{S_n(A)}{b_n}p_j\right\rVert_\infty\le2\varepsilon_j lim sup n → ∞ b n S n ( A ) p j ∞ ≤ 2 ε j . Now fix δ > 0 \delta>0 δ > 0 and choose N ∈ N N\in\mathbb{N} N ∈ N with ∑ j ≥ N 2 ε j < δ \sum_{j\ge N}2\varepsilon_j<\delta ∑ j ≥ N 2 ε j < δ . Set q N : = ⋀ j ≥ N p j ∈ P ( M ) q_N:=\bigwedge_{j\ge N}p_j\in\mathcal{P}(\mathcal{M}) q N := ⋀ j ≥ N p j ∈ P ( M ) . Now note that, τ ( 1 − q N ) ≤ ∑ j ≥ N τ ( 1 − p j ) < ∑ j ≥ N 2 ε j < δ \tau(1-q_N)\le\sum_{j\ge N}\tau(1-p_j)<\sum_{j\ge N}2\varepsilon_j<\delta τ ( 1 − q N ) ≤ ∑ j ≥ N τ ( 1 − p j ) < ∑ j ≥ N 2 ε j < δ . Since q N ≤ p j q_N\le p_j q N ≤ p j for every j ≥ N j\ge N j ≥ N , we have
lim sup n → ∞ ∥ S n ( A ) b n q N ∥ ∞ ≤ 2 ε j for every j ≥ N . \limsup_{n \to\infty}\left\lVert\frac{S_n(A)}{b_n}q_N\right\rVert_\infty\le2\varepsilon_j\text{ for every }j\ge N. n → ∞ lim sup b n S n ( A ) q N ∞ ≤ 2 ε j for every j ≥ N .
Note that, lim j → ∞ ε j = 0 \lim_{j\to\infty}\varepsilon_j=0 lim j → ∞ ε j = 0 , hence lim n → ∞ ∥ S n ( A ) b n q N ∥ ∞ = 0 \lim_{n\to\infty}\left\lVert\frac{S_n(A)}{b_n}q_N\right\rVert_\infty=0 lim n → ∞ b n S n ( A ) q N ∞ = 0 . Therefore, S n ( A ) b n → a . u . 0 \frac{S_n(A)}{b_n}\xrightarrow{\mathrm{a.u.}}0 b n S n ( A ) a.u. 0 , as n → ∞ n\to\infty n → ∞ , □
Proof of mainthm:LIL deterministic weights(b)
Fix δ ∈ ( 0 , 1 ] \delta\in(0,1] δ ∈ ( 0 , 1 ] and Δ > 0 \Delta>0 Δ > 0 . Similar as § 4.1 , for 1 ≤ k ≤ n 1\le k\le n 1 ≤ k ≤ n set x k ( c n ) : = x k 1 [ 0 , c n ) ( ∣ x k ∣ ) x_k^{(c_n)}:=x_k1_{[0,c_n)}(\lvert x_k\rvert) x k ( c n ) := x k 1 [ 0 , c n ) (∣ x k ∣) , c n : = δ b n / log n c_n:=\delta b_n/\log n c n := δ b n / log n , and write x k = x k ( c n ) + x k 1 [ c n , Δ b n ) ( ∣ x k ∣ ) + x k 1 [ Δ b n , ∞ ) ( ∣ x k ∣ ) x_k=x_k^{(c_n)}+x_k1_{[c_n,\Delta b_n)}(\lvert x_k\rvert)+x_k1_{[\Delta b_n,\infty)}(\lvert x_k\rvert) x k = x k ( c n ) + x k 1 [ c n , Δ b n ) (∣ x k ∣) + x k 1 [ Δ b n , ∞ ) (∣ x k ∣) . Then we have,
S n ( A ) = y n + m n + w n + v n , (42) S_n(A)=y_n+m_n+w_n+v_n, \tag*{(42)} S n ( A ) = y n + m n + w n + v n , (42)
where y n : = ∑ k a n , k ( x k ( c n ) − τ ( x k ( c n ) ) ) y_n:=\sum_k a_{n,k}(x_k^{(c_n)}-\tau(x_k^{(c_n)})) y n := ∑ k a n , k ( x k ( c n ) − τ ( x k ( c n ) )) , m n : = ∑ k a n , k τ ( x k ( c n ) ) m_n:=\sum_k a_{n,k}\tau(x_k^{(c_n)}) m n := ∑ k a n , k τ ( x k ( c n ) ) , w n : = ∑ k a n , k x k 1 [ c n , Δ b n ) ( ∣ x k ∣ ) w_n:=\sum_k a_{n,k}x_k1_{[c_n,\Delta b_n)}(\lvert x_k\rvert) w n := ∑ k a n , k x k 1 [ c n , Δ b n ) (∣ x k ∣) and v n : = ∑ k a n , k x k 1 [ Δ b n , ∞ ) ( ∣ x k ∣ ) v_n:=\sum_k a_{n,k}x_k1_{[\Delta b_n,\infty)}(\lvert x_k\rvert) v n := ∑ k a n , k x k 1 [ Δ b n , ∞ ) (∣ x k ∣) .
If K = 0 K=0 K = 0 , then S n ( A ) = 0 S_n(A)=0 S n ( A ) = 0 for every n n n , and the conclusion is immediate. Hence assume K > 0 K>0 K > 0 .
Lemma 4.15. For every fixed Δ > 0 \Delta>0 Δ > 0 and every ε 1 > 0 \varepsilon_1>0 ε 1 > 0 , there is p v ∈ P ( M ) p_v\in\mathcal{P}(\mathcal{M}) p v ∈ P ( M ) and n 0 ∈ N n_0\in\mathbb{N} n 0 ∈ N such that τ ( 1 − p v ) < ε 1 \tau(1-p_v)<\varepsilon_1 τ ( 1 − p v ) < ε 1 and v n p v = 0 v_np_v=0 v n p v = 0 for all n ≥ n 0 n\ge n_0 n ≥ n 0 .
Proof. For the proof see Lemma 4.12. □
Lemma 4.16. Using the definition of m n m_n m n from (42) , we have lim n → ∞ m n b n = 0 \lim_{n\to\infty}\frac{m_n}{b_n}=0 lim n → ∞ b n m n = 0 .
Proof. Using the definition of x k ( c n ) x_k^{(c_n)} x k ( c n ) , notice that τ ( x k ( c n ) ) = − τ ( x k 1 [ c n , ∞ ) ( ∣ x k ∣ ) ) \tau(x_k^{(c_n)})=-\tau(x_k1_{[c_n,\infty)}(\lvert x_k\rvert)) τ ( x k ( c n ) ) = − τ ( x k 1 [ c n , ∞ ) (∣ x k ∣)) . Therefore,
∣ m n ∣ b n ≤ K ⋅ n b n ∫ [ c n , ∞ ) s d μ ( s ) . \frac{\lvert m_n\rvert}{b_n}\le K\cdot\frac{n}{b_n}\int_{[c_n,\infty)}s\,d\mu(s). b n ∣ m n ∣ ≤ K ⋅ b n n ∫ [ c n , ∞ ) s d μ ( s ) .
By Proposition 4.7 , it follows that lim n → ∞ m n b n = 0 \lim_{n\to\infty}\frac{m_n}{b_n}=0 lim n → ∞ b n m n = 0 .
Lemma 4.17. Let θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] and η > 0 \eta>0 η > 0 . Then there exists δ = δ ( θ , η ) ∈ ( 0 , 1 ] \delta=\delta(\theta,\eta)\in(0,1] δ = δ ( θ , η ) ∈ ( 0 , 1 ] such that, if c n = δ b n log n c_n=\frac{\delta b_n}{\log n} c n = l o g n δ b n and y n y_n y n as in (42) then the following holds. For every ε 0 > 0 \varepsilon_0>0 ε 0 > 0 , there exists p y ∈ P ( M ) p_y\in\mathcal{P}(\mathcal{M}) p y ∈ P ( M ) such that τ ( 1 − p y ) < ε 0 \tau(1-p_y)<\varepsilon_0 τ ( 1 − p y ) < ε 0 and
lim sup n → ∞ ∥ y n p y ∥ ∞ b n ≤ ( 1 + θ ) 2 K ℓ 2 + η . \limsup_{n\to\infty}\frac{\lVert y_np_y\rVert_\infty}{b_n}\le(1+\theta)\sqrt{2}K\sqrt{\ell^2+\eta}. n → ∞ lim sup b n ∥ y n p y ∥ ∞ ≤ ( 1 + θ ) 2 K ℓ 2 + η . Proof. By Lemma 4.10 with δ n = δ \delta_n=\delta δ n = δ , it follows that ∥ d n , k ∥ ∞ ≤ M n \lVert d_{n,k}\rVert_\infty\le M_n ∥ d n , k ∥ ∞ ≤ M n and ∑ k E k − 1 ( d n , k 2 ) ≤ D n 2 ⋅ 1 \sum_k E_{k-1}(d_{n,k}^2)\le D_n^2\cdot1 ∑ k E k − 1 ( d n , k 2 ) ≤ D n 2 ⋅ 1 , where M n : = 2 δ K b n / log n M_n:=2\delta Kb_n/\log n M n := 2 δ K b n / log n and D n 2 : = K 2 n H ( c n ) D_n^2:=K^2nH(c_n) D n 2 := K 2 n H ( c n ) . Since 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 , we have c n ≤ b n log n c_n\le\frac{b_n}{\log n} c n ≤ l o g n b n . As H H H is non-decreasing, H ( c n ) ≤ H ( b n log n ) H(c_n)\le H\left(\frac{b_n}{\log n}\right) H ( c n ) ≤ H ( l o g n b n ) . Hence, by condition (3) of Theorem B, there exists n 0 = n 0 ( η ) n_0=n_0(\eta) n 0 = n 0 ( η ) such that, for all n ≥ n 0 n\ge n_0 n ≥ n 0 ,
H ( c n ) ≤ L η 2 b n 2 n log n , H(c_n)\le\frac{L_\eta^2b_n^2}{n\log n}, H ( c n ) ≤ n log n L η 2 b n 2 , where we take L η : = ℓ 2 + η L_\eta:=\sqrt{\ell^2+\eta} L η := ℓ 2 + η . Consequently, D n 2 ≤ K 2 L η 2 b n 2 log n D_n^2\le K^2L_\eta^2\frac{b_n^2}{\log n} D n 2 ≤ K 2 L η 2 l o g n b n 2 , n ≥ n 0 n\ge n_0 n ≥ n 0 .
Fix θ > 0 \theta>0 θ > 0 , and define t θ , η : = ( 1 + θ ) 2 K L η t_{\theta,\eta}:=(1+\theta)\sqrt{2}KL_\eta t θ , η := ( 1 + θ ) 2 K L η and λ n : = 2 log n K L η b n \lambda_n:=\frac{\sqrt{2}\log n}{KL_\eta b_n} λ n := K L η b n 2 l o g n . Then, notice that λ n M n = 2 2 δ L η \lambda_nM_n=\frac{2\sqrt{2}\delta}{L_\eta} λ n M n = L η 2 2 δ . Now choose, 0 < δ ≤ min { 1 , 3 θ L η / ( 2 2 ) } 0<\delta\le\min\{1,3\theta L_\eta/(2\sqrt{2})\} 0 < δ ≤ min { 1 , 3 θ L η / ( 2 2 )} . Then λ n M n ≤ 3 θ \lambda_nM_n\le3\theta λ n M n ≤ 3 θ . Recall that y n = ∑ k = 1 n d n , k y_n=\sum_{k=1}^n d_{n,k} y n = ∑ k = 1 n d n , k . We now apply Theorem 3.1 ,
τ ( e λ n y n ) ≤ exp ( 1 + θ 2 λ n 2 D n 2 ) . \tau(e^{\lambda_n y_n})\le\exp\left(\frac{1+\theta}{2}\lambda_n^2D_n^2\right). τ ( e λ n y n ) ≤ exp ( 2 1 + θ λ n 2 D n 2 ) . Using the bound of D n 2 D_n^2 D n 2 , for all n ≥ n 0 n\ge n_0 n ≥ n 0 we have 1 + θ 2 λ n 2 D n 2 ≤ ( 1 + θ ) log n \frac{1+\theta}{2}\lambda_n^2D_n^2\le(1+\theta)\log n 2 1 + θ λ n 2 D n 2 ≤ ( 1 + θ ) log n . Moreover, λ n t θ , η b n = 2 ( 1 + θ ) log n \lambda_nt_{\theta,\eta}b_n=2(1+\theta)\log n λ n t θ , η b n = 2 ( 1 + θ ) log n . Hence, using functional calculus, we have τ ( 1 ( t θ , η b n , ∞ ) ( y n ) ) ≤ n − ( 1 + θ ) \tau(1_{(t_{\theta,\eta}b_n,\infty)}(y_n))\le n^{-(1+\theta)} τ ( 1 ( t θ , η b n , ∞ ) ( y n )) ≤ n − ( 1 + θ ) . The same argument applied to − y n -y_n − y n . Hence, ∑ n τ ( 1 ( t θ , η b n , ∞ ) ( ∣ y n ∣ ) ) ≤ 2 ∑ n n − ( 1 + θ ) < ∞ \sum_n\tau(1_{(t_{\theta,\eta}b_n,\infty)}(|y_n|))\le2\sum_n n^{-(1+\theta)}<\infty ∑ n τ ( 1 ( t θ , η b n , ∞ ) ( ∣ y n ∣ )) ≤ 2 ∑ n n − ( 1 + θ ) < ∞ . Let ε 0 > 0 \varepsilon_0>0 ε 0 > 0 . Then, there exists N = N ( ε 0 ) N=N(\varepsilon_0) N = N ( ε 0 ) such that
∑ n ≥ N τ ( 1 ( t θ , η b n , ∞ ) ( ∣ y n ∣ ) ) < ε 0 . \sum_{n\ge N}\tau(1_{(t_{\theta,\eta}b_n,\infty)}(|y_n|))<\varepsilon_0. n ≥ N ∑ τ ( 1 ( t θ , η b n , ∞ ) ( ∣ y n ∣ )) < ε 0 . Define p y : = ⋀ n ≥ N 1 [ 0 , t θ , η b n ] ( ∣ y n ∣ ) p_y:=\bigwedge_{n\ge N}1_{[0,t_{\theta,\eta}b_n]}(|y_n|) p y := ⋀ n ≥ N 1 [ 0 , t θ , η b n ] ( ∣ y n ∣ ) . Then, τ ( 1 − p y ) < ε 0 \tau(1-p_y)<\varepsilon_0 τ ( 1 − p y ) < ε 0 . Also functional calculus gives ∥ y n p y ∥ ∞ ≤ t θ , η b n \lVert y_np_y\rVert_\infty\le t_{\theta,\eta}b_n ∥ y n p y ∥ ∞ ≤ t θ , η b n , n ≥ N n\ge N n ≥ N . Hence,
lim sup n → ∞ ∥ y n p y ∥ ∞ b n ≤ ( 1 + θ ) 2 K ℓ 2 + η . \limsup_{n\to\infty}\frac{\lVert y_np_y\rVert_\infty}{b_n}\le(1+\theta)\sqrt{2}K\sqrt{\ell^2+\eta}. n → ∞ lim sup b n ∥ y n p y ∥ ∞ ≤ ( 1 + θ ) 2 K ℓ 2 + η .
The following estimate is inspired by [Cuz95, Lemma 2.3]. The main idea is to use dyadic blocks 2 k − 1 < n ≤ 2 k 2^{k-1}<n\le2^k 2 k − 1 < n ≤ 2 k and to use the same truncation levels c k ∗ c_k^* c k ∗ and Δ b k ∗ \Delta b_k^* Δ b k ∗ on each block. We then estimate the truncated second moment using G G G and regular variation. In the critical case α = 1 / 2 \alpha=1/2 α = 1/2 , a direct application of Karamata’s theorem is not possible, since the function t h ( t ) th(t) t h ( t ) has index − 1 -1 − 1 . To avoid this endpoint difficulty, we compare t h ( t ) th(t) t h ( t ) with ( Δ b k ∗ ) 1 / 2 t 1 / 2 h ( t ) (\Delta b_k^*)^{1/2}t^{1/2}h(t) ( Δ b k ∗ ) 1/2 t 1/2 h ( t ) . The function t 1 / 2 h ( t ) t^{1/2}h(t) t 1/2 h ( t ) has index − 3 / 2 -3/2 − 3/2 , so Karamata’s theorem can be applied.
Lemma 4.18. Fix β > 0 \beta>0 β > 0 and Δ > 0 \Delta>0 Δ > 0 . For k ≥ 1 k\ge1 k ≥ 1 , define b k ∗ : = B ( 2 k ) b_k^*:=B(2^k) b k ∗ := B ( 2 k ) , c k ∗ : = β b k ∗ / ( k log 2 ) c_k^*:=\beta b_k^*/(k\log2) c k ∗ := β b k ∗ / ( k log 2 ) . Then there exists k 1 = k 1 ( β , Δ ) k_1=k_1(\beta,\Delta) k 1 = k 1 ( β , Δ ) such that c k ∗ < Δ b k ∗ c_k^*<\Delta b_k^* c k ∗ < Δ b k ∗ , ∀ k ≥ k 1 \forall k\ge k_1 ∀ k ≥ k 1 . For k ≥ k 1 k\ge k_1 k ≥ k 1 , set
D k 2 : = 2 k ∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) . D_k^2:=2^k\int_{[c_k^*,\Delta b_k^*)}s^2\,d\mu(s). D k 2 := 2 k ∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) . If α = 1 2 \alpha=\frac12 α = 2 1 , then there exists a constant C = C ( β , Δ ) < ∞ C=C(\beta,\Delta)<\infty C = C ( β , Δ ) < ∞ such that D k 2 ( b k ∗ ) 2 ≤ C k − 1 / 2 \frac{D_k^2}{(b_k^*)^2}\le Ck^{-1/2} ( b k ∗ ) 2 D k 2 ≤ C k − 1/2 , k ≥ k 1 k\ge k_1 k ≥ k 1 . Consequently,
∑ k = 1 ∞ ( D k 2 ( b k ∗ ) 2 ) M 0 < ∞ , \sum_{k=1}^{\infty}\left(\frac{D_k^2}{(b_k^*)^2}\right)^{M_0}<\infty, k = 1 ∑ ∞ ( ( b k ∗ ) 2 D k 2 ) M 0 < ∞ , for every M 0 > 2 M_0>2 M 0 > 2 .
Proof. By Lemmas 4.5 and 4.6 , applied with δ = β \delta=\beta δ = β and n = 2 k n=2^k n = 2 k , we have
c k ∗ → ∞ and c k ∗ h ( c k ∗ ) ∼ K ( 1 2 , β ) b k ∗ 2 k . (43) c_k^* \to\infty\quad\text{and}\quad c_k^*h(c_k^*) \sim K\left(\frac{1}{2},\beta\right)\frac{b_k^*}{2^k}. \tag*{(43)} c k ∗ → ∞ and c k ∗ h ( c k ∗ ) ∼ K ( 2 1 , β ) 2 k b k ∗ . (43)
Now notice that, by (36) we have ∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) = H ( Δ b k ∗ ) − H ( c k ∗ ) \int_{[c_k^*,\Delta b_k^*)}s^2\,d\mu(s)=H(\Delta b_k^*)-H(c_k^*) ∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) = H ( Δ b k ∗ ) − H ( c k ∗ ) . Hence, by Lemma 4.3 (i), note that H ( Δ b k ∗ ) − H ( c k ∗ ) ≤ c k ∗ 2 G ( c k ∗ ) + ∫ c k ∗ Δ b k ∗ 2 t G ( t ) d t H(\Delta b_k^*)-H(c_k^*)\leq c_k^{*2}G(c_k^*)+\int_{c_k^*}^{\Delta b_k^*}2tG(t)\,dt H ( Δ b k ∗ ) − H ( c k ∗ ) ≤ c k ∗ 2 G ( c k ∗ ) + ∫ c k ∗ Δ b k ∗ 2 tG ( t ) d t . Define, η ( c ) : = sup t ≥ c G ( t ) h ( t ) \eta(c):=\sup_{t\geq c}\frac{G(t)}{h(t)} η ( c ) := sup t ≥ c h ( t ) G ( t ) . By condition (2) of Theorem B, lim c → ∞ η ( c ) = 0 \lim_{c\to\infty}\eta(c)=0 lim c → ∞ η ( c ) = 0 . Now, for every t ≥ c k ∗ t\geq c_k^* t ≥ c k ∗ we have G ( t ) ≤ η ( c k ∗ ) h ( t ) G(t)\leq\eta(c_k^*)h(t) G ( t ) ≤ η ( c k ∗ ) h ( t ) . Thus,
∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) ≤ η ( c k ∗ ) ( c k ∗ 2 h ( c k ∗ ) + ∫ c k ∗ Δ b k ∗ 2 t h ( t ) d t ) . (44) \int_{[c_k^*,\Delta b_k^*)}s^2\,d\mu(s)\leq\eta(c_k^*)\left(c_k^{*2}h(c_k^*)+\int_{c_k^*}^{\Delta b_k^*}2th(t)\,dt\right). \tag*{(44)} ∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) ≤ η ( c k ∗ ) ( c k ∗ 2 h ( c k ∗ ) + ∫ c k ∗ Δ b k ∗ 2 t h ( t ) d t ) . (44)
Since, α = 1 / 2 \alpha=1/2 α = 1/2 , t h ( t ) ∈ R V − 1 th(t)\in\mathrm{RV}_{-1} t h ( t ) ∈ RV − 1 . Hence, Karamata’s theorem can not be applied. We will now use a simple trick to avoid this problem.
For c k ∗ ≤ t < Δ b k ∗ c_k^*\leq t<\Delta b_k^* c k ∗ ≤ t < Δ b k ∗ , we have t ≤ Δ b k ∗ t\leq\Delta b_k^* t ≤ Δ b k ∗ . Hence, t = t 1 / 2 t 1 / 2 ≤ ( Δ b k ∗ ) 1 / 2 t 1 / 2 t=t^{1/2}t^{1/2}\leq(\Delta b_k^*)^{1/2}t^{1/2} t = t 1/2 t 1/2 ≤ ( Δ b k ∗ ) 1/2 t 1/2 . Now observe that,
∫ c k ∗ Δ b k ∗ 2 t h ( t ) d t ≤ 2 ( Δ b k ∗ ) 1 / 2 ∫ c k ∗ Δ b k ∗ t 1 / 2 h ( t ) d t ≤ 2 ( Δ b k ∗ ) 1 / 2 ∫ c k ∗ ∞ t 1 / 2 h ( t ) d t . (45) \int_{c_k^*}^{\Delta b_k^*}2th(t)\,dt\leq2(\Delta b_k^*)^{1/2}\int_{c_k^*}^{\Delta b_k^*}t^{1/2}h(t)\,dt\leq2(\Delta b_k^*)^{1/2}\int_{c_k^*}^{\infty}t^{1/2}h(t)\,dt. \tag*{(45)} ∫ c k ∗ Δ b k ∗ 2 t h ( t ) d t ≤ 2 ( Δ b k ∗ ) 1/2 ∫ c k ∗ Δ b k ∗ t 1/2 h ( t ) d t ≤ 2 ( Δ b k ∗ ) 1/2 ∫ c k ∗ ∞ t 1/2 h ( t ) d t . (45)
Now take, ϕ ( t ) : = t 1 / 2 h ( t ) \phi(t):=t^{1/2}h(t) ϕ ( t ) := t 1/2 h ( t ) . Since α = 1 / 2 \alpha=1/2 α = 1/2 , ϕ \phi ϕ has index − 3 / 2 -3/2 − 3/2 . Note that, ϕ \phi ϕ has index strictly smaller than − 1 -1 − 1 . Thus, by Fact (RV5), for every ε > 0 \varepsilon>0 ε > 0 there exists c ε > 0 c_\varepsilon>0 c ε > 0 such that, for every c ≥ c ε c\geq c_\varepsilon c ≥ c ε , ∫ c ∞ ϕ ( t ) d t ≤ 2 ( 1 + ε ) c 3 / 2 h ( c ) \int_c^\infty\phi(t)\,dt\leq2(1+\varepsilon)c^{3/2}h(c) ∫ c ∞ ϕ ( t ) d t ≤ 2 ( 1 + ε ) c 3/2 h ( c ) . Since c k ∗ → ∞ c_k^*\to\infty c k ∗ → ∞ , there exists k ε k_\varepsilon k ε such that c k ∗ ≥ c ε c_k^*\geq c_\varepsilon c k ∗ ≥ c ε , k ≥ k ε k\geq k_\varepsilon k ≥ k ε . Therefore,
∫ c k ∗ ∞ t 1 / 2 h ( t ) d t ≤ 2 ( 1 + ε ) ( c k ∗ ) 3 / 2 h ( c k ∗ ) . (46) \int_{c_k^*}^{\infty}t^{1/2}h(t)\,dt\leq2(1+\varepsilon)(c_k^*)^{3/2}h(c_k^*). \tag*{(46)} ∫ c k ∗ ∞ t 1/2 h ( t ) d t ≤ 2 ( 1 + ε ) ( c k ∗ ) 3/2 h ( c k ∗ ) . (46)
Now using the fact c k ∗ h ( c k ∗ ) ∼ K ( 1 2 , β ) b k ∗ 2 k c_k^*h(c_k^*)\sim K\left(\frac{1}{2},\beta\right)\frac{b_k^*}{2^k} c k ∗ h ( c k ∗ ) ∼ K ( 2 1 , β ) 2 k b k ∗ , there exists C 1 > 0 C_1>0 C 1 > 0 such that for all k ≥ k ε k\geq k_\varepsilon k ≥ k ε , we have ( c k ∗ ) 3 / 2 h ( c k ∗ ) ≤ C 1 ( c k ∗ ) 1 / 2 b k ∗ 2 k (c_k^*)^{3/2}h(c_k^*)\leq C_1(c_k^*)^{1/2}\frac{b_k^*}{2^k} ( c k ∗ ) 3/2 h ( c k ∗ ) ≤ C 1 ( c k ∗ ) 1/2 2 k b k ∗ . Since c k ∗ = β b k ∗ k log 2 c_k^*=\frac{\beta b_k^*}{k\log2} c k ∗ = k l o g 2 β b k ∗ , then for some constant C 2 > 0 C_2>0 C 2 > 0 we have ( c k ∗ ) 3 / 2 h ( c k ∗ ) ≤ C 2 ( b k ∗ ) 3 / 2 2 k k 1 / 2 (c_k^*)^{3/2}h(c_k^*)\leq C_2\frac{(b_k^*)^{3/2}}{2^k k^{1/2}} ( c k ∗ ) 3/2 h ( c k ∗ ) ≤ C 2 2 k k 1/2 ( b k ∗ ) 3/2 . Hence, using (46) , there exists C 3 > 0 C_3>0 C 3 > 0 such that ∫ c k ∗ ∞ t 1 / 2 h ( t ) d t ≤ C 3 ( b k ∗ ) 3 / 2 2 k k 1 / 2 \int_{c_k^*}^{\infty}t^{1/2}h(t)\,dt\leq C_3\frac{(b_k^*)^{3/2}}{2^k k^{1/2}} ∫ c k ∗ ∞ t 1/2 h ( t ) d t ≤ C 3 2 k k 1/2 ( b k ∗ ) 3/2 , for all k ≥ k ε k\geq k_\varepsilon k ≥ k ε . Now, combining with (45) , we have
∫ c k ∗ Δ b k ∗ 2 t h ( t ) d t ≤ C 4 ⋅ ( b k ∗ ) 2 2 k k 1 / 2 , (47) \int_{c_k^*}^{\Delta b_k^*}2th(t)\,dt\leq C_4\cdot\frac{(b_k^*)^2}{2^k k^{1/2}}, \tag*{(47)} ∫ c k ∗ Δ b k ∗ 2 t h ( t ) d t ≤ C 4 ⋅ 2 k k 1/2 ( b k ∗ ) 2 , (47)
for some constant C 4 > 0 C_4>0 C 4 > 0 . Now from (43) , there exists k 0 ∈ N k_0\in\mathbb{N} k 0 ∈ N such that for all k ≥ k 0 k\geq k_0 k ≥ k 0 , we have c k ∗ 2 h ( c k ∗ ) = c k ∗ ( c k ∗ h ( c k ∗ ) ) ≤ C 5 ( b k ∗ ) 2 2 k c_k^{*2}h(c_k^*)=c_k^*(c_k^*h(c_k^*))\leq C_5\frac{(b_k^*)^2}{2^k} c k ∗ 2 h ( c k ∗ ) = c k ∗ ( c k ∗ h ( c k ∗ )) ≤ C 5 2 k ( b k ∗ ) 2 , for some constant C 5 > 0 C_5>0 C 5 > 0 . Thus using (47) , there exists k 0 < k 1 ∈ N k_0<k_1\in\mathbb{N} k 0 < k 1 ∈ N and C 6 > 0 C_6>0 C 6 > 0 such that
c k ∗ 2 h ( c k ∗ ) + ∫ c k ∗ Δ b k ∗ 2 t h ( t ) d t ≤ C 6 ( b k ∗ ) 2 2 k k 1 / 2 . c_k^{*2}h(c_k^*)+\int_{c_k^*}^{\Delta b_k^*}2th(t)\,dt\leq C_6\frac{(b_k^*)^2}{2^k k^{1/2}}. c k ∗ 2 h ( c k ∗ ) + ∫ c k ∗ Δ b k ∗ 2 t h ( t ) d t ≤ C 6 2 k k 1/2 ( b k ∗ ) 2 .
Since η ( c k ∗ ) → 0 \eta(c_k^*)\to0 η ( c k ∗ ) → 0 , there exists k 2 k_2 k 2 such that, η ( c k ∗ ) ≤ 1 \eta(c_k^*)\leq1 η ( c k ∗ ) ≤ 1 , k ≥ k 2 k\geq k_2 k ≥ k 2 . Combining this with (44) , we obtain,
∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) ≤ C 6 ⋅ ( b k ∗ ) 2 2 k k 1 / 2 , for k ≥ k 3 : = max { k 1 , k 2 } . \int_{[c_k^*,\Delta b_k^*)}s^2\,d\mu(s)\leq C_6\cdot\frac{(b_k^*)^2}{2^k k^{1/2}},\quad\text{for }k\geq k_3:=\max\{k_1,k_2\}. ∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) ≤ C 6 ⋅ 2 k k 1/2 ( b k ∗ ) 2 , for k ≥ k 3 := max { k 1 , k 2 } .
Since D k 2 = 2 k ∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) D_k^2=2^k\int_{[c_k^*,\Delta b_k^*)}s^2\,d\mu(s) D k 2 = 2 k ∫ [ c k ∗ , Δ b k ∗ ) s 2 d μ ( s ) , for all k ≥ k 3 k\geq k_3 k ≥ k 3 , we have D k 2 / ( b k ∗ ) 2 ≤ C 6 k − 1 / 2 D_k^2/(b_k^*)^2\leq C_6k^{-1/2} D k 2 / ( b k ∗ ) 2 ≤ C 6 k − 1/2 . Finally, for M 0 > 2 M_0>2 M 0 > 2 we have ( D k 2 / ( b k ∗ ) 2 ) M 0 ≤ C k − M 0 / 2 (D_k^2/(b_k^*)^2)^{M_0}\leq Ck^{-M_0/2} ( D k 2 / ( b k ∗ ) 2 ) M 0 ≤ C k − M 0 /2 , where C = C 6 M 0 C=C_6^{M_0} C = C 6 M 0 . Since M 0 > 2 M_0>2 M 0 > 2 ,
∑ k = 1 ∞ ( D k 2 ( b k ∗ ) 2 ) M 0 < ∞ . \sum_{k=1}^{\infty}\left(\frac{D_k^2}{(b_k^*)^2}\right)^{M_0}<\infty. k = 1 ∑ ∞ ( ( b k ∗ ) 2 D k 2 ) M 0 < ∞.
The following lemma provides a non-commutative counterpart of [Cuz95, Lemma 2.2] and will play a key role in the proof of the intermediate truncation term.
Lemma 4.19. Let d k , i ∈ M k , i d_{k,i}\in\mathcal{M}_{k,i} d k , i ∈ M k , i , 1 ≤ i ≤ 2 k 1\le i\le2^k 1 ≤ i ≤ 2 k , be self-adjoint elements of M \mathcal{M} M satisfying E k , i − 1 ( d k , i ) = 0 \mathbb{E}_{k,i-1}(d_{k,i})=0 E k , i − 1 ( d k , i ) = 0 , ∥ d k , i ∥ ∞ ≤ b k \lVert d_{k,i}\rVert_{\infty}\le b_k ∥ d k , i ∥ ∞ ≤ b k ( 1 ≤ i ≤ 2 k ) (1\le i\le2^k) ( 1 ≤ i ≤ 2 k ) , where ( b k ) k ≥ 1 (b_k)_{k\ge1} ( b k ) k ≥ 1 is a sequence of positive reals with b k ↑ ∞ b_k\uparrow\infty b k ↑ ∞ . Put S k : = ∑ i = 1 2 k d k , i S_k:=\sum_{i=1}^{2^k}d_{k,i} S k := ∑ i = 1 2 k d k , i . Assume that there is a sequence ( D k ) k ≥ 1 (D_k)_{k\ge1} ( D k ) k ≥ 1 of nonnegative reals such that ∑ i = 1 2 k E k , i − 1 ( d k , i 2 ) ≤ D k 2 ⋅ 1 \sum_{i=1}^{2^k}\mathbb{E}_{k,i-1}(d_{k,i}^2)\le D_k^2\cdot1 ∑ i = 1 2 k E k , i − 1 ( d k , i 2 ) ≤ D k 2 ⋅ 1 and ∑ k = 1 ∞ ( D k 2 / b k 2 ) M < ∞ \sum_{k=1}^{\infty}(D_k^2/b_k^2)^M<\infty ∑ k = 1 ∞ ( D k 2 / b k 2 ) M < ∞ , for some constant M > 0 M>0 M > 0 . Then, lim sup k → ∞ a . u . S k b k ≤ M \limsup_{k\to\infty}^{\mathrm{a.u.}}\frac{S_k}{b_k}\le M lim sup k → ∞ a.u. b k S k ≤ M .
Proof. Let r k : = D k 2 / b k 2 r_k:=D_k^2/b_k^2 r k := D k 2 / b k 2 , k ≥ 1 k\ge1 k ≥ 1 . By hypothesis, ∑ k r k M < ∞ \sum_k r_k^M<\infty ∑ k r k M < ∞ . Therefore, r k → 0 r_k\to0 r k → 0 as k → ∞ k\to\infty k → ∞ . Fix ε , δ > 0 \varepsilon,\delta>0 ε , δ > 0 and set a : = M + δ a:=M+\delta a := M + δ . Fix k k k with r k > 0 r_k>0 r k > 0 . Define x k , i : = ∑ j = 1 i d k , j x_{k,i}:=\sum_{j=1}^{i}d_{k,j} x k , i := ∑ j = 1 i d k , j for 0 ≤ i ≤ 2 k 0\le i\le2^k 0 ≤ i ≤ 2 k , so that x k , 0 = 0 x_{k,0}=0 x k , 0 = 0 and x k , 2 k = S k x_{k,2^k}=S_k x k , 2 k = S k . ( x k , i ) i = 0 2 k (x_{k,i})_{i=0}^{2^k} ( x k , i ) i = 0 2 k is a self-adjoint martingale with respect to ( M k , i , E k , i ) i = 0 2 k (\mathcal{M}_{k,i},\mathbb{E}_{k,i})_{i=0}^{2^k} ( M k , i , E k , i ) i = 0 2 k , with differences d k , i d_{k,i} d k , i satisfying ∥ d k , i ∥ ∞ ≤ b k \lVert d_{k,i}\rVert_{\infty}\le b_k ∥ d k , i ∥ ∞ ≤ b k and ∑ i E k , i − 1 ( d k , i 2 ) ≤ D k 2 ⋅ 1 \sum_i\mathbb{E}_{k,i-1}(d_{k,i}^2)\le D_k^2\cdot1 ∑ i E k , i − 1 ( d k , i 2 ) ≤ D k 2 ⋅ 1 . Now, applying Theorem 3.1 , for every λ > 0 \lambda>0 λ > 0 ,
τ ( e λ S k ) ≤ exp ( r k ( e λ b k − λ b k − 1 ) ) . (48) \tau\left(e^{\lambda S_k}\right)\le\exp\left(r_k(e^{\lambda b_k}-\lambda b_k-1)\right). \tag*{(48)} τ ( e λ S k ) ≤ exp ( r k ( e λ b k − λ b k − 1 ) ) . (48)
Define q k : = 1 ( a b k , ∞ ) ( ∣ S k ∣ ) q_k:=1_{(ab_k,\infty)}(|S_k|) q k := 1 ( a b k , ∞ ) ( ∣ S k ∣ ) . Hence using eq. (48) , τ ( 1 ( a b k , ∞ ) ( S k ) ) ≤ exp ( − λ a b k + r k ( e λ b k − λ b k − 1 ) ) \tau(1_{(ab_k,\infty)}(S_k))\le\exp(-\lambda ab_k+r_k(e^{\lambda b_k}-\lambda b_k-1)) τ ( 1 ( a b k , ∞ ) ( S k )) ≤ exp ( − λa b k + r k ( e λ b k − λ b k − 1 )) . Writing t : = λ b k > 0 t:=\lambda b_k>0 t := λ b k > 0 ,
τ ( q k ) ≤ 2 exp ( f k ( t ) ) , f k ( t ) : = − a t + r k ( e t − t − 1 ) , t > 0. (49) \tau(q_k)\le2\exp(f_k(t)),\qquad f_k(t):=-at+r_k(e^t-t-1),\quad t>0. \tag*{(49)} τ ( q k ) ≤ 2 exp ( f k ( t )) , f k ( t ) := − a t + r k ( e t − t − 1 ) , t > 0. (49)
f k f_k f k attains its minimum at t k ∗ = log ( 1 + a r k ) > 0 t_k^*=\log\left(1+\frac{a}{r_k}\right)>0 t k ∗ = log ( 1 + r k a ) > 0 . Fix θ ∈ ( 0 , 1 2 ) \theta\in\left(0,\frac{1}{2}\right) θ ∈ ( 0 , 2 1 ) such that a ( 1 − 2 θ ) > M a(1-2\theta)>M a ( 1 − 2 θ ) > M . Since r k → 0 r_k\to0 r k → 0 , there is k 1 k_1 k 1 such that r k < 1 r_k<1 r k < 1 for all k ≥ k 1 k\ge k_1 k ≥ k 1 . Hence t k ∗ ≥ log a + log 1 r k t_k^*\ge\log a+\log\frac{1}{r_k} t k ∗ ≥ log a + log r k 1 , for k ≥ k 1 k\ge k_1 k ≥ k 1 . Since log ( 1 / r k ) → ∞ \log(1/r_k)\to\infty log ( 1/ r k ) → ∞ , there is k 2 ≥ k 1 k_2\ge k_1 k 2 ≥ k 1 such that log a ≥ − θ log ( 1 / r k ) \log a\ge-\theta\log(1/r_k) log a ≥ − θ log ( 1/ r k ) for all k ≥ k 2 k\ge k_2 k ≥ k 2 , giving
t k ∗ ≥ ( 1 − θ ) log 1 r k , for k ≥ k 2 . (50) t_k^*\ge(1-\theta)\log\frac{1}{r_k},\qquad\text{for }k\ge k_2. \tag*{(50)} t k ∗ ≥ ( 1 − θ ) log r k 1 , for k ≥ k 2 . (50)
Hence by (50) and t k ∗ = log ( 1 + a r k ) t_k^*=\log\left(1+\frac{a}{r_k}\right) t k ∗ = log ( 1 + r k a ) , we have f k ( t k ∗ ) ≤ a − a ( 1 − θ ) log 1 r k f_k(t_k^*)\le a-a(1-\theta)\log\frac{1}{r_k} f k ( t k ∗ ) ≤ a − a ( 1 − θ ) log r k 1 , for k ≥ k 2 k\ge k_2 k ≥ k 2 . Since log ( 1 / r k ) → ∞ \log(1/r_k)\to\infty log ( 1/ r k ) → ∞ , choose k 3 ≥ k 2 k_3\ge k_2 k 3 ≥ k 2 so that a ≤ θ a log ( 1 / r k ) a\le\theta a\log(1/r_k) a ≤ θ a log ( 1/ r k ) for k ≥ k 3 k\ge k_3 k ≥ k 3 . Then we have, f k ( t k ∗ ) ≤ − M log 1 r k f_k(t_k^*)\le-M\log\frac{1}{r_k} f k ( t k ∗ ) ≤ − M log r k 1 for k ≥ k 3 k\ge k_3 k ≥ k 3 . Hence, using (49) , we get τ ( q k ) ≤ 2 r k M \tau(q_k)\le2r_k^M τ ( q k ) ≤ 2 r k M , for k ≥ k 3 k\ge k_3 k ≥ k 3 .
Now by hypothesis and above argument we have, ∑ k ≥ k 3 τ ( q k ) ≤ 2 ∑ k ≥ k 3 r k M < ∞ \sum_{k\ge k_3}\tau(q_k)\le2\sum_{k\ge k_3}r_k^M<\infty ∑ k ≥ k 3 τ ( q k ) ≤ 2 ∑ k ≥ k 3 r k M < ∞ . Choose N ≥ k 3 N\ge k_3 N ≥ k 3 such that ∑ k ≥ N τ ( q k ) < ε \sum_{k\ge N}\tau(q_k)<\varepsilon ∑ k ≥ N τ ( q k ) < ε . Define p : = ⋀ k ≥ N ( 1 − q k ) ∈ P ( M ) p:=\bigwedge_{k\ge N}(1-q_k)\in\mathcal{P}(\mathcal{M}) p := ⋀ k ≥ N ( 1 − q k ) ∈ P ( M ) . Fix k ≥ N k\ge N k ≥ N . Then S k p = S k ( 1 − q k ) p S_kp=S_k(1-q_k)p S k p = S k ( 1 − q k ) p and ∥ S k ( 1 − q k ) ∥ ∞ ≤ a b k \lVert S_k(1-q_k)\rVert_\infty\le ab_k ∥ S k ( 1 − q k ) ∥ ∞ ≤ a b k . Therefore, ∥ S k p b k ∥ ∞ ≤ a = M + δ \left\|\frac{S_kp}{b_k}\right\|_\infty\le a=M+\delta b k S k p ∞ ≤ a = M + δ , for all k ≥ N k\ge N k ≥ N . Hence, lim sup k → ∞ a . u . S k b k ≤ M \limsup_{k\to\infty}^{\mathrm{a.u.}}\frac{S_k}{b_k}\le M lim sup k → ∞ a.u. b k S k ≤ M . □
The dyadic decomposition used in the following proof is inspired by [Cuz95, Lemma 2.3]. In particular, the main idea is to divide the integers into dyadic blocks 2 k − 1 < n ≤ 2 k 2^{k-1}<n\le2^k 2 k − 1 < n ≤ 2 k and use the same interval on each block.
Lemma 4.20. Let θ , η \theta,\eta θ , η be as in Lemma 4.17 , and let δ \delta δ be the corresponding constant. For every ρ > 0 \rho>0 ρ > 0 , there exists Δ = Δ ( ρ ) > 0 \Delta=\Delta(\rho)>0 Δ = Δ ( ρ ) > 0 such that, for every ε 1 > 0 \varepsilon_1>0 ε 1 > 0 , there is a projection p w ∈ P ( M ) p_w\in\mathcal{P}(\mathcal{M}) p w ∈ P ( M ) satisfying τ ( 1 − p w ) < ε 1 \tau(1-p_w)<\varepsilon_1 τ ( 1 − p w ) < ε 1 and
lim sup n → ∞ ∥ p w w n p w ∥ ∞ b n ≤ ρ . \limsup_{n\to\infty}\frac{\lVert p_ww_np_w\rVert_\infty}{b_n}\le\rho. n → ∞ lim sup b n ∥ p w w n p w ∥ ∞ ≤ ρ .
Proof. Recall that b k ∗ = B ( 2 k ) b_k^* = B(2^k) b k ∗ = B ( 2 k ) for all k ∈ N k \in\mathbb{N} k ∈ N . By Lemma 4.4, applied to B ∈ RV 1 / 2 B \in\operatorname{RV}_{1/2} B ∈ RV 1/2 with the fixed ratio λ = 1 / 2 \lambda= 1/2 λ = 1/2 , there exists C B ≥ 1 C_B \ge1 C B ≥ 1 and k 0 ∈ N k_0 \in\mathbb{N} k 0 ∈ N such that C B − 1 b k ∗ ≤ b n ≤ b k ∗ C_B^{-1}b_k^* \le b_n \le b_k^* C B − 1 b k ∗ ≤ b n ≤ b k ∗ and 2 k − 1 < n ≤ 2 k 2^{k-1} < n \le2^k 2 k − 1 < n ≤ 2 k , for k ≥ k 0 k \ge k_0 k ≥ k 0 . Following Lemma 4.18 , set β : = δ / C B \beta:= \delta/C_B β := δ / C B , c k ∗ : = β b k ∗ / ( k log 2 ) c_k^* := \beta b_k^*/(k\log2) c k ∗ := β b k ∗ / ( k log 2 ) . Fix k ≥ k 0 k \ge k_0 k ≥ k 0 and let 2 k − 1 < n ≤ 2 k 2^{k-1} < n \le2^k 2 k − 1 < n ≤ 2 k . Since log n ≤ k log 2 \log n \le k\log2 log n ≤ k log 2 , we have
c n = δ b n log n ≥ δ C B − 1 b k ∗ k log 2 = c k ∗ . c_n = \frac{\delta b_n}{\log n} \ge\frac{\delta C_B^{-1}b_k^*}{k\log2} = c_k^*. c n = log n δ b n ≥ k log 2 δ C B − 1 b k ∗ = c k ∗ .
Also, Δ b n ≤ Δ b k ∗ \Delta b_n \le\Delta b_k^* Δ b n ≤ Δ b k ∗ . Hence, [ c n , Δ b n ) ⊆ [ c k ∗ , Δ b k ∗ ) [c_n,\Delta b_n) \subseteq[c_k^*,\Delta b_k^*) [ c n , Δ b n ) ⊆ [ c k ∗ , Δ b k ∗ ) . For 1 ≤ i ≤ 2 k 1 \le i \le2^k 1 ≤ i ≤ 2 k , define
z i , k : = ∣ x i ∣ 1 [ c k ∗ , Δ b k ∗ ) ( ∣ x i ∣ ) − τ ( ∣ x i ∣ 1 [ c k ∗ , Δ b k ∗ ) ( ∣ x i ∣ ) ) 1. z_{i,k} := |x_i|1_{[c_k^*,\Delta b_k^*)}(|x_i|)-\tau\left(|x_i|1_{[c_k^*,\Delta b_k^*)}(|x_i|)\right)\mathbf{1}. z i , k := ∣ x i ∣ 1 [ c k ∗ , Δ b k ∗ ) ( ∣ x i ∣ ) − τ ( ∣ x i ∣ 1 [ c k ∗ , Δ b k ∗ ) ( ∣ x i ∣ ) ) 1 .
Set, u k : = ∑ i = 1 2 k z i , k u_k := \sum_{i=1}^{2^k}z_{i,k} u k := ∑ i = 1 2 k z i , k and h k : = ∑ i = 1 2 k ∣ x i ∣ 1 [ c k ∗ , Δ b k ∗ ) ( ∣ x i ∣ ) h_k := \sum_{i=1}^{2^k}|x_i|1_{[c_k^*,\Delta b_k^*)}(|x_i|) h k := ∑ i = 1 2 k ∣ x i ∣ 1 [ c k ∗ , Δ b k ∗ ) ( ∣ x i ∣ ) . Since ( x i ) (x_i) ( x i ) are i.i.d, writing a k : = 2 k τ ( ∣ x 1 ∣ 1 [ c k ∗ , Δ b k ∗ ) ( ∣ x 1 ∣ ) ) a_k := 2^k\tau\left(|x_1|1_{[c_k^*,\Delta b_k^*)}(|x_1|)\right) a k := 2 k τ ( ∣ x 1 ∣ 1 [ c k ∗ , Δ b k ∗ ) ( ∣ x 1 ∣ ) ) , we have h k = u k + a k 1 h_k=u_k+a_k\mathbf{1} h k = u k + a k 1 . By Lemma 4.9 , we have E k , i − 1 ( z i , k ) = 0 E_{k,i-1}(z_{i,k})=0 E k , i − 1 ( z i , k ) = 0 , ∥ z i , k ∥ ∞ ≤ 2 Δ b k ∗ \|z_{i,k}\|_\infty\le2\Delta b_k^* ∥ z i , k ∥ ∞ ≤ 2Δ b k ∗ , and ∑ i = 1 2 k E k , i − 1 ( z i , k 2 ) ≤ D k 2 1 \sum_{i=1}^{2^k}E_{k,i-1}(z_{i,k}^2)\le D_k^2\mathbf{1} ∑ i = 1 2 k E k , i − 1 ( z i , k 2 ) ≤ D k 2 1 , where D k 2 D_k^2 D k 2 is from Lemma 4.18 . By Lemma 4.18 , with the above choices of β \beta β and Δ \Delta Δ , we have
∑ k = 1 ∞ ( D k 2 ( 2 Δ b k ∗ ) 2 ) 3 < ∞ . \sum_{k=1}^{\infty}\left(\frac{D_k^2}{(2\Delta b_k^*)^2}\right)^3<\infty. k = 1 ∑ ∞ ( ( 2Δ b k ∗ ) 2 D k 2 ) 3 < ∞.
Therefore, Lemma 4.19 applies to the row ( z i , k ) 1 ≤ i ≤ 2 k (z_{i,k})_{1\le i\le2^k} ( z i , k ) 1 ≤ i ≤ 2 k with bound 2 Δ b k ∗ 2\Delta b_k^* 2Δ b k ∗ . Thus, for every ε 1 > 0 \varepsilon_1>0 ε 1 > 0 , there exists a projection p w ∈ P ( M ) p_w\in\mathcal{P}(\mathcal{M}) p w ∈ P ( M ) such that τ ( 1 − p w ) < ε 1 \tau(1-p_w)<\varepsilon_1 τ ( 1 − p w ) < ε 1 and lim sup k → ∞ ∥ u k p w ∥ ∞ / ( 2 Δ b k ∗ ) ≤ 3 \limsup_{k\to\infty}\|u_kp_w\|_\infty/(2\Delta b_k^*)\le3 lim sup k → ∞ ∥ u k p w ∥ ∞ / ( 2Δ b k ∗ ) ≤ 3 . Hence,
lim sup k → ∞ ∥ u k p w ∥ ∞ b k ∗ ≤ 6 Δ . (51) \limsup_{k\to\infty}\frac{\|u_kp_w\|_\infty}{b_k^*}\le6\Delta. \tag*{(51)} k → ∞ lim sup b k ∗ ∥ u k p w ∥ ∞ ≤ 6Δ. (51)
Next, notice that
0 ≤ a k b k ∗ = 2 k b k ∗ ∫ [ c k ∗ , Δ b k ∗ ) s d μ ( s ) ≤ 2 k b k ∗ ∫ [ c k ∗ , ∞ ) s d μ ( s ) . 0\le\frac{a_k}{b_k^*}=\frac{2^k}{b_k^*}\int_{[c_k^*,\Delta b_k^*)}s\,d\mu(s)\le\frac{2^k}{b_k^*}\int_{[c_k^*,\infty)}s\,d\mu(s). 0 ≤ b k ∗ a k = b k ∗ 2 k ∫ [ c k ∗ , Δ b k ∗ ) s d μ ( s ) ≤ b k ∗ 2 k ∫ [ c k ∗ , ∞ ) s d μ ( s ) .
Now Proposition 4.7 with n = 2 k n=2^k n = 2 k and δ = β \delta=\beta δ = β gives a k / b k ∗ ⟶ 0 a_k/b_k^*\longrightarrow0 a k / b k ∗ ⟶ 0 as k → ∞ k\to\infty k → ∞ . Since a k a_k a k is a scalar, ∥ a k p w ∥ ∞ = a k \|a_kp_w\|_\infty=a_k ∥ a k p w ∥ ∞ = a k . Therefore, from h k = u k + a k 1 h_k=u_k+a_k\mathbf{1} h k = u k + a k 1 and (51) , we have
lim sup k → ∞ ∥ h k p w ∥ ∞ b k ∗ ≤ lim sup k → ∞ ∥ u k p w ∥ ∞ b k ∗ + lim sup k → ∞ a k b k ∗ ≤ 6 Δ . (52) \limsup_{k\to\infty}\frac{\|h_kp_w\|_\infty}{b_k^*}\le\limsup_{k\to\infty}\frac{\|u_kp_w\|_\infty}{b_k^*}+\limsup_{k\to\infty}\frac{a_k}{b_k^*}\le6\Delta. \tag*{(52)} k → ∞ lim sup b k ∗ ∥ h k p w ∥ ∞ ≤ k → ∞ lim sup b k ∗ ∥ u k p w ∥ ∞ + k → ∞ lim sup b k ∗ a k ≤ 6Δ. (52)
Let 2 k − 1 < n ≤ 2 k 2^{k-1}<n\le2^k 2 k − 1 < n ≤ 2 k . Since, [ c n , Δ b n ) ⊆ [ c k ∗ , Δ b k ∗ ) [c_n,\Delta b_n)\subseteq[c_k^*,\Delta b_k^*) [ c n , Δ b n ) ⊆ [ c k ∗ , Δ b k ∗ ) and K = sup n , k ∣ a n , k ∣ K=\sup_{n,k}|a_{n,k}| K = sup n , k ∣ a n , k ∣ we have,
− K h k ≤ w n ≤ K h k ⟹ ∥ p w w n p w ∥ ∞ ≤ K ∥ p w h k p w ∥ ∞ ≤ K ∥ h k p w ∥ ∞ . (53) -Kh_k\le w_n\le Kh_k\quad\Longrightarrow\quad\|p_ww_np_w\|_\infty\le K\|p_wh_kp_w\|_\infty\le K\|h_kp_w\|_\infty. \tag*{(53)} − K h k ≤ w n ≤ K h k ⟹ ∥ p w w n p w ∥ ∞ ≤ K ∥ p w h k p w ∥ ∞ ≤ K ∥ h k p w ∥ ∞ . (53)
Since, b n ≥ C B − 1 b k ∗ b_n\ge C_B^{-1}b_k^* b n ≥ C B − 1 b k ∗ , we have b n − 1 ≤ C B / b k ∗ b_n^{-1}\le C_B/b_k^* b n − 1 ≤ C B / b k ∗ . Combining this with (53) and (52) we have lim sup n → ∞ ∥ p w w n p w ∥ ∞ / b n ≤ 6 K C B Δ \limsup_{n\to\infty}\|p_ww_np_w\|_\infty/b_n\le6KC_B\Delta lim sup n → ∞ ∥ p w w n p w ∥ ∞ / b n ≤ 6 K C B Δ . Now, choose Δ : = ρ 6 K C B \Delta:=\frac{\rho}{6KC_B} Δ := 6 K C B ρ . Then, lim sup n → ∞ ∥ p w w n p w ∥ ∞ / b n ≤ ρ \limsup_{n\to\infty}\|p_ww_np_w\|_\infty/b_n\le\rho lim sup n → ∞ ∥ p w w n p w ∥ ∞ / b n ≤ ρ . Together with τ ( 1 − p w ) < ε 1 \tau(1-p_w)<\varepsilon_1 τ ( 1 − p w ) < ε 1 , this proves the lemma. □
Proof of Theorem B(b). Fix ε 0 , ε > 0 \varepsilon_0,\varepsilon>0 ε 0 , ε > 0 . Choose θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] , η > 0 \eta>0 η > 0 , ρ > 0 \rho>0 ρ > 0 so small that ( 1 + θ ) 2 K ℓ 2 + η + ρ < 2 K ℓ + ε (1+\theta)\sqrt{2K}\sqrt{\ell^2+\eta}+\rho<\sqrt{2K\ell}+\varepsilon ( 1 + θ ) 2 K ℓ 2 + η + ρ < 2 K ℓ + ε . Let δ \delta δ and Δ = Δ ( ρ ) \Delta=\Delta(\rho) Δ = Δ ( ρ ) be as given by Lemma 4.17 and Lemma 4.20 , respectively. Now notice that, using Lemma 4.17 , there exists a projection p y ∈ P ( M ) p_y\in\mathcal{P}(\mathcal{M}) p y ∈ P ( M ) and N y ∈ N N_y\in\mathbb{N} N y ∈ N such that τ ( 1 − p y ) < ε 0 3 \tau(1-p_y)<\frac{\varepsilon_0}{3} τ ( 1 − p y ) < 3 ε 0 and
∥ y n p y ∥ ∞ ≤ ( 1 + θ ) 2 K ℓ 2 + η b n , ∀ n ≥ N y . (54) \|y_np_y\|_\infty\le(1+\theta)\sqrt{2K}\sqrt{\ell^2+\eta}\,b_n,\quad\forall n\ge N_y. \tag*{(54)} ∥ y n p y ∥ ∞ ≤ ( 1 + θ ) 2 K ℓ 2 + η b n , ∀ n ≥ N y . (54)
By Lemma 4.20 , there exist p w ∈ P ( M ) p_w \in\mathcal{P}(\mathcal{M}) p w ∈ P ( M ) and N w ≥ 1 N_w \ge1 N w ≥ 1 such that τ ( 1 − p w ) < ε 0 3 \tau(1-p_w)<\frac{\varepsilon_0}{3} τ ( 1 − p w ) < 3 ε 0 and
∥ p w w n p w ∥ ∞ b n ≤ ρ + ε , ∀ n ≥ N w . (55) \frac{\lVert p_w w_n p_w\rVert_\infty}{b_n}\leq\rho+\varepsilon,\quad\forall n\geq N_w. \tag*{(55)} b n ∥ p w w n p w ∥ ∞ ≤ ρ + ε , ∀ n ≥ N w . (55)
By Lemma 4.15, there exist p v ∈ P ( M ) p_v \in\mathcal{P}(\mathcal{M}) p v ∈ P ( M ) and N v ≥ 1 N_v \geq1 N v ≥ 1 such that τ ( 1 − p v ) < ε 0 3 \tau(1-p_v)<\frac{\varepsilon_0}{3} τ ( 1 − p v ) < 3 ε 0 and
p v v n p v = 0 , ∀ n ≥ N v . (56) p_v v_n p_v=0,\quad\forall n\geq N_v. \tag*{(56)} p v v n p v = 0 , ∀ n ≥ N v . (56)
Now, define p : = p y ∧ p w ∧ p v p:=p_y\wedge p_w\wedge p_v p := p y ∧ p w ∧ p v . Then, τ ( 1 − p ) < ε 0 \tau(1-p)<\varepsilon_0 τ ( 1 − p ) < ε 0 . Set, N : = max { N y , N v , N w } N:=\max\{N_y,N_v,N_w\} N := max { N y , N v , N w } . Therefore, using (42) and (4.23), for every n ≥ N n\geq N n ≥ N , ∥ p S n ( A ) p ∥ ∞ ≤ ∥ y n p y ∥ ∞ + ∣ m n ∣ + ∥ p w w n p w ∥ ∞ \lVert pS_n(A)p\rVert_\infty\leq\lVert y_np_y\rVert_\infty+|m_n|+\lVert p_ww_np_w\rVert_\infty ∥ p S n ( A ) p ∥ ∞ ≤ ∥ y n p y ∥ ∞ + ∣ m n ∣ + ∥ p w w n p w ∥ ∞ . Now using Equations (54) and (4.22) and lemma 4.16, we have for every n ≥ N n\geq N n ≥ N , we have
∥ p S n ( A ) p ∥ ∞ b n ≤ ( 1 + θ ) 2 K ℓ 2 + η + ∣ m n ∣ b n + ∥ p w w n p w ∥ ∞ b n . \frac{\lVert pS_n(A)p\rVert_\infty}{b_n}\leq(1+\theta)\sqrt{2K}\sqrt{\ell^2+\eta}+\frac{|m_n|}{b_n}+\frac{\lVert p_ww_np_w\rVert_\infty}{b_n}. b n ∥ p S n ( A ) p ∥ ∞ ≤ ( 1 + θ ) 2 K ℓ 2 + η + b n ∣ m n ∣ + b n ∥ p w w n p w ∥ ∞ .
Taking limsup both side, and using Lemma 4.16 and eq. (4.22), we have lim sup n → ∞ ∥ p S n ( A ) p ∥ ∞ b n ≤ 2 K ℓ + ε \limsup_{n\to\infty}\frac{\lVert pS_n(A)p\rVert_\infty}{b_n}\leq\sqrt{2}K\ell+\varepsilon lim sup n → ∞ b n ∥ p S n ( A ) p ∥ ∞ ≤ 2 K ℓ + ε .
Since ε , ε 0 > 0 \varepsilon,\varepsilon_0>0 ε , ε 0 > 0 are arbitrary, we have
lim sup n → ∞ S n ( A ) b n ≤ 2 K ℓ b.a.u. \underset{n\to\infty}{\limsup}\frac{S_n(A)}{b_n}\leq\sqrt{2}K\ell
\quad\text{b.a.u.} n → ∞ lim sup b n S n ( A ) ≤ 2 K ℓ b.a.u.
Acknowledgement
Acknowledgement. DS acknowledges the Institute of Mathematics of the Polish Academy of Sciences, for providing the institutional support during the completion of this work. SP gratefully acknowledges the Indian Statistical Institute for financial support through its doctoral fellowship. He also wishes to thank his PhD supervisor, Prof. Issan Patri, for his valuable guidance and support.
References
@page 32/33 starts_mid=1 ends_mid=1 By Lemma 4.20 , there exist p w ∈ P ( M ) p_w \in\mathcal{P}(\mathcal{M}) p w ∈ P ( M ) and N w ≥ 1 N_w \ge1 N w ≥ 1 such that τ ( 1 − p w ) < ε 0 3 \tau(1-p_w)<\frac{\varepsilon_0}{3} τ ( 1 − p w ) < 3 ε 0 and
∥ p w w n p w ∥ ∞ b n ≤ ρ + ε , ∀ n ≥ N w . (57) \frac{\lVert p_w w_n p_w\rVert_\infty}{b_n}\leq\rho+\varepsilon,\quad\forall n\geq N_w. \tag*{(57)} b n ∥ p w w n p w ∥ ∞ ≤ ρ + ε , ∀ n ≥ N w . (57)
By Lemma 4.15, there exist p v ∈ P ( M ) p_v \in\mathcal{P}(\mathcal{M}) p v ∈ P ( M ) and N v ≥ 1 N_v \geq1 N v ≥ 1 such that τ ( 1 − p v ) < ε 0 3 \tau(1-p_v)<\frac{\varepsilon_0}{3} τ ( 1 − p v ) < 3 ε 0 and
p v v n p v = 0 , ∀ n ≥ N v . (58) p_v v_n p_v=0,\quad\forall n\geq N_v. \tag*{(58)} p v v n p v = 0 , ∀ n ≥ N v . (58)
Now, define p : = p y ∧ p w ∧ p v p:=p_y\wedge p_w\wedge p_v p := p y ∧ p w ∧ p v . Then, τ ( 1 − p ) < ε 0 \tau(1-p)<\varepsilon_0 τ ( 1 − p ) < ε 0 . Set, N : = max { N y , N v , N w } N:=\max\{N_y,N_v,N_w\} N := max { N y , N v , N w } . Therefore, using (42) and (4.23), for every n ≥ N n\geq N n ≥ N , ∥ p S n ( A ) p ∥ ∞ ≤ ∥ y n p y ∥ ∞ + ∣ m n ∣ + ∥ p w w n p w ∥ ∞ \lVert pS_n(A)p\rVert_\infty\leq\lVert y_np_y\rVert_\infty+|m_n|+\lVert p_ww_np_w\rVert_\infty ∥ p S n ( A ) p ∥ ∞ ≤ ∥ y n p y ∥ ∞ + ∣ m n ∣ + ∥ p w w n p w ∥ ∞ . Now using Equations (54) and (4.22) and lemma 4.16, we have for every n ≥ N n\geq N n ≥ N , we have
∥ p S n ( A ) p ∥ ∞ b n ≤ ( 1 + θ ) 2 K ℓ 2 + η + ∣ m n ∣ b n + ∥ p w w n p w ∥ ∞ b n . \frac{\lVert pS_n(A)p\rVert_\infty}{b_n}\leq(1+\theta)\sqrt{2K}\sqrt{\ell^2+\eta}+\frac{|m_n|}{b_n}+\frac{\lVert p_ww_np_w\rVert_\infty}{b_n}. b n ∥ p S n ( A ) p ∥ ∞ ≤ ( 1 + θ ) 2 K ℓ 2 + η + b n ∣ m n ∣ + b n ∥ p w w n p w ∥ ∞ .
Taking limsup both side, and using Lemma 4.16 and eq. (4.22), we have lim sup n → ∞ ∥ p S n ( A ) p ∥ ∞ b n ≤ 2 K ℓ + ε \limsup_{n\to\infty}\frac{\lVert pS_n(A)p\rVert_\infty}{b_n}\leq\sqrt{2}K\ell+\varepsilon lim sup n → ∞ b n ∥ p S n ( A ) p ∥ ∞ ≤ 2 K ℓ + ε .
Since ε , ε 0 > 0 \varepsilon,\varepsilon_0>0 ε , ε 0 > 0 are arbitrary, we have
lim sup n → ∞ S n ( A ) b n ≤ 2 K ℓ b.a.u. \underset{n\to\infty}{\limsup}\frac{S_n(A)}{b_n}\leq\sqrt{2}K\ell
\quad\text{b.a.u.} n → ∞ lim sup b n S n ( A ) ≤ 2 K ℓ b.a.u.
Acknowledgement
Acknowledgement. DS acknowledges the Institute of Mathematics of the Polish Academy of Sciences, for providing the institutional support during the completion of this work. SP gratefully acknowledges the Indian Statistical Institute for financial support through its doctoral fellowship. He also wishes to thank his PhD supervisor, Prof. Issan Patri, for his valuable guidance and support.
References
Theoretical Statistics and Mathematics Unit, Indian Statistical Institute, Delhi, India
Email address: souravpanja07@gmail.com
Theoretical Statistics and Mathematics Unit, Indian Statistical Institute, Kolkata, India
Email address: dptshs@gmail.com