betaFeedback

WorksMathematics

Weighted Logarithm Laws in non-commutative Probability

Abstract

We prove analogues of weighted laws of iterated logarithm and laws of logarithm for non-commutative martingales and independent random variables respectively. First we obtain a weighted LIL for non-commutative martingales, extending the sharp unweighted result of arXiv:2509.22037v2 to the case for polynomial weights, then by approximation to a broader class of functions. Then we proved a weighted law of logarithm for triangular array of weights and replacing the coefficient in the denominator by regularly varying functions with index αα. We show that the normalised sums converge to 0 for α>1/2α> 1/2 and for the case α=1/2α=1/2, we obtain the sharp upper bound with constant 2\sqrt{2}.

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 Sn=∑k=1nXkS_n = \sum_{k=1}^{n} X_k, where (Xk)(X_k) is a sequence of centered independent random variables. In the classical finite-variance case, the LIL uses the scale 2sn2log⁡log⁡sn2\sqrt{2s_n^2 \log\log s_n^2}, where sn2=∑k=1nE(Xk2)s_n^2 = \sum_{k=1}^{n} \mathbb{E}(X_k^2). For independent identically distributed (i.i.d) random variables with variance one, this becomes 2nlog⁡log⁡n\sqrt{2n\log\log n}. The term 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 Sn=∑k=1nXkS_n = \sum_{k=1}^{n} X_k, one considers Sn(f)=∑k=1nf(k/n)XkS_n(f) = \sum_{k=1}^{n} f(k/n)X_k, where ff is a function on [0,1][0, 1]. The case f≡1f \equiv1 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 ff, Tomkins proved in [23] that

lim sup⁡n→∞12nlog⁡log⁡n∑k=1nf(k/n)Xk=(∫01f(t)2 dt)1/2a.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.}

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) have a special structure. A general triangular array does not have this property. Let A=(an,k)A=(a_{n,k}), 1≤k≤n1\leq k\leq n, and consider Sn(A)=∑k=1nan,kXkS_n(A)=\sum_{k=1}^{n}a_{n,k}X_k. The coefficients may now change in an arbitrary way from one row to the next. Thus, Sn(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→∞Sn(A)2nlog⁡log⁡n=∞a.s.,\limsup_{n\to\infty}\frac{S_n(A)}{\sqrt{2n\log\log n}}=\infty\quad\text{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 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(X2)<∞\mathbb{E}(X^2)<\infty, and considered a bounded triangular array of weights K:=sup⁡n≥1max⁡1≤k≤n∣an,k∣<∞K:=\sup_{n\geq1}\max_{1\leq k\leq n}|a_{n,k}|<\infty. They showed that the natural scale is 2nlog⁡n\sqrt{2n\log n} and proved

lim sup⁡n→∞∣Sn(A)∣2nlog⁡n≤12K(EX2)1/2a.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)}

They call (1) the law of logarithms. Later, Chen and Hao [CH11] obtained the sharper upper bound KEX2K\sqrt{\mathbb{E}X^2} under the same finite second-moment condition.

Li and Tomkins [17] generalize (1). They replaced the specific 2nlog⁡n\sqrt{2n\log n} normalization by more general bn=B(n)b_n=B(n), where BB is regularly varying with index α∈[1/2,1)\alpha\in[1/2,1), without assuming E(X2)<∞\mathbb{E}(X^2)<\infty. When α>1/2\alpha>1/2, they proved that the normalized weighted sums converge to zero, while at the critical value α=1/2\alpha=1/2 they obtained a non-trivial upper bound:

lim sup⁡n→∞∣Sn(A)∣bn≤62 ℓsup⁡1≤k≤n∣an,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)}

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) with a faithful normal tracial state, and random variables by τ\tau-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) represents a finite von Neumann algebra with an increasing sequence (Mn)(\mathcal{M}_n) of von Neumann subalgebras. For t>et>e, set 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) be a finite von Neumann algebra and (dn)(d_n) be a sequence of self-adjoint martingale differences with τ(d1)=0\tau(d_1)=0, En−1(dn2)=1\mathbb{E}_{n-1}(d_n^2)=1 and max⁡m≤n∥dm∥≤Mn\max_{m\le n}\|d_m\|\le M_n, where (Mn)(M_n) is a sequence of positive reals such that lim⁡n→∞L(n)Mn2n=0\lim_{n\to\infty}\frac{L(n)M_n^2}{n}=0. Let pp be a polynomial on [0,1][0,1], then

lim sup⁡n→∞12nL(n)∑m=1np(m/n)dm≤∥p∥2a.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)}

When p≡1p\equiv1, 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) depends on nn. 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)=tjp(t)=t^j. The corresponding sums are ∑m=1nmjdm\sum_{m=1}^{n}m^jd_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 (xk)k≥1(x_k)_{k\ge1} be a i.i.d sequence of mean zero self-adjoint random variables on (M,τ)(\mathcal{M},\tau). Let A=(an,k)∈DA=(a_{n,k})\in\mathcal{D} with K=sup⁡n,k∣an,k∣K=\sup_{n,k}|a_{n,k}|, and let BB, (bn)(b_n) be as above with index α∈[12,1)\alpha\in[\frac{1}{2},1). Assume that,

(1) τ(B−1(∣x1∣))<∞\tau\left(B^{-1}(|x_1|)\right)<\infty

(2) B−1(tlog⁡t)log⁡tG(t)⟶0\frac{B^{-1}(t\log t)}{\log t}G(t)\longrightarrow0 as t→∞t\to\infty,

(3) if α=12\alpha=\frac{1}{2}, then there exists a constant 0≤ℓ<∞0\le\ell<\infty such that

lim sup⁡n→∞nlog⁡nbn2H(bnlog⁡n)=ℓ2.\limsup_{n\to\infty}\frac{n\log n}{b_n^2}H\left(\frac{b_n}{\log n}\right)=\ell^2.

Then:

(a) If 12<α<1\frac{1}{2}<\alpha<1 and x1∈L2(M,τ)x_1\in L_2(\mathcal{M},\tau),

Sn(A)bn⟶0,almost uniformly,where Sn(A):=∑k=1nan,kxk.\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) If α=12\alpha=\frac{1}{2},

lim sup⁡n→∞Sn(A)bn≤2Kℓb.a.u.\limsup_{n\to\infty}\frac{S_n(A)}{b_n}\le\sqrt{2}K\ell\quad\text{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 2nlog⁡n\sqrt{2n\log n}. The constant 2\sqrt{2} is the key point of this result. Under the corresponding assumptions, our bound improves the constant 626\sqrt{2} in eq. (2) to 2\sqrt{2}. This improvement is also important in the classical setting. In the finite-variance logarithmic case, take B(n)=2nlog⁡nB(n)=\sqrt{2n\log n}. Then the truncated second-moment parameter satisfies ℓ2=12E(X2)\ell^2=\frac{1}{2}\mathbb{E}(X^2). Hence, part (b) of Theorem B reduces to

References

  1. [9]Philip Hartman and Aurel Wintner. On the law of the iterated logarithm. American Journal of Mathematics, 63(1):169–176, 1941.DOI
  2. [10]Marius Junge. Doob’s inequality for non-commutative martingales. Journal für die Reine und Angewandte Mathematik, 549:149–190, 2002.DOI
  3. [11]Marius Junge and Quanhua Xu. Noncommutative maximal ergodic theorems. J. Amer. Math. Soc., 20(2):385–439, 2007.DOI
  4. [12]A Kolmogoroff. Über das gesetz des iterierten logarithmus. Mathematische Annalen, 101(1):126–135, 1929.DOI
  5. [13]Malgorzata Ewa Konwerska. The law of the iterated logarithm in noncommutative probability. University of Illinois at Urbana-Champaign, 2008.
  6. [14]Deli Li and M. Bhaskara Rao. A note on the law of the iterated logarithm for weighted sums of independent identically distributed random variables. In Convergence in ergodic theory and probability (Columbus, OH, 1993), volume 5 of Ohio State Univ. Math. Res. Inst. Publ., pages 273–276. de Gruyter, Berlin, 1996.DOI
  7. [15]De Li Li, M. Bhaskara Rao, and Xiang Chen Wang. On the strong law of large numbers and the law of the logarithm for weighted sums of independent random variables with multidimensional indices. J. Multivariate Anal., 52(2):181–198, 1995.DOI
  8. [16]Deli Li and R. J. Tomkins. Laws of the iterated logarithm for weighted sums of independent random variables. Statist. Probab. Lett., 27(3):247–254, 1996.DOI
  9. [17]D. Li and R. J. Tomkins. The law of the logarithm for weighted sums of independent random variables. J. Theoret. Probab., 16(3):519–542, 2003.DOI
  10. [18]Edward Nelson. Notes on non-commutative integration. Journal of Functional Analysis, 15(2):103–116, 1974.DOI
  11. [19]Sourav Panja, Éric Ricard, and Diptesh Saha. Non-commutative law of iterated logarithm. J. Funct. Anal., 291(1):Paper No. 111477, 21, 2026.arxiv.org/abs/2509.22037
  12. [20]William F Stout. A martingale analogue of Kolmogorov’s law of the iterated logarithm. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete, 15(4):279–290, 1970.DOI
  13. [21]Marianne Terp. Lp spaces associated with von neumann algebras. Notes, Math. Institute, Copenhagen Univ, 3(4):5, 1981.
  14. [22]R. James Tomkins. An iterated logarithm theorem for some weighted averages of independent random variables. Ann. Math. Statist., 42:760–763, 1971.DOI
  15. [23]R. J. Tomkins. A generalization of Kolmogorov’s law of the iterated logarithm. Proc. Amer. Math. Soc., 32:268–274, 1972.DOI
  16. [24]R. J. Tomkins. Iterated logarithm results for weighted averages of martingale difference sequences. Ann. Probability, 3:307–314, 1975.DOI

Paper details

vlnc:W_efz85f94
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents