Jump to ContentJump to Main Navigation
Show Summary Details
More options …

Open Mathematics

formerly Central European Journal of Mathematics

Editor-in-Chief: Vespri, Vincenzo / Marano, Salvatore Angelo

IMPACT FACTOR 2018: 0.726
5-year IMPACT FACTOR: 0.869

CiteScore 2018: 0.90

SCImago Journal Rank (SJR) 2018: 0.323
Source Normalized Impact per Paper (SNIP) 2018: 0.821

Mathematical Citation Quotient (MCQ) 2018: 0.34

ICV 2018: 152.31

Open Access
See all formats and pricing
More options …
Volume 15, Issue 1


Volume 13 (2015)

One sided strong laws for random variables with infinite mean

André Adler
  • Corresponding author
  • Department of Applied Mathematics, Illinois Institute of Technology, Chicago, Illinois, 60616, USA
  • Email
  • Other articles by this author:
  • De Gruyter OnlineGoogle Scholar
Published Online: 2017-06-22 | DOI: https://doi.org/10.1515/math-2017-0070


This paper establishes conditions that secure the almost sure upper and lower bounds for a particular normalized weighted sum of independent nonnegative random variables. These random variables do not possess a finite first moment so these results are not typical. These mild conditions allow us to show that the almost sure upper limit is infinity while the almost sure lower bound is one.

Keywords: Almost sure convergence; Weak law of large numbers; One sided strong laws; Slow variation

MSC 2010: 60F15

This paper extends work done in [1] and [2]. We use those Weak Laws to obtain our almost sure upper and lower bounds. We present conditions that allow us to achieve these bounds for our normalized weighted sums. As in [2] the random variables at the heart of these theorems are independent without a first moment where P{Xj = 0} = 1 − aj and P{Xj>x}=1x+1/aj, for x > 0. We set An=j=1naj, which must go to infinity in order to establish our theorems.

In terms of notation we use lg x = ln x, but whenever we have ln 1 in a denominator we will just set that equal to one, so we won’t be dividing by zero. Hence lg x = ln x, except lg 1 = 1. We set lg1 x = lg x, and lgk+1 x = lg(lgk x), so the i in lgi x is not the base, it is the iteration of the logarithm. Many times in the proofs and the examples we will pull the logarithm and any other slowly varying function out from an integral and the infinite summations since they are slowly varying, see [3] pages 279-284. Finally, note that the constant C will be used as a generic bound that is not necessarily the same in each appearance. From [2] we have a Weak Law that helps us produce the almost sure lower bound.

Theorem 1

If An → ∞, then j=1najXjAnlgAnP1.(1)

We next establish the almost sure upper bound for these same weighted sums. The condition that ensures that the upper bound is infinite is n=1anAnlgAn=.(2)

Theorem 2

If (2) holds, then lim supnj=1najXjAnlgAn=almostsurely.


For all M > 0 n=1P{anXnAnlgAn>M}=n=1P{Xn>MAnlgAnan}=n=11MAnlgAnan+1an=n=1anMAnlgAn+1=.

Thus lim supnanXnAnlgAn=almostsurely hence lim supnj=inajXjAnlgAnlim supnanXnAnlgAn=almostsurely which completes this proof. ☐

In order to get the lower limit we assume there is a sequence {dn, n ≥ 1} where limnj=1najlg(ajdj)AnlgAn=1(3) and n=1an2dnAn2lg2An<.(4)

Note that we need both An and an dn to go to infinity to achieve our results. But these are mild conditions.

What is fascinating is that even though we can split (3) into two pieces, we need both parts. When aj = jα, where α > −1, we select dj = j which gives us j=1najlgajAnlgAnαα+1 and j=1najlgdjAnlgAn1α+1.

This shows that both terms are necessary. Combining that with An ~ nα+1/(α + 1), we see that (3) and (4) hold, see Example 4.

Theorem 3

If there exist a sequence {dn, n ≥ 1} satisfying (3) and (4) then lim infnj=1najXjAnlgAn=1almostsurely.


From Theorem 1 we have lim infnj=1najXjAnlgAn1almostsurely.

We will show that conditions (3) and (4) establishes lim infnj=1najXjAnlgAn1almostsurely.(5)

Using our sequence {dn, n ≥ 1} we have j=1najXjAnlgAnj=1najXjI(Xjdj)AnlgAn=j=1naj[XjI(Xjdj)EXjI(Xjdj)]AnlgAn+j=1najEXjI(Xjdj)AnlgAn.

By the Khintchine-Kolmogorov Convergence Theorem, see [4] page 113, and the Kronecker lemma, the first term converges to zero almost surely since n=1an2EXn2I(Xndn)An2lg2An=n=1an2An2lg2An0dnx2dx(x+1/an)2n=1an2dnAn2lg2An< by (4). We now examine our expectation j=1najEXjI(Xjdj)AnlgAn=1AnlgAnj=1naj0djxdx(x+1/aj)2=1AnlgAnj=1naj1/ajdj+1/aj(u1/aj)duu2=1AnlgAnj=1naj1/ajdj+1/ajduu1AnlgAnj=1n1/ajdj+1/ajduu2.

The first term converges to one by (3) since 1AnlgAnj=1naj1/ajdj+1/ajduu=1AnlgAnj=1naj[lg(dj+1/aj)lg(1/aj)]=1AnlgAnj=1najlg(ajdj+1)1 while the second term goes to zero since 1AnlgAnj=1n1/ajdj+1/ajduu21AnlgAnj=1n1/ajduu2=1AnlgAnj=1naj=AnAnlgAn=1lgAn0.

Hence (5) does hold, concluding the proof. ☐

We conclude with three examples, each showing how to select {dn, n ≥ 1} so that (3) and (4) hold. What is quite remarkable is that if we select {dn, n ≥ 1} correctly we see that all three conditions hold even though (2) and (4) seem contradictory.

Example 4

Let L(x) be any slowly varying function and α > −1, thenlim infnj=1njαL(j)Xjnα+1L(n)lgn=1almostsurelyandlim supnj=1njαL(j)Xjnα+1L(n)lgn=almostsurely.


Since aj = jαL(j) we have An ~ nα+1L(n)/(α + 1) and lg An ~ (α + 1) and lg n, thus n=1anAnlgAnCn=1nαL(n)nα+1L(n)lgn=Cn=11nlgn=.

In this case we set dj = j. Thus ajdj = jα+1L(j) and lg(ajdj) ~ (α + 1) lg j, hence j=1najlg(ajdj)AnlgAn(α+1)j=1njαL(j)lgjnα+1L(n)lgn1 and n=1an2dnAn2lg2AnCn=1n2α+1L2(n)n2α+2L2(n)lg2n=Cn=11nlg2n<.

Next we look at a borderline case, where once again An → ∞ is necessary.

Example 5

If α > −1, then for all δliminfnj=1n(lgj)α(lg2j)δjXj(lgn)α+1(lg2n)δ+1=1almostsurelyandlimsupnj=1n(lgj)α(lg2j)δjXj(lgn)α+1(lg2n)δ+1=almostsurely.


Since aj = (lgj)α(lg2 j)δ/j we have An ~ (lgn)α+1(lg2n)δ /(α + 1) and lg An ~ (α + 1) lg2 n, thus n=1anAnlgAnCn=1(lgn)α(lg2n)δ/n(lgn)α+1(lg2n)δ+1=Cn=11nlgnlg2n=.

In this case we set dj = j lg j. Thus ajdj = (lg j)α+1(lg2 j)δ and lg(aj dj) ~ (α + 1) lg2 j, hence j=1najlg(ajdj)AnlgAn(α+1)j=1n(lgj)α(lg2j)δ+1/j(lgn)α+1(lg2n)δ+11 and n=1an2dnAn2lg2AnCn=1((lgn)α(lg2n)δn)2(nlgn)((lgn)α+1(lg2n)δ+1)2=Cn=11nlgnlg22n<.

Example 6

If α > −1, then for all δliminfnj=1n(lg2j)α(lg3j)δjlgjXj(lg2n)α+1(lg3n)δ+1=1almostsurelyandlimsupnj=1n(lg2j)α(lg3j)δjlgjXj(lg2n)α+1(lg3n)δ+1=almostsurely.


Since aj = (lg2 j)α(lg3 j)δ/(j lg j) we have An ~ (lg2n)α+1(lg3 n)δ/(α + 1) and lg An ~ (α + 1) lg3 n, thus n=1anAnlgAnCn=1(lg2n)α(lg3n)δ/(nlgn)(lg2n)α+1(lg3n)δ+1=Cn=11nlgnlg2nlg3n=.

In this case we set dj = j lgj lg2 j. Thus ajdj = (lg2 j)α+1(lg3 j)δ and lg(ajdj) ~ (α + 1) lg3 j, hence j=1najlg(ajdj)AnlgAn(α+1)j=1n(lg2j)α(lg3j)δ+1/(jlgj)(lg2n)α+1(lg3n)δ+11 and n=1an2dnAn2lg2AnCn=1((lg2n)α(lg3n)δnlgn)2(nlgnlg2n)((lg2n)α+1(lg3n)δ+1)2=Cn=11nlgnlg2nlg32n<.

In each case we see that even though (2) and (4) are quite similar series and even though one must be finite and the other infinite, if one selects {dn, n ≥ 1} correctly, these two equations along with (3) hold every time. We can continue with an=(lgd+1n)α(lgd+2n)δni=1dlgin where α > −1, as long as we set dn=ni=1d+1lgin.

Remark 7

It would be nice to replace (2) with the milder condition An → ∞, but they arent equivalent. If we have rapidly growing sequences we can see the difference. For example, let An = en2, then clearly An → ∞. Howeveran=AnAn1=en2e(n1)2<en2andn=1anAnlgAn<n=1en2en2lg(en2)=n=11n2<.


  • [1]

    Adler A., An exact weak law of large numbers, Bull. Inst. Math. Acad. Sinica, 2012, 7, 417-422 Google Scholar

  • [2]

    Nakata T., Weak law of large numbers for weighted independent random variables with infinite mean. Statist. Probab. Letters, 2016, 109, 124-129 CrossrefGoogle Scholar

  • [3]

    Feller W., An Introduction to Probability Theory and its Applications, Vol 2, 2nd edition, 1996, Wiley and Sons Google Scholar

  • [4]

    Chow Y.S., Teicher H., Probability Theory: Independence, Interchangeability, Martingales, 3rd edition, 1997, Springer-Verlag Google Scholar

About the article

Received: 2016-12-12

Accepted: 2017-05-16

Published Online: 2017-06-22

Citation Information: Open Mathematics, Volume 15, Issue 1, Pages 828–832, ISSN (Online) 2391-5455, DOI: https://doi.org/10.1515/math-2017-0070.

Export Citation

© 2017 Adler. This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 License. BY-NC-ND 3.0

Citing Articles

Here you can find all Crossref-listed publications in which this article is cited. If you would like to receive automatic email messages as soon as this article is cited in other publications, simply activate the “Citation Alert” on the top of this page.

Wenzhi Yang, Lei Yang, Da Wei, and Shuhe Hu
Communications in Statistics - Theory and Methods, 2019, Volume 48, Number 12, Page 3044

Comments (0)

Please log in or register to comment.
Log in