Professional Documents
Culture Documents
A Supplement To The Laws of Large Numbers and The Large Deviations
A Supplement To The Laws of Large Numbers and The Large Deviations
To cite this article: Deli Li & Yu Miao (2021) A supplement to the laws of large numbers and the
large deviations, Stochastics, 93:8, 1261-1280, DOI: 10.1080/17442508.2021.1903465
Article views: 28
1. Introduction
Throughout this paper, let (B, · ) be a real separable Banach space equipped with its
Borel σ -algebra B ( = the σ -algebra generated by the class of open subsets of B determined
by · ). Let {X, Xn ; n ≥ 1} be a sequence of independent and identically distributed (i.i.d.)
B-valued
n random variables defined on a probability space (, F , P). As usual, let Sn =
k=1 k n ≥ 1 denote their partial sums.
X ,
Many questions in probability theory can be formulated as a law of large numbers. We
begin the statements of the classical laws of large numbers. If 0 < p < 2 and if {X, Xn ; n ≥ 1}
is a sequence of i.i.d. real-valued random variables (that is, if B = R), then
Sn
lim = 0 almost surely (a.s.) if and only if
n→∞ n1/p
and
⎧
⎪
⎪ lim nP |X| > n1/p = 0 if 0 < p < 1,
Sn ⎨n→∞
→P 0 if and only if lim (EXI(|X| ≤ n) + nP (|X| > n)) = 0 if p = 1,
n1/p ⎪
⎩EX = 0 and lim nP |X| > n1/p = 0
n→∞
⎪
if 1 < p < 2,
n→∞
(2)
where ‘→P ’ stands for convergence in probability. Statement (1) is the famous Kol-
mogorov–Marcinkiewicz–Zygmund strong law of large numbers (SLLN) (see [12] for
p = 1 and [17] for p = 1) and statement (2) is the celebrated Kolmogorov–Feller weak
law of large numbers (WLLN) (see the books by Gnedenko and Kolmogorov [9] and Feller
[7]).
The classical Kolmogorov SLLN in real separable Banach spaces was established by
Mourier [20]. The extension of the Kolmogorov–Marcinkiewicz–Zygmund SLLN to B-
valued random variables is independently due to Azlarov and Volodin [1] and de Acosta
[5, Theorem 3.1] who showed that, if 0 < p < 2 and if {X, Xn ; n ≥ 1} is a sequence of i.i.d.
B-valued random variables, then
Sn Sn
lim 1/p
= 0 a.s. if and only if EXp < ∞ and 1/p →P 0.
n→∞ n n
In addition, de Acosta [5, Theorem 4.1] provided a remarkable characterization of
Rademacher-type p (1 ≤ p < 2) Banach spaces via the SLLN by showing that the following
two statements are equivalent:
Obviously, the SLLN and WLLN results say little or nothing about the rate of convergence,
however, which is often important for many applications of probability theory.
Let {X, Xn ; n ≥ 1} be a sequence of i.i.d. real-valued random variables. Cramér [4] and
Chernoff [3] showed that if
M(t) ≡ E etX < ∞ ∀ t ∈ R, (3)
STOCHASTICS 1263
then
1
lim sup log P (Sn /n ∈ A) ≤ −(A),
n→∞ n
1
lim inf log P (Sn /n ∈ A) ≥ −(A),
n→∞ n
where for x ∈ R and A ⊆ R, λ(x) = supt∈R (tx − log M(t)), (A) = inf x∈A λ(x). This
fundamental result is what we call the large deviation principle (LDP) for partial sums
{Sn , n ≥ 1}. Clearly, under condition (3), the LDP characterizes the exponential concen-
tration behaviour, as n → ∞, of a sequence of probabilities {P(Sn /n ∈ A); n ≥ 1} in terms
of a rate function (A), A ⊆ R. Donsker and Varadhan [6] and Bahadur and Zabell [2]
established an LDP for sums of i.i.d. B-valued random variables. The large deviations the-
ory has applications in many different scientific fields, ranging from queuing theory to
statistics and from finance to engineering.
Note that (3) is the moment generating function of X which is a very strong moment
condition of X. Thus, one may ask, naturally, whether there is some LDP analogue under
Sn /n1/p →P 0 (i.e. WLLN condition) only. This paper will provide a positive answer (see
Theorem 2.1 in Section 2).
The second motivation of this paper comes from the following logarithmic asymp-
totic behaviours obtained by Gantert [8] for the partial sums of i.i.d. non-negative random
variables.
Theorem 1.1 (Gantert [8, Theorem 2]): Let {X, Xn ; n ≥ 1} be a sequence of non-negative
i.i.d. random variables and set Sn = ni=1 Xi , n ≥ 1. Let α > 1. Then the following holds.
(a) If EX α = ∞ and P(X > t) ≥ L(t)/t α for some slowly varying function L, then for every
m > 0 and x ≥ 1,
1
lim inf log P Sn ≥ nx m ≥ −αx + 1.
n→∞ log n
(b) If EX p < ∞ for each p < α, then for every m > EX and x ≥ 1,
1
lim sup log P Sn ≥ nx m ≤ −αx + 1.
n→∞ log n
To further investigate the logarithmic asymptotic behaviours for the partial sums Sn ,
Hu and Nyrhinen [11] introduced the following two parameters for non-negative random
1264 D. LI AND Y. MIAO
variable X, namely
1
ᾱ = − lim sup log P(log X > t) ∈ [0, ∞]
t→∞ t
and
1
α = − lim inf log P(log X > t) ∈ [0, ∞].
t→∞ t
ᾱ = sup λ ≥ 0; EX λ < ∞ .
The proof can be found in Rolski et al. [23, p. 39]. Hence, if ᾱ < ∞, then X is heavy tailed,
namely, EeλX = ∞ for every λ > 0.
Write x̄ = max{1, 1/ᾱ} if ᾱ > 0 and x = max{1, 1/α}. Hu and Nyrhinen [11] estab-
lished the following large deviations for the partial sums of non-negative i.i.d. random
variables.
Theorem 1.2 ([11, Theorems 2.1 and 2.2]): Let {X, Xn ; n ≥ 1} be a sequence of non-
negative i.i.d. random variables. We have:
(i) Assume that ᾱ ∈ (0, ∞). Then for every x > x̄,
1
lim sup log P Sn > nx = −ᾱx + 1.
n→∞ log n
Motivated by the above-mentioned results, Miao et al. [19] established the logarithmic
asymptotic behaviours for the cases of m-stationary sequences and stationary negatively
associated sequences. Recently, Wang [24] not only improved the logarithmic asymptotic
results obtained by Miao et al. [19] for the partial sums of stationary sequences, but also
generalized the results to classes of acceptable random variables and widely acceptable
random variables.
Inspired by the above discovery by Gantert [8] and Hu and Nyrhinen [11], in the current
work a supplement to the classical laws of large numbers and the classical large deviations
STOCHASTICS 1265
will be provided. We will establish the large deviations for the tail probabilities P(Sn >
sn1/p ) for all s > 0 by giving the exact values for
1 1
lim sup log P Sn > sn1/p and lim inf log P Sn > sn1/p ,
n→∞ log n n→∞ log n
where {X, Xn ; n ≥ 1} is a sequence of i.i.d. B-valued random variables and 0 < p < 2.
The plan of the paper is as follows. The main results are stated in Section 2. Our main
results Theorems 2.1–2.3 are stated in Section 2. For the case 1 ≤ p < 2, our results are
new even in the case where the Banach space is the real line. Some preliminary lemmas
are presented in Section 3 and the proofs of our main results are given in Section 4. The
main tools employed in proving the main results are the remarkable Hoffmann–Jørgensen
[10] inequality, de Acosta [5] inequality, the symmetrization technique, and an inequality
of Ledoux and Talagrand [13].
and
1
β = − lim inf log P(log X > t) ∈ [0, ∞]. (5)
t→∞ t
For any real numbers x ∈ (0, ∞) and y ∈ (−∞, ∞), by convention, (±∞ + y)/x = ±∞.
We now state our main results.
Theorem 2.1: Let 0 < p < 2. Let {X, Xn ; n ≥ 1} be a sequence of i.i.d. B-valued random
variables and set Sn = ni=1 Xi , n ≥ 1. If
Sn
→P 0, (6)
n1/p
then, for all s > 0,
1
lim sup log P Sn > sn1/p = − β̄ − p /p (7)
n→∞ log n
and
1
lim inf log P Sn > sn1/p = − β − p /p. (8)
n→∞ log n
Hence, under condition (6), for all s > 0,
1
lim log P Sn > sn1/p = − β̂ − p /p if and only if β̄ = β = β̂. (9)
n→∞ log n
1266 D. LI AND Y. MIAO
Remark 2.1: Let 0 < p < 2 and {bn ; n ≥ 1} be an increasing sequence of positive real
numbers such that
according as
where γn = nE(XI{X ≤ bn }), n ≥ 1. For the special case bn = n1/p , it can be easily
deduced from the equivalence between (10) and (11) that, if B is of stable type p, then
for every sequence {X, Xn ; n ≥ 1} of i.i.d. B-valued random variables, (6) holds if and only
if
⎧
⎪
⎪ lim nP X > n1/p = 0 if 0 < p < 1,
⎪
⎨n→∞
EX = 0 and lim nP X > n1/p = 0 if 1 < p < 2,
⎪
⎪ n→∞
⎪
⎩ lim (EXI(X ≤ n) + nP (X > n)) = 0 if p = 1.
n→∞
Every real separable Banach space B is of stable type p for all p ∈ (0, 1). It is well
known that all real separable Hilbert spaces, Lp spaces (p ≥ 2), and real separable finite-
dimensional Banach spaces are of stable type p for all 0 < p ≤ 2. It now follows from
Theorem 2.1 and Remark 2.1 that the following theorem provides large deviations for sums
of heavy-tailed i.i.d. real-valued random variables.
Theorem 2.2: Let 0 < p < 2. Let {X, Xn ; n ≥ 1} be a sequence of i.i.d. real-valued random
variables such that
⎧
⎪
⎪ lim nP |X| > n1/p = 0 if 0 < p < 1,
⎪
⎨n→∞
lim (EXI(|X| ≤ n) + nP (|X| > n)) = 0 if p = 1, (12)
⎪
⎪
n→∞
⎪
⎩EX = 0 and lim nP |X| > n 1/p
=0 if 1 < p < 2.
n→∞
and
1
lim inf log P |Sn | > sn1/p = − β − p /p,
n→∞ log n
STOCHASTICS 1267
where β̄ and β are defined by (4) and (5), respectively, when X is replaced by |X|. Hence,
under condition (12), for all s > 0,
1
lim log P |Sn | > sn1/p = − β̂ − p /p if and only if β̄ = β = β̂. (13)
n→∞ log n
Using Theorem 2.2, we can establish large deviations for sums of heavy-tailed i.i.d. non-
negative random variables.
and
1
lim inf log P Sn > sn1/p = − α − p /p.
n→∞ log n
Hence, under condition (14), for all s > 0,
1
lim log P Sn > sn1/p = − α̂ − p /p if and only if ᾱ = α = α̂.
n→∞ log n
(ii) If 1 ≤ p < 2,
EX = μ ∈ (0, ∞) and lim nP X > n1/p = 0, (15)
n→∞
and
1
lim inf log P Sn > nμ + sn1/p = − α − p /p. (17)
n→∞ log n
If ᾱ > 0, then for every x > max{1, 1/ᾱ}, we have 0 < 1/x < min{1, ᾱ} and hence (14)
holds with p = 1/x. Hence, by Theorem 2.3 (i), Theorem 1.2 (i) and Theorem 1.2 (iii)
follow if 0 < ᾱ < ∞ and ᾱ = ∞, respectively.
Obviously, the case where ᾱ ∈ (0, ∞) and α = ∞ is not covered by Theorem 1.2.
However, for this case, by Theorem 2.3 (i), for all x > max{1, 1/ᾱ},
1 1
lim sup log P Sn > snx = lim sup log P Sn > sn1/p
n→∞ log n n→∞ log n
= − ᾱ − p /p = −ᾱx + 1
and
1 1
lim inf log P Sn > snx = lim inf log P Sn > sn1/p = −∞.
n→∞ log n n→∞ log n
The assertion of Theorem 1.2 (ii) is simple. In fact, since {X, Xn ; n ≥ 1} is a sequence of
non-negative i.i.d. random variables, for all s > 0,
n
P Sn > sn ≥ P max Xk > sn = 1 − 1 − P X > snx .
x x
1≤k≤n
which ensures
n
lim inf 1 − P X > snx = 0,
n→∞
and hence,
lim sup P Sn > snx = 1,
n→∞
i.e. Theorem 1.2 (ii) follows.
3. Preliminary lemmas
To prove our main results, we use the following preliminary lemmas.
1
β = − lim inf log P(log X > t) = sup r ≥ 0; lim inf t r P(X > t) = 0 . (19)
t→∞ t t→∞
Proof: We will only give the proof of (18) since a proof of (19) can be culled from the proof
of (18) with obvious modifications.
Replacing t by log t, we have
1 1
β̄ = − lim sup log P(log X > t) = − lim sup log P(X > t).
t→∞ t t→∞ log t
and there exists a sequence {tn, ; n ≥ 1} of increasing positive numbers such that
−β̄−
lim tn, = ∞ and P X > tn, ≥ tn, , n ≥ 1. (21)
n→∞
which ensures
sup r ≥ 0; lim t P(X > t) = 0 ≥ M.
r
t→∞
Remark 3.1: Let X be a B-valued random variable. Let β̄ and β be defined in Lemma 3.1.
By Lemma 3.1, it is easy to see that, for all s > 0,
β̄ = sup r ≥ 0; lim t r P(X > t) = 0 = sup r ≥ 0; lim t r P(X > st) = 0
t→∞ t→∞
and
β = sup r ≥ 0; lim inf t r P(X > t) = 0
t→∞
= sup r ≥ 0; lim inf t r P(X > st) = 0 ,
t→∞
and hence,
β̄ = sup r ≥ 0; lim t P X > st 1/r
=0
t→∞
1270 D. LI AND Y. MIAO
= sup r ≥ 0; lim nP X > sn1/r = 0
n→∞
and
β = sup r ≥ 0; lim inf t P X > st 1/r
=0
t→∞
= sup r ≥ 0; lim inf nP X > sn1/r = 0 .
n→∞
and
lim inf t r P(X > t) = 0 if and only if lim inf t r P X − X
> t = 0.
t→∞ t→∞
and
1 1
β = − lim inf log P(log X > t) = − lim inf log P(log X − X
> t).
t→∞ t t→∞ t
The following lemma, which provides a symmetrization procedure for the large devia-
tion for B-valued random variables, may be of independent interest.
Lemma 3.2: Let {Yn ; n ≥ 1} be a sequence of B-valued random variables such that
Yn →P 0. (22)
Let {Yn
; n ≥ 1} be an independent copy of {Yn ; n ≥ 1}. Write Ŷn = Yn − Yn
, n ≥ 1. Let
{an ; n ≥ 1} be a sequence of positive numbers such that limn→∞ an = ∞. If there exist two
constants 0 ≤ γ̄ ≤ γ ≤ ∞ such that, for all s > 0,
1 1
lim sup log P Ŷn > s = −γ̄ and lim inf log P Ŷn > s = −γ , (23)
n→∞ an n→∞ an
then, for all s > 0,
1 1
lim sup log P (Yn > s) = −γ̄ and lim inf log P (Yn > s) = −γ . (24)
n→∞ an n→∞ an
STOCHASTICS 1271
Proof: For any given s > 0, it follows from (22) that there exists a positive integer n0 (which
depends on s only) such that, for all n ≥ n0 ,
1
P Ŷn > 2s ≤ P (Yn > s) ≤ 2P Ŷn > s/2 ,
2
and hence,
1 1 1 1
log P Ŷn > 2s ≤ log P (Yn > s) ≤ log 2P Ŷn > s/2 . (25)
an 2 an an
The first part of the following lemma is one of Lévy’s inequalities (see, e.g.[13, p.47]),
the second part is a version of the Hoffmann–Jørgensen inequality due to Li et al. [15,
Lemma 2.2], and the third part is the second part of Proposition 6.8 of [13, p.156].
(ii) for each integer j ≥ 1, there exist positive numbers Cj and Dj , depending only on j, such
that
P Tn > 2jt ≤ Cj P max Vk > t + Dj (P(Tn > t))j ;
1≤k≤n
Lemma 3.4: Let {X, Xn ; n ≥ 1} be a sequence of i.i.d. B-valued random variables. Let p > 0.
Then, for any s > 0,
1
lim sup log P max Xk > sn 1/p
= 0 ∧ −(β̄ − p)/p (26)
n→∞ log n 1≤k≤n
and
1
lim inf log P max Xk > sn 1/p
= 0 ∧ −(β − p)/p . (27)
n→∞ log n 1≤k≤n
and hence,
1
lim sup log P max Xk > sn1/p
n→∞ log n 1≤k≤n
1
= lim sup log 1 ∧ nP X > sn1/p
n→∞ log n
1
= 0 ∧ lim sup log nP X > sn 1/p
n→∞ log n
1
= 0 ∧ 1 + lim sup log P X > sn 1/p
n→∞ log n
STOCHASTICS 1273
log sn1/p 1
= 0 ∧ 1 + lim sup × log P X > sn1/p
n→∞ log n log sn1/p
1 1
= 0 ∧ 1 + lim sup log P(X > t)
p t→∞ log t
1
=0∧ 1+ −β̄
p
= 0 ∧ −(β̄ − p)/p ,
proving (26). Similarly (27) follows when lim supn→∞ is replaced by lim inf n→∞ .
Lemma 3.5: Let 0 < p < 2. Let {X, Xn ; n ≥ 1} be a sequence of i.i.d. B-valued random
variables such that (6) holds. Then
Ŝn
→P 0,
n1/p
n
where Ŝn = k=1 X̂k , n≥ 1. Thus, applying Lemma 3.3 (i), we have
P max X̂k > n1/p
≤ 2P Ŝn > n1/p → 0 as n → ∞,
1≤k≤n
and hence,
n
P max X̂k > n1/p
= 1 − 1 − P X̂ > n1/p → 0 as n → ∞
1≤k≤n
which is equivalent to
nP X̂ > n1/p → 0 as n → ∞.
Lemma 3.6: For any separable Banach space B and any finite sequence {Vk ; 1 ≤ k ≤ n} of
independent B-valued random variables with EVk 2 < ∞, 1 ≤ k ≤ n, we have
n
E |Tn − ETn | ≤ 4
2
EVk 2 ,
k=1
where Tn = V1 + · · · + Vn .
1274 D. LI AND Y. MIAO
Let m(Y) denote a median for a real-valued random variable Y. The following lemma is
used to prove Theorem 2.3. The first part was obtained by Li and Rosalsky [14, Lemma 3.1]
and the second part was established by Petrov [21] (also see [22, Theorem 2.1]).
Proof of Theorem 2.1: Applying Remark 3.2 and Lemma 3.2, without loss of generality, we
can assume that {X, Xn ; n ≥ 1} is a sequence of i.i.d. symmetric B-valued random variables.
Thus, by Lemma 3.3 (i), for all s > 0,
P max Xk > sn 1/p
≤ 2P Sn > sn1/p , n ≥ 1.
1≤k≤n
and
1
lim inf log P Sn > sn1/p
n→∞ log n
1
≥ lim inf log P max Xk > sn1/p = 0 ∧ −(β − p)/p .
n→∞ log n 1≤k≤n
Since for each n ≥ 1, Xk , 1 ≤ k ≤ n are symmetric B-valued random variables, for any
c ≥ 0 we have
d
Xk I (|Xk | ≤ c) − Xk I (|Xk | > c) = Xk , 1 ≤ k ≤ n,
d
where ‘=’ stands for ‘equal in distribution’. Thus, for each n ≥ 1,
d
Xk I |Xk | ≤ n1/q − Xk I |Xk | > n1/q ; 1 ≤ k ≤ n = {Xk ; 1 ≤ k ≤ n} . (34)
Note that, for each n ≥ 1,
n
k=1 Xk I Xk ≤ n
1/q − X I X > n1/q
1 k k Sn
Tn = + 1/p .
2 n1/p n
Thus, it follows from (34) and (6) that, for any > 0,
n
1/q − X I X > n1/q
k=1 Xk I Xk ≤ n k k
P (Tn > ) ≤ P >
n1/p
Sn
+P >
n1/p
Sn
= 2P 1/p >
n
→0 as n → ∞
and hence,
tn = inf {t > 0; P (Tn > t) ≤ 1/24} → 0 as n → ∞.
By Lemma 3.3 (iii) with r = 1, we have
X I X ≤ n1/q
k k
E Tn ≤ 6E max + 6tn
1≤k≤n n1/p
≤ 6 n−(q−p)/(pq) + tn
→ 0 as n → ∞. (35)
p (2−p)(q−p) 2 2−p+δ
Let δ = 2 ∧ 2p and τ = p −1− q . Then δ > 0 and
2q − pq − 2p + p2 − δp (q − p)(2 − p) − δp (q − p)(2 − p)
τ= = ≥ > 0.
pq pq 2pq
By Lemma 3.5, (6) implies that
EXp−δ < ∞.
For given s > 0, it follows from (35), Markov’s inequality, and Lemma 3.6 that, for all
sufficiently large n,
s
P (Tn > s) ≤ P |Tn − ETn | >
2
1276 D. LI AND Y. MIAO
s
For given M > 0, let j = [M/τ ] + 1 and t = 2j . Then, it follows from Lemma 3.3 (ii)
and (36) that, for all sufficiently large n,
P (Tn > s) = P Tn > 2jt
≤ Cj P max Vn,k > tn1/p
+ Dj (P(Tn > t))j
1≤k≤n
−
q−p
≤ Cj P n pq > t + O n−jτ
= O n−M ,
and hence,
1
lim sup log P (Tn > s) ≤ −M.
n→∞ log n
Letting M → ∞, we have, for all s > 0,
1
lim log P (Tn > s) = −∞. (37)
n→∞ log n
we have
log P Sn > sn1/p ≤ log 2 + log P (Tn > s) ∨ log nP X > n1/q , n ≥ 1.
Note that
1
lim sup log nP X > n1/q
n→∞ log n
1
= 1 + lim sup log P X > t1/q
t→∞ log t
log t 1/q 1
= 1 + lim sup log P X > t 1/q
t→∞ log t log t 1/q
= − β̄ − q /q,
and similarly,
1
lim inf log nP X > n1/q = − β − q /q.
n→∞ log n
Thus, it follows from (37) that, for all s > 0,
1
lim sup log P Sn > sn1/p
n→∞ log n
1
≤ lim sup log 2 + log P (Tn > s) ∨ log nP X > n1/q
n→∞ log n
1 1
≤ lim log P (Tn > s) ∨ lim sup log nP X > n 1/q
n→∞ log n n→∞ log n
= (−∞) ∨ (β̄ − q)/q
= −(β̄ − q)/q,
and similarly,
1
lim inf log P Sn > sn1/p ≤ −(β − q)/q.
n→∞ log n
Letting q p, we have, for all s > 0,
1
lim sup log P Sn > sn1/p ≤ −(β̄ − p)/p (38)
n→∞ log n
and
1
lim inf log P Sn > sn1/p ≤ −(β − p)/p. (39)
n→∞ log n
Thus, (7) follows from (32) and (38) and (8) follows from (33) and (39). The proof of
Theorem 2.1 is complete.
Proof of Theorem 2.2: Since {X, Xn ; n ≥ 1} is a sequence of i.i.d. real-valued random vari-
ables, we have for 0 < p < 2, (6) and (12) are equivalent and refer to (2), Theorem 2.2 follows
from Theorem 2.1 immediately.
1278 D. LI AND Y. MIAO
Proof of Theorem 2.3: For 0 < p < 1, since {X, Xn ; n ≥ 1} is a sequence of non-negative
i.i.d. random variables, we have, for all s > 0,
P Sn > sn1/p = P |Sn | > sn1/p , n ≥ 1,
and hence, the conclusions of Theorem 2.3 (i) follow from (14) and Theorem 2.2.
For 1 ≤ p < 2, since X is a non-negative random variable, under condition (15), we
have
Sn,μ
→P 0, (40)
n1/p
where Xμ = X − μ, Xn,μ = Xn − μ, and Sn,μ = X1,μ + · · · + Xn,μ , n ≥ 1. It is easy to see
that
1 1
lim sup log P(log |X − μ| > t) = lim sup log P(log X > t) = −ᾱ
t→∞ t t→∞ t
and
1 1
lim inf log P(log |X − μ| > t) = lim inf log P(log X > t) = −α.
t→∞ t t→∞ t
Thus, either refer to (2) and Theorem 2.2 or directly to Theorem 2.1, for all s > 0,
1
lim sup log P Sn > nμ + sn1/p
n→∞ log n
1
≤ lim sup log P Sn,μ > sn1/p = − ᾱ − p /p (41)
n→∞ log n
and
1
lim inf log P Sn > nμ + sn1/p
n→∞ log n
1
≤ lim inf log P Sn,μ > sn1/p = − α − p /p. (42)
n→∞ log n
Recall that m(Y) is a median for a real-valued random variable Y. Write S0,μ = 0 and
mn = min m Sk,μ , n ≥ 1,
0≤k≤n
≤P max Xk,μ > sn1/p − 2mn
1≤k≤n
=P max Xk,μ + min m Sk,μ > sn1/p − mn
1≤k≤n 0≤k≤n
≤P max Xk,μ + min m Sk−1,μ > sn 1/p
− mn
1≤k≤n 1≤k≤n
≤P max Xk,μ + m Sk−1,μ > sn 1/p
− mn
1≤k≤n
≤ 2P max Sk,μ > sn 1/p
− mn
1≤k≤n
≤ 2P max Sk,μ + m Sn,μ − Sk,μ > sn1/p
1≤k≤n
≤ 4P Sn,μ > sn1/p . (43)
Recall that X is a non-negative random variable; by Lemma 3.1 (replace X with X), it
follows from (15) that p ≤ ᾱ ≤ α. Thus, by Lemma 3.4, (43) ensures that, for all s > 0,
1 1
lim sup log P Sn > nμ + sn1/p = lim sup log P Sn,μ > sn1/p
n→∞ log n n→∞ log n
1
≥ lim sup log P max Xk > 3sn 1/p
n→∞ log n 1≤k≤n
= 0 ∧ − ᾱ − p /p
= − ᾱ − p /p, (44)
and similarly,
1
lim inf log P Sn > nμ + sn1/p ≥ − α − p /p. (45)
n→∞ log n
Thus, (16) follows from (41) and (44), and (17) follows from (42) and (45). This completes
the proof of Theorem 2.3.
Acknowledgments
The authors are extremely grateful to the referee for very carefully reading the manuscript and for
offering numerous comments and suggestions which enabled them to improve the presentation of
the paper.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Funding
The research of Deli Li was partially supported by a grant from the Natural Sciences and Engineer-
ing Research Council of Canada [grant no. RGPIN-2019-06065], and the research of Yu Miao was
1280 D. LI AND Y. MIAO
partially supported by a grant from the National Natural Science Foundation of China [grant no.
NSFC-11971154].
References
[1] T.A. Azlarov and N.A. Volodin, Laws of large numbers for identically distributed Banach-space
valued random variables, Teor. Veroyatnost.Primen. 26 (1981), pp. 584–590, in Russian. English
translation in Theory Probab. Appl. 26 (1981), pp. 573–580.
[2] R.R. Bahadur and S.L. Zabell, Large deviations of the sample mean in general vector spaces, Ann.
Probab. 7 (1979), pp. 587–621.
[3] H. Chernoff, A measure of asymptotic efficiency for tests of a hypothesis based on the sum of
observations, Ann. Math. Stat. 23 (1952), pp. 493–507.
[4] H. Cramér, Sur un nouveau théorème-limite de la thorie des probabilités, Actual. Sci. Ind. 736
(1938), pp. 5–23.
[5] A. de Acosta, Inequalities for B-valued random vectors with applications to the law of large
numbers, Ann. Probab. 9 (1981), pp. 157–161.
[6] M.D. Donsker and S.R.S. Varadhan, Asymptotic evaluation of certain Markov process expecta-
tions for large time, III, Comm. Pure Appl. Math. 29 (1976), pp. 389–461.
[7] W. Feller, An Introduction to Probability Theory and Its Applications, 2nd ed., Vol. 2, Wiley, New
York, 1971.
[8] N. Gantert, A note on logarithmic tail asymptotics and mixing, Statist. Probab. Lett. 49 (2000),
pp. 113–118.
[9] B.V. Gnedenko and A.N. Kolmogorov, Limit Distributions for Sums of Independent Random
Variables, 2nd ed., Addison-Wesley, Reading, MA, 1968.
[10] J. Hoffmann-Jørgensen, Sums of independent Banach space valued random variables, Studia
Math.52 (1974), pp. 159–186.
[11] Y.J. Hu and H. Nyrhinen, Large deviations view points for heavy-tailed random walks, J. Theoret.
Probab. 17 (2004), pp. 761–768.
[12] A. Kolmogoroff, Sur la loi forte des grands nombres, C. R. Acad. Sci. Paris Ser. Math. 191 (1930),
pp. 910–912.
[13] M. Ledoux and M. Talagrand, Probability in Banach Spaces: Isoperimetry and Processes,
Springer-Verlag, Berlin, 1991.
[14] D. Li and A. Rosalsky, Precise lim sup behavior of probabilities of large deviations for sums of
i.i.d. random variables, Int. J. Math. Math. Sci. 2004 (2004), pp. 3565–3576.
[15] D. Li, M.B. Rao, T. Jiang, and X. Wang, Complete convergence and almost sure convergence of
weighted sums of random variables, J. Theoret. Probab. 8 (1995), pp. 49–76.
[16] D. Li, H.-Y. Liang, and A. Rosalsky, A probability inequality for sums of independent Banach
space valued random variables, Stochastics 90 (2018), pp. 214–223.
[17] J. Marcinkiewicz and A. Zygmund, Sur les fonctions indépendantes, Fund. Math. 29 (1937),
pp. 60–90.
[18] M.B. Marcus and W.A. Woyczyński, Stable measures and central limit theorems in spaces of stable
type, Trans. Amer. Math. Soc. 251 (1979), pp. 71–102.
[19] Y. Miao, T.Y. Xue, K. Wang, and F.F. Zhao, Large deviations for dependent heavy tailed random
variables, J. Korean Statist. Soc. 41 (2012), pp. 235–245.
[20] E. Mourier, Eléments aléatoires dans un espace de Banach, Ann. Inst. H. Poincaré 13 (1953),
pp. 161–244.
[21] V.V. Petrov, A generalization of an inequality of Lévy, Teor Veroyatn.Primen. 20 (1975), pp.
140–144, in Russian, translated in Theory Probab. Appl. 20 (1975), pp. 141–145.
[22] V.V. Petrov, Limit Theorems of Probability Theory: Sequences of Independent Random Variables,
Clarendon Press, Oxford, 1995.
[23] T. Rolski, H. Schmidli, V. Schmidt, and J. Teugels, Stochastic Processes for Insurance and Finance,
John Wiley, Chichester, 1999.
[24] X.J. Wang, Upper and lower bounds of large deviations for some dependent sequences, Acta Math.
Hungar. 153 (2017), pp. 490–508.