Deviation inequalities for sums of weakly dependent time series 493
2.3 Extensions to the product space
X
q
, q 1.
To consider conditional coupling schemes of length q 1, we define the notion of ϕ-coefficients
for X = X
t r
≤tr+q
∈ X
q
.
Definition 2.2. For any q ≥ 1, any X ∈ X
q
and any σ-algebra M of A let us define the coefficients
ϕM , X = sup{kE f X |M − E f X k
∞
, f ∈ F
q
}, where
F
q
is the set of 1-Lipschitz functions with values in [
−12, 12] of X
d
equipped with the metric d
q
x, y = q
−1
P
q i=
1
dx
i
, y
i
. Let us discuss the consequences of the choice of the metric d
q
: • The coefficient τ
∞
is defined for X
q
equipped with the same metric d
q
, see [6]. Thus we have
τ
∞
M , X = ϕM , X under A and τ
∞
M , f X
1
, . . . , f X
q
≤ ϕM , X
1
, . . . , X
q
for all f ∈ F .
• As d
q
x, y ≤ 11
x 6= y
then ϕM , X ≤ φM , σX ; our definition of the coefficient ϕ differs
from the one of Rio in [21] where X
q
is equipped with d
∞
x, y = max
1 ≤i≤q
dx
i
, y
i
.
3 A deviation inequality under
ϕ-weak dependence
Assume that there exists a non increasing sequence δ
r r
satisfying sup
1 ≤ j≤n−2r+1
ϕM
j
, , X
r+ j
, . . . , X
2r+ j −1
≤ δ
r
for all r ≥ 1.
3.1
3.1 A Bernstein type inequality
Theorem 3.1. If condition 3.1 is satisfied, for any f
∈ F such that E f X
1
= 0 we have P
S f
≥ 5.8 Æ
n σ
2 k
∗
f x + 1.5 k
∗
x
≤ e
−x
, 3.2
where k
∗
= min{1 ≤ k ≤ n kδ
k
≤ σ
2 k
f } and σ
2 k
∗
f = max{σ
2 k
f k
∗
≤ k ≤ n}. The proof of this Theorem is given in Subsection 6.1. We adopt the convention min
; = +∞ and the estimate is non trivial when r
δ
r
→ 0 and nδ
n
≥ σ
n
f , i.e. for not too small values of n.
3.2 The variance terms
σ
2 k
f
Before giving some remarks on Theorem 3.1, the next proposition give estimates of the quantity σ
2 k
f = k
−1
Var P
k i=
1
f X
i
:
Proposition 3.2. Under the assumption of Theorem 3.1, for any 1 ≤ k ≤ n we have
σ
2 k
f ≤ σ
2 1
f + 2E| f X
1
|
k −1
X
r= 1
δ
r
. See Subsection 6.2 for a proof of Proposition 3.2. The estimate given in Proposition 3.2 can be
rough, for example in the degenerate case when σ
2 k
f tends to 0 with k. Note also that this estimate is often useless when the correlations terms are summable as the inequality
σ
2 k
f ≤ σ
2 1
f P
k ≥1
| Corr f X , f X
k
| may lead to better estimates.
494 Electronic Communications in Probability
3.3 Remarks on Theorem 3.1