Deviation inequalities for separately Lipschitz functionals of composition of random functions

(1)

HAL Id: hal-02176147

https://hal.archives-ouvertes.fr/hal-02176147

Preprint submitted on 7 Jul 2019

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires

Deviation inequalities for separately Lipschitz functionals of composition of random functions

Jérôme Dedecker, Paul Doukhan, Xiequan Fan

To cite this version:

Jérôme Dedecker, Paul Doukhan, Xiequan Fan. Deviation inequalities for separately Lipschitz func-

tionals of composition of random functions. 2019. �hal-02176147�

(2)

Deviation inequalities for separately Lipschitz functionals of composition of random functions

J´ erˆ ome Dedecker,

^∗^∗

Paul Doukhan

^††

and Xiequan Fan

^‡‡§

Abstract

We consider a class of non-homogeneous Markov chains, that contains many natural examples.

Next, using martingale methods, we establish some deviation and moment inequalities for separately Lipschitz functions of such a chain, under moment conditions on some dominating random variables.

Keywords: Non-homogeneous Markov chains, martingales, exponential inequalities, moment inequalities.

Mathematics Subject Classification (2010): 60G42, 60J05, 60E15.

1 Introduction

Concentration inequalities are essential tools for ensuring the validity of many statistical procedures; let us cite for instance [18] for classification problems, [21] for model selection, and [3] for high dimensional procedures (see also [2] and [33] in a dependent framework).

These inequalities are not easy to prove in a dependent context; up to now this has been done under quite restrictive assumptions, and mainly for bounded functionals of the variables in a stationary/homogeneous context. As a non exhaustive list, let us quote [31], [27], [12], [1], [10]

and [11]. Among these references, the case of unbounded functionals has been investigated in [1]

for geometrically ergodic Markov chains, and in [10] for iterated random functions satisfying a mean-contraction condition (see condition (2.8) below, with F

n

= F ).

In the paper [10], the authors obtained very precise inequalities for Lipschitz functionals of the chain, by assuming moreover a Lipschitz condition on the function that generates the chain (see condition (2.9) below). However, this last condition is in fact quite restrictive, and does not hold for many natural models satisfying the mean-contraction property.

In the present paper we enlarge the class of Markov chains studied in [10], by considering non-homogeneous Markov chains obtained through composition of random functions (see the recursive mechanism (2.7) below), and by making no extra assumptions than the (uniform) mean- contraction (2.8). As in [10], we shall use the decomposition of the functional of the chain in terms of martingale differences, as first introduced by Yurinskii [32]. This method is well adapted to the non-homogeneous Markov context, because it is intrinsically a non-stationary method. Following this approach, we obtain deviation and moment inequalities for separately Lipschitz functionals of the chain that are driven by the distribution of some dominating random variables.

The present article was conceived within the general framework of non stationary time series, which is now widely discussed in the context global warming [6]. Besides temperatures or

∗Universit´e Paris Descartes, Sorbonne Paris Cit´e, Laboratoire MAP5 and CNRS UMR 8145, 75016 Paris, France.

E-mail: [email protected]

†AGM UMR 8088, Universit´e Paris Seine, UCP site Saint-Martin, 2 Bd. Adolphe Chauvin 95000 Cergy-Pontoise, France, and CIMFAV, Universidad de Valparaiso, Chile. Email: [email protected]

‡Center for Applied Mathematics, Tianjin University, Tianjin, China. E-mail: [email protected]

§All authors contributed equally to this work

(3)

ozone concentration, most of the real life phenomena present trends and periodicities. A first excellent view of those questions may be found in [5], but this is a linear view of time series analysis. It appears important to consider cases where the dynamic of the models itself is non time-homogeneous. For instance [8] and [4] provide different views for a more relevant dynamical approach including local stationarity and non-periodic features. The present paper aims at devel- oping reasonable concentration and probability inequalities for non-homogeneous Markov chains able to model some of the above features.

Before going into further details, let us give a simple class of examples to which our results apply. We consider a generalized

R^k

-valued auto-regressive processes

X

n

= A

n

X

n−1

+ B

n

, (1.1)

where A

n

is a random k × k matrix and B

n

is an

R^k

-valued random variable. Here ε

n

= (A

n

, B

n

) are independent random variables, and (ε

n

)

n≥2

is independent of the initial random variable X

1

. Let | · | be a norm on

R^k

. Then, the Markov chain X

n

satisfies the mean contraction condition (2.8) for the norm | · | as soon as

sup

n≥2E

[|A

n

|] ≤ ρ for some ρ < 1, (1.2) where as usual |A

n

| = sup

_|x|=1

|A

n

x|.

Model (1.1) contains a lot of natural examples (see for instance [9], Sections 2.2 and 2.3), but does not fit within the framework of [10]; moreover, it has no reason to be mixing in the sense of Rosenblatt [30] without further assumptions on the distribution of (ε

n

)

n≥2

. Recall that the chain X

n

is non-homegeneous, since we do not assume here that the variables ε

n

are iid.

Let now f : (

R^k

)

ⁿ

7→

R

be a separately Lipschitz function, such that

|f(x

1

, x

2

, . . . , x

n

) − f(x

⁰₁

, x

⁰₂

, . . . , x

⁰_n

)| ≤ |x

1

− x

⁰₁

| + · · · + |x

n

− x

⁰_n

| , (1.3) and let also S

n

= f(X

1

, . . . , X

n

) −

E

[f(X

1

, . . . , X

n

)].

For simplicity, let us consider the case where the chain starts at X

1

= 0. Assuming that kA

n

k

^pp

:=

E

[|A

n

|

^p

] < ∞ and kB

n

k

^pp

:=

E

[|B

n

|

^p

] < ∞ for any n ≥ 2 and some p > 1, we infer from (2.11) (control of the

L^p

-norm of the dominating variables H

k

(X

k−1

, ε

k

) defined in (2.10)) and Propositions 5.1 and 5.3 that

kS

n

k

^pp

≤ C

1

(p, ρ)

n

X

k=3

kX

k−1

k

²p

kA

k

²p

+

n

X

k=2

kB

k

²p

!2

if p ≥ 2, and

kS

n

k

^pp

≤ C

2

(p, ρ)

n

X

k=3

kX

k−1

k

^pp

kA

k

^pp

+

n

X

k=2

kB

k

^pp

!

if p ∈ (1, 2),

for some constants C

1

(p, ρ), C

2

(p, ρ) depending only on (p, ρ). These inequalities are satisfactory, because if X

n

= B

n

for n ≥ 2 (case A

n

= 0), we recover for p ≥ 2 the usual Marcinkiewicz- Zygmund inequalities (see [28]) for

L^p

-norms of sums of independent random variables, and for p ∈ (1, 2) the usual von Bahr-Esseen inequalities (see [26]). Note that, under the stronger condition than (1.2): sup

_n≥2

kA

n

k

p

≤ ρ (

L^p

-contraction), and if sup

_n≥2

kB

n

k

p

< ∞, we obtain that kS

n

k

p

= O( √

n) if p ≥ 2 and kS

n

k

p

= n

^1/p

if p ∈ (1, 2), which is exactly what we could expect for

L^p

-norms of partial sums in a quasi-stationary regime.

Under more restrictive conditions on (A

n

, B

n

)

n≥2

, one can also obtain some semi-exponential bounds for the deviation of S

n

. For the sake of simplicity, let us assume that sup

_n≥2

kA

n

k

∞

≤ ρ (uniform contraction) and that there exist κ > 0, α ∈ (0, 1) such that

sup

k≥2E h

exp

n

κ|B

k

|

^1−α^2α oi

< ∞ . (1.4)

(4)

It is then easy to see that dominating variables H

k

(X

k−1

, ε

k

) defined in (2.10) also satisfy the uniform bound (1.4) (for the same α and a different κ, say κ

⁰

). Hence, it follows from Proposition 4.2 that

P

(|S

n

| ≥ nx) ≤ C(x) exp

−Kx

^2α

n

^α

, (1.5)

for any x > 0, where the positive constant K depends only on (ρ, α, κ

⁰

), and C(x) = 2 + c(α, ρ, κ

⁰

)

1 x

^2α

+ 1

x

²

.

In particular, we obtain from (1.5) the following moderate deviation behavior: for any δ ∈ (1/2, 1], there exist a > 0 such that

P

|S

n

| ≥ n

^δ

= O exp

n

−an

^α(2δ−1)o

. (1.6)

Note that, for δ = 1, this is in accordance with the best possible rate for large deviation of sums of martingale differences (see Theorem 2.1 in [15]).

2 Composition of random functions

Let (Ω, A,

P

) be a probability space. Let (X , d) and (Y, δ) be two complete separable metric spaces. Let (ε

i

)

i≥2

be a sequence of independent Y-valued random variables. Let X

1

be a X - valued random variable independent of (ε

i

)

i≥2

. We consider the Markov chain (X

i

)

i≥1

such that

X

n

= F

n

(X

n−1

, ε

n

), n ≥ 2, (2.7)

where F

n

: X × Y → X is such that

E

d F

n

(x, ε

n

), F

n

(x

⁰

, ε

n

)

≤ ρ d(x, x

⁰

) (2.8)

for some constant ρ ∈ [0, 1) not depending on n.

In the paper [10], the authors studied a class of homogeneous Markov chains (that is, with F

n

= F and (ε

i

)

i≥2

a sequence of i.i.d. random variables) satisfying (2.8) and the condition

d(F (x, y), F (x, y

⁰

)) ≤ C δ(y, y

⁰

) (2.9) for some positive constant C. Under this additional constraint, they obtained very precise upper bounds for the deviation of separately Lipschitz functionals of the chain; this is possible, because in that case, the martingales differences M

k

from McDiarmid’s decomposition are bounded by a function of ε

k

, which is then independent of the past σ-field of the chain.

However, condition (2.9) is quite restrictive, and is not satisfied for many natural models (a short list of such models is presented below). In the present paper, we shall not assume that (2.9) is satisfied. In this more general setting, the dominating random variables are

H

k

(X

k−1

, ε

k

) where H

k

(x, y) =

Z

d(F

k

(x, y), F

k

(x, y

⁰

))P

ε_k

(dy

⁰

) (2.10) (see Proposition 3.1 below). The main difference with [10] is that these dominating random variables are no longer independent from the past σ-field of the chain. Hence, the deviations bounds that we obtain are not as precise as in [10], but apply to a much larger class of (non homogeneous) Markov chains.

Remark 2.1. Note that if (2.8) holds for the distance d, then, for any α ∈ (0, 1], it also hold

for the distance d

α

(x, y) = (d(x, y))

^α

with ρ

^α

instead of ρ. This is elementary, but nevertheless

important: it means that we can also obtain concentration inequalities for separately Lipschitz

functions with respect to d

α

by controlling the behavior of H

k,α

(X

k−1

, ε

k

) (whose definition is as

in (2.10) for the distance d

α

). Note that separately Lipschitz functions with respect to d

α

are less

and less regular as α approaches 0.

(5)

Remark 2.2. Let us quote an error in the paper [10]. The inequality (1.4) of that paper gives an upper bound for the quantity

E

[H(d(X

n

, x

0

))] when H is any increasing function from

R⁺

to

R⁺

. However this upper bound is not true in general under the assumption (1.2) of [10] (which is similar to our assumption (2.8)), but it holds under the much more restrictive assumption d(F(x, y), F (x

⁰

, y)) ≤ ρ d(x, x

⁰

). The error comes from the fact that the first version of the paper [10] was written under this more restrictive assumption. Note that this wrong inequality was not used at any points in the proofs of the main results in [10], but only in Items 4 of Remarks 3.1 and 3.2 (which are therefore not correct).

2.1 Examples

In this subsection, we give a non exhaustive list of models satisfying condition (2.8), and we show how to control the moments of the dominating variables H

k

(X

k−1

, ε

k

) defined by (2.10). For the sake of simplicity, we shall only deal with the moments of order p of H

k

(X

k−1

, ε

k

), but similar computations may be done for exponential moments. We refer to [13] and [8] for more examples.

• ARCH-type models. For which

F

n

(x, y) = M

θ_n

(x, y), with M

θ

(x, y) =

p

a

²

x

²

+ b

²

· y, θ = (a, b) .

Thus the non stationarity appears simply from changes in the parameter θ

n

= (a

n

, b

n

). In that case (2.8) is satisfied for d(x, x

⁰

) = |x − x

⁰

| provided sup

_n≥2

|a

n

|

E

[|ε

n

|] = ρ for some ρ < 1.

For these models

H

k

(X

k−1

, ε

k

) =

q

a

²_k

X

_k−1²

+ b

²_k Z

|ε

k

− y|P

ε_k

(dy) , and the moments of order p of H

k

(X

k−1

, ε

k

) satisfy

kH

k

(X

k−1

, ε

k

)k

^pp

≤ 2

^p−1E h

a

²k

X

k−1²

+ b

²k

^p/2i

kε

k

^pp

, p ≥ 1.

Those models are easy to extend in an

R^k

-valued framework. For instance, one can consider X

n

= A

n

(X

n−1

) ε

n

, where A

n

(x) is a k × k matrix and ε

n

are

R^k

-valued random variables. Let | · | be a norm on

R^k

, and |A| = sup

_|x|=1

|A

n

x| be the associated matrix norm. Now, if |A

n

(x) − A

n

(x

⁰

)| ≤ a

n

|x − x

⁰

|, then the condition (2.8) is satisfied as soon as sup

_n≥2

a

nE

[|ε

n

|] = ρ for some ρ < 1.

• Switching models. Many analogous models can be provided with a switching, e.g. for the first ARCH-model, such a parametric model is given with X =

R

, Y =

R

× {0, 1}, a parameter θ = (a, b, a

⁰

, b

⁰

) ∈

R⁴

and

M

θ

(x, y) = y

2

p

a

²

x

²

+ b

²

· y

1

+ (1 − y

2

)

p

a

⁰²

x

²

+ b

⁰²

· y

1

.

Here (ε

n

)

n≥2

is a sequence of independent random variables with values in

R

× {0, 1}. Using the notation ε

n

= (ε

⁽¹⁾n

, ε

⁽²⁾n

), we see that condition (2.8) is satisfied as soon as

sup

n≥2

(1 −

E

[ε

⁽²⁾_n

])|a

n

| +

E

[ε

⁽²⁾_n

]|a

⁰n

|

E

[|ε

⁽¹⁾n

|] = ρ for some ρ < 1.

Now, similar computations as for the first example lead to kH

k

(X

k−1

, ε

k

)k

^pp

≤ 4

^p−1E

h

a

²_k

X

k−1²

+ b

²_kp/2i

ε

⁽¹⁾_k

ε

⁽²⁾_k

p p

+ 4

^p−1E h

a

⁰²_k

X

k−1²

+ b

⁰²_kp/2i

ε

⁽¹⁾_k

(1 − ε

⁽²⁾_k

)

p p

, p ≥ 1.

(6)

• Generalized

R^k

-valued auto-regressive processes. We consider here the Model (1.1) presented in the introduction. Recall that A

n

is a random k × k matrix and B

n

is an

R^k

- valued random variable. Here ε

n

= (A

n

, B

n

) are independent random variables, and (ε

n

)

n≥2

is independent of the initial random variable X

1

. Model (1.1) is a composition of random functions as in (2.7), with

F

n

(x, y) = F (x, y) = y

1

x + y

2

.

Let | · | be a norm on

R^k

, and let as usual |A

n

| = sup

_|x|=1

|A

n

x|. The condition (2.8) is satisfied as soon as (1.2) holds.

For these models

H

k

(X

k−1

, ε

k

) ≤ |X

k−1

|

Z

|A

k

− y|P

A_k

(dy) +

Z

|B

k

− y|P

B_k

(dy) , and the moments of order p of H

k

(X

k−1

, ε

k

) satisfy

kH

k

(X

k−1

, ε

k

)k

^pp

≤ 4

^p−1E

[|X

k−1

|

^p

]

E

[|A

k

|

^p

] + 4

^p−1E

[|B

k

|

^p

] , p ≥ 1. (2.11)

• INAR(1) type models. In this case, let y = (y

0

, y

1

, y

2

, . . . , y

p

, . . .) ∈ Y =

N^N

and ε = (ε

⁽⁰⁾

, ε

⁽¹⁾

, . . . , ε

^(p)

, . . .), where (ε

⁽¹⁾

, . . . , ε

^(p)

, . . .) is a sequence of i.i.d. integer valued random variables. The function F is then given by

F (x, y) = y

0

+ 1

{x6=0}

x

X

k=1

y

k

.

Here, (ε

n

)

n≥2

is an i.i.d. sequence distributed as ε. It is then easy to see that (2.8) is satisfied provided ρ =

E

[ε

⁽¹⁾

] < 1.

We shall now give some hints to control the moments of the dominating random variables H

k

(X

k−1

, ε

k

). Let ˜ ε

k

be distributed as ε

k

and independent of (ε

k

, X

k−1

). We then have that

H

k

(X

k−1

, ε

k

) =

E

d(F

k

(X

k−1

, ε

k

), F

k

(X

k−1

, ε ˜

k

))

Xk−1

, ε

k

. For the INAR(1) model, we have

H

k

(X

k−1

, ε

k

) =

E

(ε

⁽⁰⁾_k

− ε ˜

⁽⁰⁾_k

) + 1

{X_k−16=0}

X_k−1

X

i=1

(ε

⁽ⁱ⁾_k

− ε ˜

⁽ⁱ⁾_k

)

X

k−1

, ε

k

By contraction, we get that kH

k

(X

k−1

, ε

k

)k

^pp

≤

(ε

⁽⁰⁾_k

− ε ˜

⁽⁰⁾_k

) + 1

{X_k−16=0}

X_k−1

X

i=1

(ε

⁽ⁱ⁾_k

− ε ˜

⁽ⁱ⁾_k

)

p

, p ≥ 1.

For p ≥ 2, applying the Marcinkiewicz-Zygmund inequality given in [28], we get that kH

k

(X

k−1

, ε

k

)k

^p_p

≤ 2

^(p−2)/2

kε

⁽⁰⁾

− ε ˜

⁽⁰⁾

k

^p_p

+ (p − 1)

^p/2

2

^(p−2)/2E

h

X

_k−1^p/2i

kε

⁽¹⁾

− ε ˜

⁽¹⁾

k

^p_p

. For p ∈ (1, 2), applying the von-Bahr Essen inequality given in [26], we get that

kH

k

(X

k−1

, ε

k

)k

^pp

≤ kε

⁽⁰⁾

− ε ˜

⁽⁰⁾

k

^pp

+ 2

^2−pE

[X

k−1

] kε

⁽¹⁾

− ˜ ε

⁽¹⁾

k

^pp

.

Note that non-stationary variants of this model can be obtained by considering independent

(but non i.i.d.) ε

n

’s, with the constraint: sup

_n≥2E

[ε

⁽¹⁾n

] < 1.

(7)

• GLM-Poisson models. Besides the standard ARCH-models the simplest case is that of Poisson ARCH–models, where (ε

n

)

n≥2

is a sequence of i.i.d. unit Poisson processes. Consider a sequence of functions f

n

:

N

→

R⁺

and set

F

n

(x, y) = y(f

n

(x)),

where y :

R⁺

→

N

denotes a function. In that case, the condition (2.8) is satisfied if

|f

n

(x) − f

n

(x

⁰

)| ≤ ρ|x − x

⁰

| for any n ≥ 2, any x, x

⁰

∈

N

, and some ρ < 1.

For these models,

H

k

(X

k−1

, ε

k

) =

E

|ε(f

k

(X

k−1

)) − ε

k

(f

k

(X

k−1

))|

εk

, X

k−1

,

where ε is a unit Poisson process independent of (X

k−1

, ε

k

). By contraction, we get that kH

k

(X

k−1

, ε

k

)k

^pp

≤ 2

^p−1

kε

k

(f

k

(X

k−1

))k

^pp

= 2

^p−1E

[Q

p

(f

k

(X

k−1

))] ,

where Q

p

(t) = kε(t)k

^pp

. Note that, when p is an integer, Q

p

denotes the Stirling polynomial defined through Stirling numbers (see the Lemma A-1 in [13] from [17]).

• GLM–GARCH Poisson models. One can give numerous extensions of the previous model. Keeping the same notations, one can consider f

n

:

R⁺

×

N

→

R⁺

, and

F

n

(x, y) = (f

n

(x), y(f

n

(x)), where x = (λ, z) ∈

R⁺

×

N

. Let |x| = |λ| + a|z|. Then

E

[|F

n

(x, ε) − F

n

(x

⁰

, ε

n

)|] ≤ (1 + a)||f

n

(x) − f

n

(x

⁰

)| ,

and (2.8) is true provided that |f

n

(x) − f

n

(x

⁰

)| ≤ L

n

|x − x

⁰

| and sup

_n≥2

L

n

(1 + a) = ρ for some ρ < 1.

For these models,

H

k

(X

k−1

, ε

k

) = a

E

|ε(f

k

(X

k−1

)) − ε

k

(f

k

(X

k−1

))|

εk

, X

k−1

,

where ε is a unit Poisson process independent of (X

k−1

, ε

k

). Hence, the moments of order p of H

k

(X

k−1

, ε

k

) can be controlled exactly as in the previous example.

3 Separately Lipschitz functions of X ₁ , . . . , X _n

Let f : X

ⁿ

7→

R

be a separately Lipschitz function, such that

|f(x

1

, x

2

, . . . , x

n

) − f(x

⁰₁

, x

⁰₂

, . . . , x

⁰_n

)| ≤ d(x

1

, x

⁰₁

) + · · · + d(x

n

, x

⁰_n

) . (3.1) Let

S

n

:= f(X

1

, . . . , X

n

) −

E

[f(X

1

, . . . , X

n

)] . (3.2) We introduce the natural filtration of the chain, that is F

0

= {∅, Ω} and for all k ∈

N^∗

, F

k

= σ(X

1

, X

2

, . . . , X

k

). Define

g

k

(X

1

, . . . , X

k

) =

E

[f(X

1

, . . . , X

n

)|F

k

] (3.3) and

M

k

= g

k

(X

1

, . . . , X

k

) − g

k−1

(X

1

, . . . , X

k−1

). (3.4) For all k ∈ [1, n − 1], let

S

k

:= M

1

+ M

2

+ · · · + M

k

,

and notice that, by the definition of M

k

’s, the functional S

n

introduced in (3.2) satisfies

S

n

= M

1

+ M

2

+ · · · + M

n

.

(8)

Thus S

k

is a martingale adapted to the natural filtration F

k

. This representation appears in Yurinskii [32] and in p. 33 of the monograph by Milman and Schechtman [24]. In the setting of separately Lipschitz functions of independent random variables (i.e. when X

i

= ε

i

) it has been used by McDiarmid [23] to get an exponential bound on tail probabilities

P

(S

n

≥ x), x ≥ 0.

The following Proposition, similar to Proposition 2.1 in [10], collects some interesting proper- ties of the function g

k

and of the martingale difference M

k

.

Proposition 3.1. For all k ∈

N

and any ρ in [0, 1), denote K

k

(ρ) = (1 − ρ

^k+1

)/(1 − ρ) = 1 + ρ + · · · + ρ

^k

. Let (X

i

)

i≥1

be a Markov chain satisfying (2.7) for some functions F

n

satisfying (2.8). Let g

k

and M

k

be defined by (3.3) and (3.4) respectively.

1. The function g

k

is separately Lipschitz, and satisfies

gk

(x

1

, x

2

, . . . , x

k

) −g

k

(x

⁰1

, x

⁰2

, . . . , x

⁰k

)

≤ d(x

1

, x

⁰1

) +· · ·+ d(x

k−1

, x

⁰k−1

) +K

n−k

(ρ)d(x

k

, x

⁰k

) . 2. Denote by P

X₁

and P

ε_k

the distribution of X

1

and the distribution of ε

k

respectively. Let

G

X₁

and H

k

be two functions defined as follows

G

X₁

(x) =

Z

d(x, x

⁰

)P

X₁

(dx

⁰

) and H

k

(x, y) =

Z

d(F

k

(x, y), F

k

(x, y

⁰

))P

ε_k

(dy

⁰

) . Then, the martingale difference M

k

satisfies

|M

1

| ≤ K

n−1

(ρ)G

X₁

(X

1

) and |M

k

| ≤ K

n−k

(ρ)H

k

(X

k−1

, ε

k

), k ∈ [2, n].

Remark 3.1. Assume moreover that F

n

satisfies

d(F

n

(x, y), F

n

(x, y

⁰

)) ≤ C(x) δ(y, y

⁰

) (3.5) for some function C(x) ≥ 0 not depending on n, and let G

k

be the function defined by

G

k

(y) =

Z

δ(y, y

⁰

)P

ε_k

(dy

⁰

) for all k ∈ [2, n].

Then H

k

(x, y) ≤ C(x)G

k

(y) and, consequently,

|M

k

| ≤ K

n−k

(ρ)C(X

k−1

)G

k

(ε

k

) for all k ∈ [2, n].

Note that (3.5) is a non-uniform version of (2.9), which is satisfied for many examples (for instance the three first examples of Section 2). However, it seems quite difficult to check for INAR or GLM type models, while the moments of the dominating variables H

k

(X

k−1

, ε

k

) are easy to control for such models (see Section 2).

Proof. The first point will be proved by recurrence in the backward sense. For k = n, the result is obvious due to g

n

= f. Suppose it is true at step k, and let us prove it at step k − 1. By definition

g

k−1

(X

1

, . . . , X

k−1

) =

E

[g

k

(X

1

, . . . , X

k

)|F

k−1

] =

Z

g

k

(X

k

, . . . , X

k−1

, F

k

(X

k−1

, y))P

ε_k

(dy) . Then it is easy to see that

|g

k−1

(x

1

, x

2

, . . . , x

k−1

) − g

k−1

(x

⁰1

, x

⁰2

, . . . , x

⁰k−1

)|

≤

Z

g

k

(x

1

, x

2

, . . . , F

k

(x

k−1

, y)) − g

k

(x

⁰₁

, x

⁰₂

, . . . , F

k

(x

⁰k−1

, y))

P

ε_k

(dy) . (3.6) Now, by assumption and condition (2.8),

Z

g

k

(x

1

, x

2

, . . . , F

k

(x

k−1

, y)) − g

k

(x

⁰₁

, x

⁰₂

, . . . , F

k

(x

⁰k−1

, y))

P

ε_k

(dy)

≤ d(x

1

, x

⁰₁

) + · · · + d(x

k−1

, x

⁰k−1

) + K

n−k

(ρ)

Z

d(F

k

(x

k−1

, y), F

k

(x

⁰k−1

, y))P

ε_k

(dy)

≤ d(x

1

, x

⁰₁

) + · · · + (1 + ρK

n−k

(ρ))d(x

k−1

, x

⁰k−1

)

≤ d(x

1

, x

⁰₁

) + · · · + K

n−k+1

(ρ)d(x

k−1

, x

⁰k−1

) . (3.7)

(9)

The point 1 follows from (3.6) and (3.7).

Next, we prove the point 2. First notice that

|M

1

| =

g

1

(X

1

) −

Z

g

1

(x)P

X₁

(dx)

≤ K

n−1

(ρ)

Z

d(X

1

, x)P

X₁

(dx) = K

n−1

(ρ)G

X₁

(X

1

) . Similarly, for all k ≥ 2,

|M

k

| =

g

k

(X

1

, · · · , X

k

) −

E

[g

k

(X

1

, · · · , X

k

)|F

k−1

]

≤

Z

g

k

(X

1

, · · · , F

k

(X

k−1

, ε

k

)) − g

k

(X

1

, · · · , F

k

(X

k−1

, y))

P

ε_k

(dy)

≤ K

n−k

(ρ)

Z

d(F

k

(X

k−1

, ε

k

), F

k

(X

k−1

, y))P

ε_k

(dy) = K

n−k

(ρ)H

k

(X

k−1

, ε

k

) . This completes the proof of Proposition 3.1.

4 Deviation inequalities for the functional S _n

Let (X

i

)

i≥1

be a Markov chain satisfying (2.7) for some functions F

n

satisfying (2.8). In this section, we apply inequalities for martingales to bound up the deviation of the functional S

n

defined by (3.2). Some of these inequalities are direct applications of known inequalities, and some deserve a short proof.

Denote by S

2,n

= S

n

−M

1

, and let a

n

be a sequence of positive numbers. Then, for any x > 0,

P

S

n

≥ a

n

x

≤

P

M

1

≥ a

n

x/2 +

P

S

2,n

≥ a

n

x/2

≤

P

G

X₁

(X

1

) ≥ a

n

x 2K

n−1

(ρ)

+

P

S

2,n

≥ a

n

x/2

=: I

1

(a

n

, x) + I

2

(a

n

, x) , (4.1) and note that the same bound is valid for

P

(−S

n

≥ a

n

x) by replacing the term I

2

(a

n

, x) by I ˜

2

(a

n

, x) :=

P

(−S

2,n

≥ a

n

x/2).

The term I

1

(a

n

, x) will be most of the time negligible, and represents the direct influence of the initial distribution of the chain. For instance, when the chain starts from a point X

1

= x

1

, then G

X1

(X

1

) = 0 and I

1

(a

n

, x) = 0. The main difficulty is to give an upper bound for I

2

(a

n

, x), which is the purpose of the present paper.

4.1 A first exponential bound

Under a sub-Gaussian type condition, we obtain the following proposition.

Proposition 4.1. Assume that there exists a positive constant such that, for any integer k ≥ 2,

E

h

H

k

(X

k−1

, ε

k

)

li

≤ 1 2

l!

^l−2

(l − 1)

^l/2E

h

H

k

(X

k−1

, ε

k

)

2i

for all l ≥ 2. (4.2) Then, for any x > 0,

P

± S

n

≥ xV

n

≤ I

1

(V

n

, x) + exp

− (x/2)

²

1 +

p

1 + xK

n−2

(ρ)/σ

n

+ xK

n−2

(ρ)/2σ

n

(4.3)

≤ I

1

(V

n

, x) + exp

− (x/2)

²

2 1 + xK

n−2

(ρ)/2σ

n

, (4.4)

where

V

n²

=

n

X

k=2

K

n−k²

(ρ)

E

H

k

(X

k−1

, ε

k

)

2

and σ

n²

= 1

n V

n²

.

(10)

Remark 4.1. Let us give some comments on Proposition 4.1.

1. Condition (4.2) is in fact a sub-Gaussian condition. On can check that it is satisfied provided inf

k≥2E

H

k

(X

k−1

, ε

k

)

2

> 0 and

sup

k≥2E h

exp

n

c H

k

(X

k−1

, ε

k

)

2oi

< ∞ for some positive constant c not depending on k.

2. Assume that

E h

exp

n

c

p

G

X₁

(X

1

)

o i

< ∞ for some positive constant c, and that

0 < lim inf

n→∞

σ

n

≤ lim sup

n→∞

σ

n

< ∞.

Then, it follows from Proposition 4.1 that

P

(±S

n

≥ n) = O exp

−C √

n (4.5)

for some positive constant C.

Proof. By Taylor’s expansion of e

^x

and the fact that

E

[S

2,n

] = 0, we have, for all t ≥ 0,

E

exp

t S

2,n

√ n

= 1 +

∞

X

k=2

t

^k

k!

E

h

S

2,n

√ n

ki

. (4.6)

Using Rio’s inequality (see Theorem 2.1 of [28]): for any p ≥ 2,

E

[|S

2,n

|

^p

]

2/p

≤ (p − 1)

n

X

i=2

E

[|M

i

|

^p

]

2/p

, (4.7)

we get, for all k ≥ 2,

E

[|S

2,n

|

^k

] ≤ (k − 1)

^k/2Xⁿ

i=2

E

[|M

i

|

^k

]

2/kk/2

. (4.8)

Hence, by H¨ older’s inequality, inequality (4.8) implies that, for all k ≥ 2,

E

[|S

2,n

|

^k

] ≤ (k − 1)

^k/2

n

^k/2−1

n

X

i=2

E

[|M

i

|

^k

]. (4.9)

Applying the last inequality to (4.6), we obtain

E

exp

t S

2,n

√ n

≤ 1 +

∞

X

k=2

t

^k

k! (k − 1)

^k/2

n

⁻¹

n

X

i=2

E

[|M

i

|

^k

]

. (4.10)

By points 2 of Proposition 3.1 and (4.2), we deduce that, for any integer i ≥ 2,

E

[|M

i

|

^l

] ≤

E

[|K

n−i

(ρ)H

i

(X

i−1

, ε

i

)|

^l

]

≤ 1 2

l! (K

n−2

(ρ))

^l−2

(l − 1)

^l/2 E

[(K

n−i

(ρ)H

i

(X

i−1

, ε

i

))

²

] for all l ≥ 2.

Hence condition (4.2) implies that, for all 0 ≤ t < (K

n−2

(ρ))

⁻¹

,

E

exp

t S

2,n

√ n

≤ 1 +

∞

X

k=2

σ

²n

2 t

^k

(K

n−2

(ρ))

^k−2

= 1 + t

²

σ

²n

2 (1 − tK

n−2

(ρ)) . (4.11)

(11)

By the inequality 1 + x ≤ e

^x

, it follows that, for all 0 ≤ t < (K

n−2

(ρ))

⁻¹

,

E

exp

t S

2,n

√ n

≤ exp

(

t

²

σ

n²

2 (1 − tK

n−2

(ρ))

)

.

Applying Markov’s inequality, it is then easy to see that, for all 0 ≤ t < σ

n

(K

n−2

(ρ))

⁻¹

and x ≥ 0,

P

(S

2,n

≥ xV

n

/2) ≤ exp

n

− tx/2

o E

exp

t S

2,n

V

n

.

Hence

P

(S

2,n

≥ xV

n

/2) ≤ inf

0≤t<σ_n(Kn−2(ρ))⁻¹

exp

− tx/2 + t

²

2 (1 − t K

n−2

(ρ)/σ

n

)

= exp

− (x/2)

²

1 +

p

1 + xK

n−2

(ρ)/σ

n

+ xK

n−2

(ρ)/2σ

n

,

which gives (4.3). Using the inequality

p

1 + xK

n−2

(ρ)/σ

n

≤ 1 + xK

n−2

(ρ)/2σ

n

, we get (4.4) from (4.3).

4.2 Semi-exponential bounds

In the case where the variables H

k

(X

k−1

, ε

k

) have semi-exponential moments, the following proposition holds.

Proposition 4.2. Let α ∈ (0, 1). Assume that there exists a constant C

1

such that, for any integer k ≥ 2,

E h

exp

n

H

k

(X

k−1

, ε

k

)

_1−α^2α oi

≤ C

1

. (4.12)

Then, for any x > 0,

P

(±S

n

≥ nx) ≤ I

1

(n, x) + C(α, x) exp

(

−

x

8K

n−2

(ρ)

2α

n

^α )

, (4.13)

where

C(α, x) = 2 + 35C

1

K

n−2^2α

(ρ)

x

^2α

4

^2−3α

+ 4K

n−2²

(ρ) x

²

3(1 − α) 2α

^1−α_α !

depends on n only through the term K

n−2

(ρ).

Remark 4.2. Let us comment on inequality (4.13). Assume moreover that

E

h

exp

n

c

G

X₁

(X

1

)

αoi

≤ C

2

, (4.14)

for two positive constants c, C

2

. Then, it follows from (4.13) that

P

(±S

n

≥ n) = O (exp {−Cn

^α

}) , (4.15)

for some positive constant C. This rate is in accordance with the best possible rate for large devi-

ation of partial sums of martingales differences, as proved in Theorem 2.1 of [15]. For partial sums

of independent random variables, the rate (4.15) holds under weaker conditions on exponential

moments, see Lanzinger and Stadtm¨ uller [22].

(12)

Proof. From point 2 of Proposition 3.1 and condition (4.12), it is easy to see that, for any k ∈ [2, n],

E

h

exp

n

|K

_n−2⁻¹

(ρ)M

k

|

^1−α^2α oi

≤ C

1

. (4.16)

Applying Theorem 2.1 of Fan et al. [15] to the martingale sequence (K

_n−2⁻¹

(ρ)M

k

, F

k

)

k=2,..,n

, we get, for any x > 0,

I

2

(n, x) ≤ C(α, x) exp

(

−

x

8K

n−2

(ρ)

2α

n

^α )

. (4.17)

Combining the inequalities (4.1) and (4.17), we obtain the desired inequality.

For the next proposition, let us introduce the random variables L

k

(X

k−1

), where L

k

(x) =

Z

(H

k

(x, y))

²

P

ε_k

(dy) , (4.18) and note that

L

k

(X

k−1

) =

E

(H

k

(X

k−1

, ε

k

))

² Xk−1

.

According to Proposition 3.1, for any k ≥ 2,

E

[M

_k²

|F

k−1

] ≤ K

n−k

(ρ)L

k

(X

k−1

).

Proposition 4.3. Assume (3.5), and let α ∈ (0, 1). Assume that there exist two constants C

1

and C

2

such that, for any integer k ≥ 2, n ≥ 2,

E

"

exp 1

n

X

k=2

L

k

(X

k−1

)

_1−α^α #

≤ C

1

(4.19)

and

E

exp

H

k

(X

k−1

, ε

k

)

_1−α^α

≤ C

2

. (4.20)

Then, for all x > 0,

P

(±S

n

≥ nx) ≤ I

1

(n, x) + exp

(

− (xK

_n−2⁻¹

(ρ)/2)

^1+α

2 1 + xK

_n−2⁻¹

(ρ)/6 n

^α

)

+ (C

1

+ nC

2

) exp

n

− xK

n−2⁻¹

(ρ)/2

^α

n

^αo

. (4.21) Remark 4.3. According to Remark 4.2, under the conditions (4.19), (4.20) and (4.14), we have

P

(±S

n

≥ n) = O (exp {−Cn

^α

}) , (4.22)

for some positive constant C. This rate is in accordance with the best possible rate for large deviation of partial sums of martingales differences, as proved in Corollary 2.3 of [15]. Note that if α ∈ [1/2, 1), the condition (4.19) is true provided that

sup

k≥2E

exp

(L

k

(X

k−1

))

^1−α^α

≤ C

1

. (4.23)

For α ∈ [1/2, 1), the two conditions (4.23) and (4.20) are clearly less restrictive than (4.12), so Proposition 4.3 is more precise than Proposition 4.2 in the regime of large deviation. However, it does not allow to control moderate deviations

P

±S

n

≥ n

^δ

for δ close to 1/2, which is possible via Proposition 4.2 (see for instance (1.6)).

Proof. From point 2 of Proposition 3.1 and condition (4.20), it is easy to see that, for any k ∈ [2, n],

E

h

exp

n

|K

n−2⁻¹

(ρ)M

k

|

^1−α^α oi

≤ C

2

.

(13)

For any k ∈ [2, n],

E

h

|K

n−2⁻¹

(ρ)M

k

|

²

F

k−1

i

≤

E h

K

_n−2⁻¹

(ρ)K

n−k

(ρ)H

k

(X

k−1

, ε

k

)

2

F

k−1

i

≤

E h

H

k

(X

k−1

, ε

k

)

2

F

k−1

i

. (4.24)

Thus

n

X

k=2

E

h

|K

n−2⁻¹

(ρ)M

k

|

²

F

k−1

i

≤

n

X

k=2

E h

H

k

(X

k−1

, ε

k

)

2

F

k−1

i

=

n

X

k=2

L

k

(X

k−1

).

Using Theorem 2.2 of Fan et al. [15], we have

P

K

_n−2⁻¹

(ρ)S

2,n

≥ nK

_n−2⁻¹

(ρ)x/2 and

n

X

k=2

L

k

(X

k−1

) ≤ nv

²

≤ exp

(

− (K

_n−2⁻¹

(ρ)x/2)

²

2 n

^α−1

v

²

+

¹₃

(K

⁻¹_n−2

(ρ)x/2)

^2−α

n

^α )

+ nC

2

exp

(

− x

2K

n−2

(ρ)

α

n

^α )

.

From the last inequality, we deduce that I

2

(n, x) ≤ exp

(

− (K

_n−2⁻¹

(ρ)x/2)

²

2 n

^α−1

v

²

+

¹₃

(K

_n−2⁻¹

(ρ)x/2)

^2−α

n

^α )

+ nC

2

exp

(

− x

2K

n−2

(ρ)

α

n

^α )

+

P n

X

k=2

L

k

(X

k−1

) > nv

²

.

Using the exponential Markov inequality and the condition (4.19), we get, for all v > 0,

P

n

X

k=2

L

k

(X

k−1

) > nv

²

≤ C

1

exp

n

− v

^1−α^2α o

.

Taking v

²

= (xnK

_n−2⁻¹

(ρ)/2)

^(1−α)/2

, we have, for all x > 0, I

2

(n, x) ≤ exp

(

− (xK

_n−2⁻¹

(ρ)/2)

^1+α

2 1 + xK

_n−2⁻¹

(ρ)/6 n

^α

)

+ (C

1

+ nC

2

) exp

n

− xK

_n−2⁻¹

(ρ)/2

α

n

^αo

.

Combining the last inequality and (4.1), we obtain the desired inequality.

4.3 Fuk-Nagaev type bound

We now consider the case where the random variables H

k

(X

k−1

, ε

k

), k ≥ 2, have only a weak moment of order p > 2. For any real-valued random variable Z and any p ≥ 1, define the weak moment of order p by

kZ k

^pw,p

= sup

x>0

x

^pP

(|Z| > x) . (4.25)

The following proposition is a Fuk-Nagaev type inequality (cf. Fuk [19] and Nagaev [25]; see also Fan et al. [16] and Rio [29] for martingales).

Proposition 4.4. Let p ≥ 2 and δ > 0, and consider the variables L

k

(X

k−1

) defined in (4.18).

Assume that there exist two constants C

1

and C

2

such that, for any integer k ≥ 2,

1 n

n

X

k=2

L

k

(X

k−1

)

p+δ

w,p+δ

≤ C

1

(4.26)

and

kH

k

(X

k−1

, ε

k

)k

^p+δ_w,p+δ

≤ C

2

. (4.27)

(14)

Then, for all x > 0,

P

(±S

n

≥ nx) ≤ I

1

(n, x) + exp

(

− (K

_n−2⁻¹

(ρ)/2)

²

2 n

^−1/(p+δ)

x

⁻¹

+

¹₆

K

_n−2⁻¹

(ρ) (nx)

^δ/(p+δ) )

+ C

1

+ C

2

n

^p−1

x

^p

. (4.28) Remark 4.4. Let us comment on Proposition 4.4.

1. If there exists a constant C

3

such that

||G

X₁

(X

1

)||

^p−1_w,p−1

≤ C

3

, (4.29) then, for any x > 0,

I

1

(n, x) =

P

G

X₁

(X

1

) ≥ nx 2K

n−1

(ρ)

≤ (2K

n−1

(ρ))

^p−1

C

3

(nx)

^p−1

. (4.30) Thus under conditions (4.26), (4.27) and (4.29), we have

P

(±S

n

≥ n) = O

1 n

^p−1

.

2. Assume moreover that F

n

satisfies (3.5). Then, according to Remark 3.1, Proposition 4.4 remains valid when H

k

(X

k−1

, ε

k

) is replaced by C(X

k−1

)G

k

(ε

k

). Since C(X

k−1

) and G

k

(ε

k

) are independent, we easily see that

L

k

(X

k−1

) ≤ (C(X

k−1

))

²E

(G

k

(ε

k

))

²

and

kC(X

k−1

)G

k

(ε

k

)k

^p+δ_w,p+δ

≤

E h

C(X

k−1

)

p+δi

kG

k

(ε

k

)k

^p+δ_w,p+δ

. (4.31) Thus if

E h

C(X

k−1

)

p+δi

≤ C

3

and kG

k

(ε

k

)k

^p+δ_w,p+δ

≤ C

4

,

then condition (4.27) is satisfied with C

2

= C

3

C

4

. Of course, the same computations may be done by interchanging C(X

k−1

) and G

k

(ε

k

). Hence, if

E h

G

k

(ε

k

)

p+δi

≤ C

3

and kC(X

k−1

)k

^p+δ_w,p+δ

≤ C

4

, then condition (4.27) still holds with C

2

= C

3

C

4

.

Proof. To prove Proposition 4.4, we need the following inequality whose proof can be found in Fan et al. [14] (cf. Corollary 2.3 and Remark 2.1 therein).

Lemma 4.1. Assume that (ξ

i

, G

i

)

i≥1

are square integrable martingale differences, and let Z

n

= ξ

1

+ · · · + ξ

n

and hZ i

n

=

Pn

k=1E

[ξ

k²

|G

k−1

]. Then, for all x, y, v > 0,

P

Z

n

≥ x and hZ i

n

≤ v

²

≤ exp

− x

²

2(v

²

+

¹₃

xy)

+

P

1≤i≤n

max ξ

i

> y

.

By Lemma 4.1 and Markov’s inequality, it follows that, for all x, y, v > 0,

P

K

n−2⁻¹

(ρ)S

2,n

≥ nK

n−2⁻¹

(ρ)x/2 and

n

X

k=2

L

k

(X

k−1

) ≤ nv

²

!

≤ exp

(

− (nK

_n−2⁻¹

(ρ)x/2)

²

2 nv

²

+

¹₆

nK

_n−2⁻¹

(ρ)xy

)

+

P

2≤i≤n

max K

n−2⁻¹

(ρ)M

i

≥ y

.