, , , December 25, 2004.

In this paper, we study the moderate deviation principle (in short: MDP) for a family

(S_{t}^{κ})_{t \to \infty}

κ \in (\frac{1}{2}, 1)

S_{t}^{κ} = \frac{1}{t^{κ}} \int_{0}^{t} H (X_{s}) d s,

where

X = (X_{t})_{t \geq 0}

is an ergodic diffusion process (

X_{t} \in R^{d}

d \geq 1

) (with the unique invariant measure

μ (d z)

, obeying the density

p (z)

relative to Lebesgue measure over

R^{d}

The function

H : R^{d} \to R^{q}

is assumed to be integrable relative to

μ (d z)

and has zero barycenter

\begin{matrix} \int_{R^{d}} H (z) p (z) d z = 0 . \end{matrix}

(1.1)

We restrict ourselves by consideration of the strong (unique) solution of Itô's equation

\begin{matrix} d X_{t} = b (X_{t}) d t + σ (X_{t}) d W_{t} \end{matrix}

(1.2)

generated by a standard vector-valued Wiener process

W = (W_{t})_{t \geq 0}

and subject to a fixed initial point,

X_{0} = x

. We also include into the consideration a linear version of 1.2 (here

A, B

are matrices):

\begin{matrix} d X_{t} = A X_{t} d t + B d W_{t} \end{matrix}

(1.3)

being popular in engineering.

In a nonlinear case, we use Veretennikov Khasminskii's condition (see, [14] and [28] ): for some positive numbers

r

C

and

α

, (here

〈 〈 \cdot 〉 〉

denotes the inner product)

〈 〈 z, b (z) 〉 〉 \leq - r ∥ z ∥^{1 + α}, ∥ z ∥ > C

and assume that the diffusion matrix

a (x) = σ σ^{*} (x)

is nonsingular and bounded.

In a linear case, proper assumptions are given in terms of the pair

(A, B)

1) eigenvalues of

A

have negative real parts; 2)

(A, B)

satisfies Kalman's controllability condition from [12] , i.e. a singularity of

a (x) \equiv B B^{*}

is permissible.

For the MDP analysis, we apply well known method employed for the central limit theorem (in short CLT) proof of a family

{(\frac{1}{\sqrt{t}} \int_{0}^{t} H (X_{s}) d s)}_{t \to \infty}

(see, e.g. Papanicolaou, Stroock and Varadhan [20] , Ethier and Kurtz [7] , Bhattacharya [3] , Pardoux and Veretennikov [21] , [22] and citations therein, see also Ch. 9, §3 in [16] ) based on a decomposition with corrector:

\frac{1}{\sqrt{t}} \int_{0}^{t} H (X_{s}) d s = \frac{1}{\sqrt{t}} {\underset{︸}{[U (x) - U (X_{t})]}}_{c o r r e c t o r} + \frac{1}{\sqrt{t}} {\underset{︸}{M_{t}}}_{m a r t i n g a l e},

where

U (x) = \int_{0}^{\infty} \int_{R^{d}} H (y) P_{x}^{(t)} (d y) d t,

P_{x}^{(t)}

is the transition probability kernel of

X

, and

M_{t}

is a continuous martingale with the variation process

〈 M 〉_{t}

. In the above mentioned papers, the corrector is negligible in a sense

\frac{1}{\sqrt{t}} [U (x) - U (X_{t})] - - - \to p r o b . t \to \infty 0

and the main contribution to a limit distribution brings

\frac{1}{\sqrt{t}} M_{t}

. It is well known (see, e.g. Ch. 5 in [16] ) the following implication: with nonnegative definite matrix

\frac{1}{t} 〈 M 〉_{t} - - - \to p r o b . t \to \infty Q \Rightarrow E e^{〈 〈 λ, \frac{1}{\sqrt{t}} M_{t} 〉 〉} - - - \to t \to \infty e^{- \frac{1}{2} 〈 〈 λ, Q λ 〉 〉}, \forall λ \in R^{q},

where

Q = \int_{0}^{\infty} \int_{R^{d}} [(P_{z}^{(t)} H) H^{*} (z) + (P_{z}^{(t)} H)^{*} H (z)] p (z) d z d t,

Summarizing these remarks, we may claim that the CLT holds provided that

U (x)

and

Q

exist and for any

ɛ > 0

\begin{matrix} {lim}_{t \to \infty} P (| U (x) - U (X_{t}) | > \sqrt{t} ɛ) = 0 \end{matrix}

\begin{matrix} {lim}_{t \to \infty} P (| 〈 M 〉_{t} - t Q | > t ɛ) = 0 . \end{matrix}

We develop the same method for MDP analysis. Replacing

\frac{1}{\sqrt{t}}

\frac{1}{t^{κ}}

, we keep the CLT framework with the same

U (x)

M_{t}

and

Q

, i.e.,

\frac{1}{t^{κ}} \int_{0}^{t} H (X_{s}) d s = \frac{1}{t^{κ}} {\underset{︸}{[U (x) - U (X_{t})]}}_{c o r r e c t o r} + \frac{1}{t^{κ}} {\underset{︸}{M_{t}}}_{m a r t i n g a l e}

and claim that (Theorem 2.1 ) the MDP holds, with the rate of speed

ϱ (t) = \frac{1}{t^{2 κ - 1}}

provided that

U (x)

and

Q

exist and for any

ɛ > 0

\begin{matrix} \begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (| U (x) - U (X_{t}) | > t^{κ} ɛ) = - \infty \\ {lim}_{t \to \infty} ϱ (t) log P (| 〈 M 〉_{t} - t Q | > t ɛ) = - \infty . \end{matrix} \end{matrix}

(1.4)

A choice of

ϱ (t)

is imposed by

\frac{1}{t^{κ}}

. As in the CLT proof, the corrector negligibility is required but exponentially fast with the rate of speed

ϱ (t)

. The main contribution in the MDP brings the family

{(\frac{1}{t^{κ}} M_{t})}_{t \to \infty}

Most probably, Dembo, [5] , was one of the first who introduced a condition of 1.4 (second) type. We found in Puhalskii, [25] (Theorem 2.3) and [24] , [26] that, in our setting with nonsingular (!) matrix

Q

, 1.4 provides MDP for the family

{(\frac{1}{t^{κ}} M_{t})}_{t \to \infty}

with the rate of speed

ϱ (t)

and the rate function

J (Y) = \frac{1}{2} ∥ Y ∥_{Q^{- 1}}^{2}, Y \in R^{q} .

We prove in Theorem 2.1 that the same statement remains valid for a singular

Q

too with the rate function

J (Y) = {\begin{matrix} \frac{1}{2} ∥ Y ∥_{Q^{\oplus}}^{2}, & Y = Q Q^{\oplus} Y, \\ \infty, & otherwise, \end{matrix}

where

Q^{\oplus}

is the Moore-Penrose pseudoinverse matrix (see, Albert, [1] ).

It would be noted that seeming simplicity of 1.4 is delusive with the exception of the eigenvalue gap case (in short EG, see Gong and Wu, [8] ) for

P_{x}^{(t)}

(a corresponding scenario can be found in [4] ). Unfortunately, the EG fails for diffusion processes. For instance, under

P_{x}^{(t)}

associated with Ornstein-Uhlenbeck's process

d X_{t} = - X_{t} d t + d W_{t}

having

(0, \frac{1}{2})

-Gaussian invariant measure

μ

, if EG were valid, then for bounded centered

H

| P_{x}^{(t)} H | \leq cons. e^{- λ t}, \forall t \geq 0, \exists λ > 0 .

However, direct computations show that for

H (x) = sign (x)

and sufficiently large

| x |

, we have

| P_{x}^{(t)} H | d t \leq υ (x) e^{- λ t}

where

υ (x)

is a positive function,

υ (x) < \infty

over

R^{d}

and

υ (x) \to \infty

with

| x | \to \infty

. The condition of this type: for any bounded and measurable

H

| P_{x}^{(t)} H - μ H | \leq υ (x) e^{- λ t}

describes the geometric ergodicity (see, Down, Meyn and Tweedie, [6] and citations therein). The geometric ergodicity is a helpful tool for the verification of

U (x)

and

Q

existence and even for the first part of 1.4 verification, although, a crude choice of

υ (x)

, say

υ (x) ≍ | x |^{m}, m > 2

, may to render this verification impossible (CLT analysis is not so sensitive to a choice of

υ

). The second part of 1.4 verification is very sensitive to properties of

U

, owing to

〈 M 〉_{t} = \int_{0}^{t} \nabla^{*} U (X_{s}) (a (X_{s}) \nabla U (X_{s}) d s

, so that, the geometric ergodicity framework is not a “foreground” tool. Following Pardoux and Veretennikov, [21] , we combine a property of

H

with a polynomial ergodicity

| P_{x}^{(t)} H - μ H | \leq \frac{υ (x)}{(1 + t)^{γ}}

γ > 1

with

H

-depending

υ

admitting an effective verification of 1.4 . In this connection, we mention here some result (see, Theorem A.1 ), in Appendix, interesting by itself, which is helpful in 1.4 verification. Let

X

be a diffusion process with the generator

L

and

V (x)

is Lyapunov's function belonging to the range of definition of

L

. Then,

N_{t} = V (X_{t}) - V (x_{0}) - \int_{0}^{t} L V (X_{s}) d s

is a continuous martingale and denote by

〈 N 〉_{t}

its variation process. Assume:

L V \leq - c V^{ℓ} + c, \exists q > 0 and 〈 N 〉_{t} \leq \int_{0}^{t} c (1 + V^{r} (X_{s})) d s, \exists r \leq ℓ .

Then, for any

ɛ > 0

and sufficiently large number

n

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (V (X_{t}) > t^{2 κ} ɛ) = - \infty, \end{matrix}

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (\int_{0}^{t} V^{ℓ} (X_{s}) d s > t n) = - \infty . \end{matrix}

Our method of the MDP analysis differs from Wu [29] - [33] where the Laplace transform technique dominates, or Guillin [9] , [10] based on discrete time approximation and Markov chains. In our approach, we deal with the above-mentioned Puhalskii's results obtained with the help of, so called, stochastic exponential as an alternative to Laplace's transform technique (see, e.g. [4] for more detailed explanation in the discrete time case).

The paper is organized as follows. In Section 2 , all notations are given and Theorem 2.1 , generalized Puhalskii's for singular

Q

, is formulated and proved.

In Section 3 , all results and examples are presented focusing on the existence and properties of the corrector and martingale variation process. The proofs are gathered in Section 4 . A simple example showing how the MDP may help in a statistical inference (for more information on statistical applications see, Inglot and Kallenberg, [11] ) is given in Section 5 . The technical tools are gathered in Appendix A .

2 Preliminaries

We fix the following notations and assumptions which are in force through the paper. The random process

X = (X_{t})_{t \geq 0}

is defined on some stochastic basis

(Ω, F, F = (F_{t})_{t \geq}, P)

satisfying the usual conditions.

∥ \cdot ∥

| \cdot |

, and

〈 〈 \cdot, \cdot 〉 〉

are Euclidean's and

L

norms respectively in

R^{d}

and the inner product.

^{*}

is transposition symbol.

a (z) : = σ σ^{*} (z) .

c, c, c \in R_{+}

\dots,

are generic constants.

P_{x}^{(t)} (d y)

is the transition probability kernel of

X

E_{x}

denotes the expectation relative to

P_{x}^{(t)} (d y)

μ (d z)

is the invariant measure.

L = \frac{1}{2} \sum_{i, j = 1}^{d} a_{i j} (z) \frac{\partial^{2}}{\partial z_{i} \partial z_{j}} + \sum_{i = 1}^{d} b_{i} (z) \frac{\partial}{\partial z_{i}}

is the generator of

X

(F_{t}^{X})_{t \geq 0}

is the filtration, with the general conditions, generated by

(X_{t})

〈 L 〉_{t}

is the variation process of a continuous martingale

(L_{t})_{t \geq 0} .

\nabla f (x)

is the gradient of

f (x)

(row vector).

ρ

is Euclidean's metric in

R^{d}

ϱ (t) = \frac{1}{t^{2 κ - 1}}

I

denotes the identical matrix of an appropriate size.

“

>

”, “

\geq

” denote also the standard inequalities for nonnegative definite matrices.

As was mentioned in Introduction, the existence of

\begin{matrix} Q & = & \int_{0}^{\infty} \int_{R^{d}} [(P_{z}^{(t)} H) H^{*} (z) + (P_{z}^{(t)} H)^{*} H (z)] d t p (z) d z, \end{matrix}

(2.1)

\begin{matrix} U (x) & = & \int_{0}^{\infty} \int_{R^{d}} H (y) P_{x}^{(t)} (d y) d t \end{matrix}

(2.2)

is required. We emphasize that

M_{t} = U (X_{t}) - U (x) + \int_{0}^{t} H (X_{s}) d s

is the martingale relative to

(F_{t}^{X})_{t \geq 0}

The theorem below is a “master-key” for MDP analysis.

Theorem 2.1. For any

x \in R^{d}

and any

ɛ > 0

, assume (i)

{lim}_{t \to \infty} ϱ (t) log P (| U (x) - U (X_{t}) | > t^{κ} ɛ) = - \infty

(ii)

{lim}_{t \to \infty} ϱ (t) log P (| 〈 M 〉_{t} - t Q | > t ɛ) = - \infty

Then, the family

(S_{t}^{κ})_{t \to \infty}

obeys the MDP in

(R^{q}, ρ)

with the rate of speed

ϱ (t)

and the rate function

\begin{matrix} J (Y) = {\begin{matrix} \frac{1}{2} ∥ Y ∥_{Q^{\oplus}}^{2}, & Y = Q Q^{\oplus} Y, \\ \infty, & otherwise, \end{matrix} \end{matrix}

(2.3)

where

Q^{\oplus}

is the Moore-Penrose pseudoinverse matrix (see, Albert, [1] ).

Proof. From the definition of

M_{t}

, it follows that

S_{t}^{κ} = \frac{1}{t^{κ}} [U (x) - U (X_{t})] + \frac{1}{t^{κ}} M_{t} .

(i) provides the negligibility of

{(\frac{1}{t^{κ}} [U (x) - U (X_{t})])}_{t \to \infty}

ϱ

-MDP scale.

(ii) provides

ϱ

-MDP, , under positive definite matrix, with the rate function

J (Y) = \frac{1}{2} ∥ Y ∥_{Q^{- 1}}^{2}

for the family

{(\frac{1}{t^{κ}} M_{t})}_{t \to \infty}

(due to result similar to Puhalskii, [25] (Theorem 2.3) and [26] ).

Q

is nonnegative definite only, the above result is no longer valid. This remark necessitates to turn to the general approach in large deviation analysis adapted to our setting. The family

{(\frac{1}{t^{κ}} M_{t})}_{t \to \infty}

is said to obey the large deviation principle (in our terminology: MDP) with the rate of speed

ϱ (t)

and some (good) rate function

J (Y), Y \in R^{q}

, provided that this family is

ϱ

-exponentially tight in

(R^{q}, ρ)

\begin{matrix} {lim}_{K \to \infty} l i m_{t \to \infty} ϱ (t) log P (| \frac{1}{t^{κ}} M_{t} | > K) = - \infty \end{matrix}

(2.4)

and obeys

(ϱ, J)

-local large deviation principle with the rate function

J (Y)

: for any

Y \in R^{q}

\begin{matrix} \begin{matrix} l i m_{δ \to 0} l i m_{t \to \infty} ϱ (t) log P (| \frac{1}{t^{κ}} M_{t} - Y | \leq δ) \leq - J (Y) \\ l i m_{δ \to 0} l i m_{t \to \infty} ϱ (t) log P (| \frac{1}{t^{κ}} M_{t} - Y | \leq δ) \geq - J (Y) . \end{matrix} \end{matrix}

(2.5)

A direct verification of 2.4 and 2.5 would be difficult. So, it is reasonable to verify 2.4 by applying the following regularization procedure. We introduce a new family

{(\frac{1}{t^{κ}} M_{t}^{γ})}_{t \to \infty}

with

M_{t}^{γ} = M_{t} + \sqrt{γ} d W_{t}^{'},

where

γ

is a positive number and

W_{t}^{'} (\in R^{q})

is a standard Wiener process independent of

M_{t}

. The random process

M_{t}^{γ}

is continuous martingale with

〈 M^{γ} 〉_{t} = 〈 M 〉_{t} + γ I t,

where

I = I_{q \times q}

. For the family

{(\frac{1}{t^{κ}} M_{t}^{γ})}_{t \to \infty}

, (ii) reads as:

{lim}_{t \to \infty} ϱ (t) log P (| \frac{1}{t} 〈 M^{δ} 〉_{t} - Q_{γ} | > ɛ) = - \infty,

where

Q_{γ} = Q + γ I

. Since

Q_{γ}

is the nonsingular matrix, the family

{(\frac{1}{t^{κ}} M_{t}^{γ})}_{t \to \infty}

obeys

(ϱ, J_{γ})

-MDP, where

J_{γ} (Y) = \frac{1}{2} ∥ Y ∥_{Q_{γ}^{- 1}}^{2}

Now, we apply the basic Puhalskii theorem from [23] which, being adapted to our case, states that the family

{(\frac{1}{t^{κ}} M_{t}^{γ})}_{t \to \infty}

ϱ (t)

-exponentially tight, in

(R^{q}, ρ)

\begin{matrix} l i m_{K \to \infty} l i m_{t \to \infty} ϱ (t) log P (| \frac{1}{t^{κ}} M_{t}^{γ} | > K) = - \infty, \end{matrix}

(2.6)

and obeys

(ϱ, J_{γ})

-local deviation principle:

\begin{matrix} \begin{matrix} l i m_{δ \to 0} l i m_{t \to \infty} ϱ (t) log P (| \frac{1}{t^{κ}} M_{t}^{γ} - Y | \leq δ) \leq - J_{γ} (Y) \\ l i m_{δ \to 0} l i m_{t \to \infty} ϱ (t) log P (| \frac{1}{t^{κ}} M_{t}^{γ} - Y | \leq δ) \geq - J_{γ} (Y) . \end{matrix} \end{matrix}

(2.7)

Obviously, 2.6 and 2.7 imply 2.4 and 2.5 provided that

\begin{matrix} {lim}_{δ \to 0} l i m_{t \to \infty} ϱ (t) P (| \frac{\sqrt{γ}}{t^{κ}} W_{t}^{'} | \geq η) = - \infty, \forall η > 0 \end{matrix}

(2.8)

and

\begin{matrix} {lim}_{γ \to 0} J_{γ} (V) = {\begin{matrix} \frac{1}{2} ∥ Y ∥_{Q^{\oplus}}^{2}, & Q^{\oplus} Q Y = Y \\ \infty, & otherwise . \end{matrix} \end{matrix}

(2.9)

2.8 holds true, since the family

\frac{\sqrt{γ}}{t^{κ}} W_{t}^{'}

obeys the

ϱ

-MDP with the rate function

\frac{1}{2 γ} ∥ Y ∥^{2}

, so that,

l i m_{t \to \infty} \frac{1}{t^{2 κ - 1}} P (∥ \frac{\sqrt{γ}}{t^{κ}} W_{t}^{'} ∥ \geq η) \leq - {inf}_{{Y : ∥ Y ∥ \geq \frac{η}{\sqrt{γ}}}} \frac{1}{2} ∥ Y ∥^{2} = - \frac{η^{2}}{2 γ} - - - \to γ \to 0 - \infty .

2.9 is verified with an utilization of the pseudoinverse matrix properties. Let

T

be an orthogonal matrix transforming

Q

to the diagonal form:

diag (Q) = T^{*} Q T .

Due to

2 J_{γ} (Y) = Y^{*} {[γ I + Q]}^{- 1} Y = Y^{*} T {[γ I + diag (Q)]}^{- 1} T^{*} Y,

for

Y = Q^{\oplus} Q Y

we have (recall that

Q^{\oplus} Q Q^{\oplus} = Q^{\oplus}

, see [1] )

\begin{matrix} 2 J_{γ} (Y) & = Y^{*} Q^{\oplus} Q T {[γ I + diag (Q)]}^{- 1} T^{*} Y \end{matrix}

\begin{matrix} = Y^{*} Q^{\oplus} T T^{*} Q T {[γ I + diag (Q)]}^{- 1} T^{*} Y \end{matrix}

\begin{matrix} = Y^{*} Q^{\oplus} T diag (Q) [γ I + diag (Q))^{- 1} T^{*} Y \end{matrix}

\begin{matrix} - - - \to γ \to 0 Y^{*} Q^{\oplus} T diag (Q) {diag}^{\oplus} (Q) T^{*} Y \end{matrix}

\begin{matrix} = Y^{*} Q^{\oplus} T diag (Q) T^{*} T {diag}^{\oplus} (Q) T^{*} Y \end{matrix}

\begin{matrix} = Y^{*} Q^{\oplus} Q Q^{\oplus} Y = Y^{*} Q^{\oplus} Y = ∥ Y ∥_{Q^{\oplus}}^{2} = 2 J (Y) . \end{matrix}

For

Y \neq Q^{\oplus} Q Y

{lim}_{γ \to 0} J_{γ} (Y) = \infty

. □

3 Main results

3.1 Nonlinear model, I

X_{t}

solves 1.2 subject to

X_{0} = x

(A_{b})

b

is locally Lipschitz continuous; for some

α \geq 1

and

C > 0

there exists

r > 0

, depending on

α, C

, such that

〈 〈 z, b (z) 〉 〉 \leq - r ∥ z ∥^{1 + α}, ∥ z ∥ > C .

(A_{σ, a})

σ

is Lipschitz continuous; for some

Λ > λ > 0

λ I \leq a (z) \leq Λ I .

From Pardoux and Veretennikov [21] , it follows that, under

(A_{b})

and

(A_{σ, a})

, the diffusion process

X

is ergodic with the unique invariant measure

μ (d z)

possessing a density

p (z)

relative to

d z

. Moreover, for

α > 1

and any

β < 0

\int_{R^{d}} (1 + ∥ z ∥)^{α - 1 + β} p (z) d z < \infty .

(A_{H})

H

is measurable function,

\int_{R^{d}} H (z) p (z) d z \equiv 0

; for

α \geq 1

, sufficiently small

δ > 0

and any

β < 0 \land \frac{1}{2} (3 - α - δ)

∥ H (x) ∥ \leq c (1 + ∥ x ∥)^{α - 1 + β} .

Remark 1. Under

(A_{b})

(A_{σ, a})

and

(A_{H}))

, from Pardoux and Veretennikov, [21] Theorem 2, it follows that

U (x)

, given in 2.2 , is bounded and solves the Poisson equation

L U = - H

in the class of functions with Sobolev's partial second derivatives locally integrable in any power and a polynomial growth. With all this going on,

\begin{matrix} | \nabla U (x) | \leq c (1 + ∥ x ∥)^{(β + α - 1)^{+}} \end{matrix}

(3.1)

and, by embedding theorems [15] , all entries of

\nabla U

are continuous functions. So, the Krylov generalization of Itô's formula (see [13] ) is applicable to

U (X_{t})

\begin{matrix} U (X_{t}) = U (x) - \int_{0}^{t} H (X_{s}) d s + \int_{0}^{t} \nabla U (X_{s}) σ (X_{s}) d W_{s} . \end{matrix}

(3.2)

Theorem 3.1. Under

(A_{b})

(A_{σ, a})

and

(A_{H})

, the family

(S_{t}^{κ})_{t \geq 0}

obeys the MDP in

(R^{q}, ρ)

with the rate of speed

ϱ (t)

and the rate function given in 2.3 with

Q

defined in 2.1 .

3.2 Nonlinear model, II

Though Theorem 3.1 serves a wide class of bounded and unbounded functions

H

, it is far from to be universal especially for

α = 1

So, we fix the next set of stronger assumptions.

(A_{b, σ}^{'})

b (x)

and

σ (x)

are Lipschitz continuous; for any

x^{'}, x^{''} \in R^{d}

there exists a positive number

ν

such that

2 〈 〈 (x^{'} - x^{''}, b (x^{'}) - b (x^{''}) 〉 〉 + trace [σ (x^{'}) - σ (x^{''})] [σ (x^{'}) - σ (x^{''})]^{*} \leq - ν ∥ x^{'} - x^{''} ∥^{2} .

(A_{a}^{'})

λ I \leq a (z) \leq Λ I,

for some

Λ > λ > 0

(A_{H}^{'})

H (x)

is Lipschitz continuous function.

Theorem 3.2. Under

(A_{b, σ}^{'})

(A_{a}^{'})

and

(A_{H}^{'})

, the statement of Theorem 3.1 remains valid.

3.3 Linear model

The diffusion process

X_{t}

solves 1.3 ,

A = A_{d \times d}

B = B_{d \times d}

and

(W_{t})_{t \geq 0}

is a standard vector-valued Wiener process of the corresponding size.

For this setting,

(A_{b})

(A_{b, σ}^{'})

, and

(A_{a})

are too restrictive. We replace them by the following assumptions.

(A)

Eigenvalues of

A

have negative real parts.

(A_{B})

D : = B B^{*} + A^{*} B B^{*} A + \dots + (A^{*})^{d - 1} B B^{*} A^{d - 1}

is nonsingular matrix.

(A_{H}^{''})

Suppose either 1)

H

possesses continuous and bounded partial derivatives, 2)

H

is bounded,

B B^{*} > 0

Theorem 3.3. Under

(A)

(A_{B})

and

(A_{H}^{''})

, the family

(S_{t}^{κ})_{t \to \infty}

obeys the MDP in

(R^{d}, ρ)

with rate of speed

ϱ (t)

and the rate function given in 2.3 with

Q

defined in 2.1 .

The next result deals with quadratic function

H

. Under

(A)

and

(A_{B})

, the invariant measure

μ

is zero mean Gaussian with nonsingular covariance matrix

P

solving the Lyapunov equation

\begin{matrix} A^{*} P + P A + B B^{*} = 0 . \end{matrix}

(3.3)

We introduce also a positive definite matrix

Γ = Γ_{q \times q}

and a matrix

ϒ = ϒ_{q \times q}

solving the Lyapunov equation

A^{*} ϒ + A ϒ + Γ = 0 .

Theorem 3.4. Assume

(A)

and

B B^{*} > 0

and

H (x) = 〈 〈 x, Γ x 〉 〉 - trace (Γ^{1 / 2} P Γ^{1 / 2}) .

Then, the family

(S_{t}^{κ})_{t \to \infty}

obeys the MDP in

(R^{d}, ρ)

with rate of speed

ϱ (t)

and the rate function given in 2.3 with

Q = 4 trace (ϒ B P B^{*} ϒ) > 0 .

3.4 More examples

In this section, we give examples which are not explicitly compatible with Theorems 3.1 - 3.4 .

Example 3.1. Let

d = 1

H (x) = x^{3}

and

\begin{matrix} d X_{t} = - X_{t}^{3} d t + d W_{t} . \end{matrix}

(3.4)

Though

(A_{b})

holds with

α = 3

, Theorem 3.1 is not applicable since by

(A_{H})

only

H

with property

∥ H (x) ∥ \leq c (1 + ∥ x ∥)^{γ}

γ < 2

is admissible.

Nevertheless, the MDP holds and is trivially verified. Indeed, 3.4 is nothing but 3.2 with

U (x) \equiv x

. Hence,

\nabla U (x) = 1

and

Q = 1

Consequently, (ii) from Theorem 2.1 automatically holds.

(i) is reduced to

{lim}_{t \to \infty} ϱ (t) log P (X_{t}^{2} \geq t^{2 κ} ɛ) = - \infty

and is verified with the help of Theorem A.1 with

V (x) \equiv x^{2}

. Actually, by Itô's formula we find that

d V (X_{t}) = [- 2 V^{2} (X_{t}) + 1] d t + d N_{t},

where

N_{t} = \int_{0}^{t} 2 X_{s} d W_{s}

. Hence,

L V (x) \leq - V^{2} (x) + 1 and 〈 N 〉_{t} = \int_{0}^{t} 4 V (X_{s}) d s .

Example 3.2. Let

d = 1

and

d X_{t} = b (x_{t}) d t + d W_{t},

where

b (x)

is Lipschitz continuous and symmetric,

b (x) = - b (- x),

function (obviously

b (0) = 0

). Under

(A_{b, σ}^{'})

, providing

(A_{b})

X_{t}

is an ergodic diffusion process with the symmetric invariant density,

p (z) = p (- z)

. So, any bounded

H (x)

, with

H (x) = - H (- x)

, possesses 1.1 . We choose

H (x) = sign (x), letting sign (0) = 0 .

However, neither Theorem 3.3 nor Theorem 3.1 are compatible with the setting owing to

H (x)

does not satisfy neither

(A_{H}^{'})

nor

(A_{H})

. Nevertheless, we show that the standard MDP holds. A computational trick proposes to use a decomposition

H = H^{'} + H^{''}

for

H^{'} (x) = {\begin{matrix} e^{- x}, & x > 0 \\ 0, & x = 0 \\ - e^{x}, & x < 0 \end{matrix}

since

H^{'}

satisfies

(A_{H})

and

H^{''} (x) = {\begin{matrix} 1 - e^{- x}, & x \geq 0 \\ - 1 + e^{x}, & x < 0 \end{matrix}

satisfies

(A_{H}^{'})

. Then,

U^{'} (x)

and

\nabla U^{'} (x)

are well defined and both are bounded; at the same time

U^{''} (x)

and

\nabla U^{''} (x)

are also well defined and

\nabla U^{''} (x)

is bounded, i.e.

| U^{''} (x) | \leq c (1 + | x |)

Taking

U (x) = U^{'} (x) + U^{''} (x)

we get bounded

\nabla U (x) = \nabla U^{'} (x) + \nabla U^{''} (x)

and

U (x)

satisfying the linear growth condition. Moreover, due to

M_{t} = M_{t}^{'} + M_{t}^{''},

we have

M_{t} = \int_{0}^{t} \nabla U^{'} (X_{s}) d W_{s} + \int_{0}^{t} \nabla U^{''} (X_{s}) d W_{s} = \int_{0}^{t} \nabla U (X_{s}) d W_{s},

providing

〈 M 〉_{t} = \int_{0}^{t} {(\nabla U (X_{s}))}^{2} d s

with bounded

{(\nabla U (x))}^{2}

Now, (i) and (ii) from Theorem 2.1 are verified in a standard way with the help of Theorems A.1 , A.2 .

Example 3.3. (Linear version of Langevin model.) A nonlinear Langevin's model, including our linear one, is studied in Wu, [33] . The result from [33] seems not to be accomplished. At least, we could not adapt assumptions from there to verify the MDP for the following setting.

Let

X_{t} = (\begin{matrix} q_{t} \\ p_{t} \end{matrix}) \in R^{2 d}

with (

q_{t}, p_{t} \in R^{d}

) and

\begin{matrix} \begin{matrix} d q_{t} & = p_{t} d t \\ d p_{t} & = - Γ p_{t} d t - \nabla F (q_{t}) d t + σ d W_{t}, \end{matrix} \end{matrix}

(3.5)

where

\nabla F (q) = Λ q

and matrices

Λ

Γ

and

σ σ^{*}

are positive definite. We verify the MDP with the help of Theorem 3.3 .

It is expedient to write 3.5 to the form 1.3 with matrices (in a block form)

A = (\begin{matrix} 0 & I \\ - Λ & - Γ \end{matrix}) and B = (\begin{matrix} 0 & 0 \\ 0 & σ \end{matrix}) .

In accordance with Theorem 3.3 , we have to verify only two conditions:

1) eigenvalues of

A

have negative real parts, 2) the matrix

D

(see

(A_{B})

) is nonsingular.

1) fulfils since free of noise 3.5 :

\begin{matrix} {\dot{q}}_{t} & = p_{t} \end{matrix}

\begin{matrix} \dot{p_{t}} & = - Γ p_{t} - \nabla F (q_{t}) \end{matrix}

is asymptotically stable. Traditionally for the Langevin equation, this result is easily verified with the help of Lyapunov's function

V_{t} = \frac{1}{2} ∥ p_{t} ∥^{2} + F (q_{t})

and is omitted here.

2) holds since

D^{'} : = B B^{*} + A^{*} B B^{*} A (\leq D)

is nonsingular. Indeed,

D^{'} = (\begin{matrix} Λ σ^{*} σ Λ & Λ σ^{*} σ Γ \\ Γ σ^{*} σ Λ & Γ σ^{*} σ Γ + σ^{*} σ \end{matrix}),

that is, with a vector

v = (\begin{matrix} v_{1} \\ v_{2} \end{matrix}) \neq 0,

we have

\begin{matrix} 〈 〈 v, D^{*} D v 〉 〉 & = 〈 〈 v_{1}, Λ σ^{*} σ Λ v_{1} 〉 〉 + 〈 〈 v_{2}, Γ σ^{*} σ Γ v_{2} 〉 〉 + 2 〈 〈 v_{1}, Λ σ^{*} σ Γ v_{2} 〉 〉 \end{matrix}

\begin{matrix} + 〈 〈 v_{2}, σ^{*} σ v_{2} 〉 〉 . \end{matrix}

By virtue of the well known inequality

\begin{matrix} 2 〈 〈 z_{1}, z_{2} 〉 〉 \geq - 〈 〈 z_{1}, z_{1} 〉 〉 - 〈 〈 z_{2}, z_{2} 〉 〉, \end{matrix}

(3.6)

we get

〈 〈 v_{1}, Λ σ^{*} σ Λ v_{1} 〉 〉 + 〈 〈 v_{2}, Γ σ^{*} σ Γ v_{2} 〉 〉 + 2 〈 〈 v_{1}, Λ σ^{*} σ Γ v_{2} 〉 〉 \geq 0 .

Consequently, under

v_{2} \neq 0

, we have

〈 〈 v, (D^{'})^{*} D^{'} v 〉 〉 \geq 〈 〈 v_{2}, σ^{*} σ v_{2} 〉 〉 > 0 .

Even though

v_{2} = 0

, and so

v_{1} \neq 0

, we also have

〈 〈 v, (D^{'})^{*} D^{'} v 〉 〉 = 〈 〈 v_{1}, Λ σ^{*} σ v_{1} Λ 〉 〉 > 0

Thus, under

(A_{H}^{''})

, the MDP holds.

Example 3.4. (MDP for a smooth component of diffusion process.) Let

X_{t}^{(1)}

be the first component of a diffusion process

X_{t}

with entries

X_{t}^{(i)}

i = 1, \dots, d

\begin{matrix} \begin{matrix} {\dot{X}}_{t}^{(1)} & = X_{t}^{(2)} \\ {\dot{X}}_{t}^{(i)} & = X_{t}^{(i + 1)}, i = 2, \dots, d - 1 \\ d X_{t}^{(d)} & = - \sum_{i = 1}^{d} a_{i} X_{t}^{(d - i)} d t + b d W_{t}, \end{matrix} \end{matrix}

(3.7)

where

a_{1}, a_{2}, \dots, a_{d}

and

b

are positive numbers and

W_{t}

is a Wiener process.

As in the previous example, we rewrite 3.7 to the form of 1.3 with

A = (\begin{matrix} A_{11} & A_{12} \\ A_{21} & A_{22} \end{matrix}) and B = (\begin{matrix} B_{11} & B_{12} \\ B_{21} & B_{22} \end{matrix}),

where

A_{11} = {(\begin{matrix} 0 & 1 & 0 & . . . & 0 & 0 \\ 0 & 0 & 1 & 0 & . . . & 0 \\ . . . & . . . & . . . & . . . & . . . & . . . \\ 0 & 0 & 0 & 0 & 0 & 1 \end{matrix})}_{(d - 1) \times (d - 1)},

A_{12} = 0

A_{22} = - a_{d}

A_{21} = {(\begin{matrix} - a_{1} & - a_{2} & . . . & - a_{d - 1} \end{matrix})}_{1 \times (d - 1)}

and, analogously,

B_{11} = 0_{(d - 1) \times (d - 1)}

B_{12} = 0

B_{22} = b

B_{21} = 0_{1 \times (d - 1)}

We verify the MDP with the help of Theorem 3.3 . In order to guarantee

(A_{H})

, suffice it to assume that roots of the polynomial

φ (z) = z^{d} + a_{1} z^{d - 1} + \dots + a_{d - 1} z + a_{d}

have negative real parts owing to the noise free version of 3.7 is nothing but the differential equation

x_{t}^{(d)} + \sum_{i = 1}^{d - 1} a_{i} x_{t}^{(d - i)} + a_{d} x_{t} = 0 .

Notice that

(A_{B})

is fulfilled too since

D^{'} = B B^{*} + A^{*} B B^{*} A (\leq D)

is a nonsingular matrix. Actually,

D^{'} = b^{2} (\begin{matrix} A_{21}^{*} A_{21} & A_{21}^{*} A_{22} \\ A_{21} A_{22} & A_{22}^{2} + 1 \end{matrix})

and so, we have

\begin{matrix} 〈 〈 v, D^{*} D v 〉 〉 = b^{2} [v_{1}^{2} ∥ A_{21} ∥^{2} + (A_{22}^{2} + 1) ∥ v_{2} ∥^{2} + 2 v_{1} A_{22} 〈 〈 v_{2}, A_{21} 〉 〉] . \end{matrix}

Taking

v = (\begin{matrix} v_{1} \\ v_{2} \end{matrix}) \neq 0,

where

v_{1}

is a number and

v_{2}

is a vector of the size

d - 1

, for

v_{2} = 0

, and then

v_{1} \neq 0

, we have

〈 〈 v, D^{*} D v 〉 〉 > 0

. Even though

v_{2} \neq 0

, the use of 3.6

provides

〈 〈 v, D^{*} D v 〉 〉 \geq b^{2} A_{22}^{2} ∥ v_{2} ∥^{2} > 0 .

In order to establish the MDP for the family

{(\frac{1}{t^{κ}} \int_{0}^{t} H (X_{s}^{(1)}) d s)}_{t \to \infty},

we redefine the function

H

as:

H (x^{(1)}) \equiv H (x^{(1)}, x^{(2)}, \dots, x^{(d)})

and assume

H

satisfies

(A_{H}^{''})

. Then, the family

{(\frac{1}{t^{κ}} \int_{0}^{t} H (X_{s}) d s)}_{t \to \infty}

obeys the MDP with the rate of speed

ϱ (t)

and the rate function

J (Y) = J (Y^{(1)}, \dots, Y^{(d)})

of the standard form 2.3 .

Now, the desired MDP holds by Varadhan's contraction principle, [27] , with the same rate of speed and the rate function

j (y) = {inf}_{{Y^{(2)}, \dots, Y^{(d)}} \in R^{d - 1}} J (y, Y^{(2)}, \dots, Y^{(d)}) .

Example 3.5. Let

X_{t} (\in R)

be Gaussian diffusion with

d X_{t} = - X_{t} d t + d W_{t}

and

H (x) = x^{2} sign (x)

. This function satisfies 1.1 and, at the same time, is not compatible with Theorems 3.1 - 3.4 . So, we suppose to embed this setting to a new one with a vector function

H (x)

with entries:

H_{1} (x) = \frac{1}{2} sign (x) and H_{2} (x) = x^{2} sign (x) - \frac{1}{2} sign (x),

which is MDP verifiable. Applying arguments from the proof of the Theorem 3.3 , one can show the existence of

U_{1} (x)

with bounded

\nabla U_{1} (x)

such that

\begin{matrix} U_{1} (X_{t}) & = U_{1} (x) - \int_{0}^{t} H_{1} (X_{s}) d s + M_{t}^{(1)} \end{matrix}

\begin{matrix} 〈 M^{(1)} 〉_{t} & = \int_{0}^{t} (\nabla U_{1} (X_{s}))^{2} d s . \end{matrix}

Now, we establish similar property of

H_{2} (x)

. By the Krylov-Itô formula (see [13] ), we find that

d H (X_{t}) = - H_{2} (X_{t}) d t + | X_{t} | d W_{t} .

Consequently,

U_{2} (x) \equiv H (x)

and

〈 M^{(2)} 〉_{t} = \int_{0}^{2} X_{s}^{2} d s

Now, we may verify (i), (ii) from Theorem 2.1 .

(i): Since

\nabla U_{1}

is bounded,

U_{1}

satisfies the linear growth condition.

Thus,

| U_{1} | \leq c (1 + | U_{2} |) = c (1 + | H (x) |) \leq c (1 + x^{2}) .

Hence, (i) is reduced to

{lim}_{t \to \infty} ϱ (t) log P (X_{t}^{2} \geq t^{κ} ɛ) = - \infty .

The latter holds owing to

X_{t}^{2}

possesses an exponential moment:

E e^{λ X_{t}^{2}} < \infty

uniformly in

t

over

R_{+}

and sufficiently small

λ

and, therefore, the Chernoff inequality is effective. Write

\begin{matrix} \frac{1}{t^{2 κ - 1}} log P (X_{t}^{2} > t^{κ} ɛ) & \leq \frac{1}{t^{2 κ - 1}} log (e^{- λ t^{κ} ɛ + log E e^{λ X_{t}^{2}}}) \end{matrix}

\begin{matrix} \leq - λ t^{1 - κ} ɛ + \frac{log E e^{λ X_{t}^{2}}}{t^{2 κ - 1}} - - - \to t \to \infty - \infty . \end{matrix}

Notice that

| U_{2} (x) | = x^{2}

, so that, the (i) verification is the same as for

U_{1}

(ii): The martingale

M_{t}

is vector-valued process with two entries

M_{t}^{(1)}

and

M_{t}^{(2)}

. Hence, its variation process is a matrix

〈 M 〉_{t} = (\begin{matrix} 〈 M^{(1)} 〉_{t} & 〈 M^{(1)}, M^{(2)} 〉_{t} \\ 〈 M^{(1)}, M^{(2)} 〉_{t} & 〈 M^{(2)} 〉_{t}, \end{matrix})

so that, the entries of

Q

are defined in the following way:

\begin{matrix} Q_{11} & = \int_{R} {(\nabla U_{1} (z))}^{2} p (z) d z, \end{matrix}

\begin{matrix} Q_{22} & = \int_{R} z^{2} p (z) d z, \end{matrix}

\begin{matrix} Q_{12} & = \int_{R} \nabla U_{1} (z) | z | p (z) d z . \end{matrix}

Thus, (ii) is reduced to

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (| \int_{0}^{t} h (X_{s}) d s | \geq t ɛ) = - \infty, \end{matrix}

(3.8)

where

h (x)

is continuous function satisfying 1.1 and either 1)

| h (x) | \leq c | x |

, 2)

h (x) = x^{2} - \frac{1}{2}

In 1), we apply

h (x) = h_{l}^{'} (x) + h_{l}^{''} (x),

borrowed from 4.1 , and verify versions of 3.8 with

h_{l}^{'}

and

h_{l}^{''}

separately.

h_{l}^{'}

-version holds owing to by Theorem 3.1 (

α = 1

)

{(\frac{1}{t^{κ}} \int_{0}^{t} h_{l}^{'} (X_{s}) d s)}_{t \to \infty}

obeys

ϱ

-MDP with a nondegenerate rate function and

κ < 1

h_{l}^{''}

-version holds owing to

| h_{l}^{''} (x) | \leq I (| x | > l) | x | \leq \frac{x^{2}}{l}

and for sufficiently large

l

{lim}_{t \to \infty} ϱ (t) log P (\int_{0}^{t} X_{s}^{2} d s > t l ɛ) = - \infty

verified with the help of Theorem A.1 for

V (x) = x^{2}

In 2), by Theorem 3.4 ,

{(\frac{1}{t^{κ}} \int_{0}^{t} [X_{s}^{2} - \frac{1}{2}] d s)}_{t \to \infty}

obeys

ϱ

-MDP with a nondegenerate rate function. So, it remains to recall that

κ < 1

Thus,

ϱ

-MDP for new family holds true with the rate function

J (Y)

Y \in R^{2}

, defined in 2.3 . Hence, the original family possesses the MDP with the quadratic rate function

j (y) = {inf}_{{Y_{1}, Y_{2} : Y_{1} + Y_{2} = y}} J (Y) .

4 Proof of Theorems from Section 3

4.1 The proof of Theorem 3.1

Denote by

M_{t} = \int_{0}^{t} \nabla U (X_{s}) σ (X_{s}) d W_{s}

the martingale from 3.2 having

〈 M 〉_{t} = \int_{0}^{t} \nabla U (X_{s}) a (X_{s}) \nabla^{*} U (X_{s}) d s

We shall verify (i) and (ii) from Theorem 2.1 .

(i) holds since, by Remark 1 ,

U

is bounded.

(ii) is verified in a few steps.

Step 1:

Q

identification. We show that

\int_{R^{d}} \nabla U (z) a (z) \nabla^{*} U (z) p (z) d z = Q .

This fact is well known and is given here for a reader convenience only. Notice that, by 2.2 ,

Q = E [H (X_{0}^{μ}) U^{*} (X_{0}^{μ}) + U (X_{0}^{μ}) H^{*} (X_{0}^{μ})],

where

X_{t}^{μ}

the stationary version of

X_{t}

, that is, the version solving 1.2 subject to

X_{0}^{μ}

the random vector, independent of

W_{t}

, with the distribution provided by the invariant measure

μ

. Hence, suffice it to show that

\begin{matrix} E [\nabla U (X_{0}^{μ}) a (X_{0}^{μ}) \nabla^{*} U (X^{)} μ_{0}] = E [H (X_{0}^{μ}) U^{*} (X_{0}^{μ}) + U (X_{0}^{μ}) H^{*} (X_{0}^{μ})] . \end{matrix}

(4.1)

We verify 4.1 with the help of Itô's formula

U (X_{t}^{μ}) U^{*} (X_{t}^{μ}) = U (X_{0}^{μ}) U^{*} (X_{0}^{μ}) - \int_{0}^{t} [H (X_{s}^{μ}) U^{*} (X_{s}^{μ}) + U (X_{s}^{μ}) H^{*} (X_{0}^{μ})] d s + \int_{0}^{t} [U (X_{s}^{μ}) d M_{s}^{*} + d M_{s} U^{*} (X_{0}^{μ})] + \int_{0}^{t} \nabla (X_{s}^{μ}) a (X_{s}^{ν}) \nabla^{*} (X_{s}^{μ}) d s

by taking the expectation.

Step 2. Preliminaries.Set

H (x) = \nabla U (x) a (x) \nabla^{*} (x) - Q

and let

h (x)

denotes any entry of

H (x)

. For (ii) to be valid suffice it to show that

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (| \int_{0}^{t} h (X_{s}) d s | > t ɛ) = - \infty . \end{matrix}

(4.2)

Recall

\int_{R^{d}} h (z) p (z) d z = 0

. By 3.1 ,

| h (x) | \leq c (1 + ∥ x ∥)^{2 (β + α - 1)^{+}} .

We consider separately two cases provided by a special choice of

β < 0 \land \frac{1}{2} (3 - α - δ) for sufficiently small δ > 0 .

(see,

(A_{H})

(α = 1) :

| h (x) |

is bounded;

(α > 1) :

| h (x) | \leq c (1 + ∥ x ∥)^{1 + α - δ}

1 + α - δ \geq 2 .

Step 3.

α = 1

For sufficiently large number

l

, set

h_{l}^{'} (x) = {\begin{matrix} h (x) & ∥ x ∥ \leq l, \\ v_{l} (x) & l < ∥ x ∥ \leq l + 1 \\ 0 & ∥ x ∥ > l + 1, \end{matrix}

where

v_{l} (x)

is bounded continuous function such that

h_{l}^{'} (x)

is continuous function with

\int_{R^{d}} h_{l}^{'} (z) p (z) d z = 0

. In contrast to

h

, the function

h_{l}^{'}

decreases fast to zero with

∥ x ∥ \to \infty

, so that, a negative constant

β^{'}

can be chosen such that

| h_{l}^{'} (x) | \leq c (1 + | x |)^{β^{'} + α - 1} \equiv c (1 + | x |)^{β^{'}} .

In accordance with this property,

u (x) = - \int_{0}^{\infty} E_{x} h^{'} (X_{t}) d t

solves the Poisson equation

L u = - h_{l}^{'}

and is bounded jointly with

\nabla u (x)

(see, Remark 1 ). Hence,

u (X_{t}) = u (x) - \int_{0}^{t} h^{'} (X_{s}) d s + m_{t}

with the martingale

m_{t} = \int_{0}^{t} \nabla u (X_{s}) σ (X_{s}) d W_{s}

having

〈 m 〉_{t} = \int_{0}^{t} \nabla u (X_{s}) a (X_{s}) \nabla^{*} u (X_{s}) d s .

The negligibility of

\frac{u (x) - u (X_{t})}{t}

ϱ

-MDP scale is provided by the boundedness of

u (x)

. The same type negligibility of

\frac{1}{t} m_{t}

is provided by the boundedness of

\nabla u^{*} (x) a (x) \nabla u (x)

, due to Theorem A.2 .

Consequently, a version of 4.2 with

h_{l}^{'}

holds true.

Set

h_{l}^{''} = h - h_{l}^{'}

. Since

h

is bounded,

| h_{l}^{''} (x) | \leq c I (∥ x ∥ > l) \leq \frac{c}{l^{2}} ∥ x ∥^{2} .

Consequently a version of 4.2 with

h_{l}^{''}

is reduced to

{lim}_{t \to \infty} ϱ (t) log P (\int_{0}^{t} ∥ X_{s} ∥^{2} d s > t (l^{2} ɛ)) = - \infty

and is verified with the help of Theorem A.2 for

V (x) = ∥ x ∥^{2}

owing to

L V (x) \leq - c V (x) + c, and 〈 N_{t} 〉 \leq \int_{0}^{t} c (1 + V (X_{s})) d s

are fulfilled under

(A_{b})

and

(A_{σ, a})

(a verification of these facts is accomplished with the help of Itô's formula).

Step 4.

α > 1

We apply again the decomposition

h = h_{l}^{'} + h_{l}^{''}

. With chosen

l

| h_{l}^{'} |

is decreasing fast to zero, with

∥ x ∥ \to \infty

, and is bounded by

c (1 + l)^{1 + α - δ}

. So, the version of 4.2 with

h_{l}^{'}

is verified as in the case “

α = 1

”.

Notice that

| h_{l}^{''} (x) | \leq c (1 + ∥ x ∥)^{1 + α - δ)} I (∥ x ∥ > l) \leq \frac{c}{l^{δ}} (1 + ∥ x ∥)^{1 + α} \leq \frac{c}{l^{δ}} (1 + V (x)),

where

V (x) = \frac{∥ x ∥^{4 + 2 α}}{1 + ∥ x ∥^{3 + α}} .

Hence, the version of 4.2 with

h_{l}^{''}

is reduced to

{lim}_{t \to \infty} ϱ (t) log P (\int_{0}^{t} V (X_{s}) d s > t l^{δ}) = - \infty .

To this end, we apply Theorem A.1 .

First, taking into account that

∥ x ∥^{3 + α} = {(∥ x ∥^{2})}^{\frac{3 + α}{2}}

∥ x ∥^{4 + 2 α} = (∥ x ∥^{2})^{2 + α}

and

\frac{3 + α}{2} > 2

, by the Itô formula we find that

\begin{matrix} d ∥ X_{t} ∥^{2} & = [2 〈 〈 X_{t}, b (X_{t} 〉 〉 + trace (a (X_{t}))] d t + 2 〈 〈 X_{t}, σ (X_{t}) d W_{t} 〉 〉, \end{matrix}

\begin{matrix} d ∥ X_{t} ∥^{3 + α} & = (\frac{3 + α}{2} - 1) {(∥ X_{t} ∥^{2})}^{\frac{3 + α}{2} - 1} {2 〈 〈 X_{t}, b (X_{t}) 〉 〉 + trace (a (X_{t}))} \end{matrix}

\begin{matrix} + 2 [\frac{3 + α}{2} - 1] [\frac{3 + α}{2} - 2] {(∥ X_{t} ∥^{2})}^{\frac{3 + α}{2} - 2} 〈 〈 X_{t}, a (X_{t}) X_{t} 〉 〉] d t \end{matrix}

\begin{matrix} + (\frac{3 + α}{2} - 1) (∥ X_{t} ∥^{2})^{\frac{3 + α}{2} - 1} 2 〈 〈 X_{t}, σ (X_{t}) d W_{t} 〉 〉, \end{matrix}

\begin{matrix} d ∥ X_{t} ∥^{4 + 2 α} & = (1 + α) {(∥ X_{t} ∥^{2})}^{1 + α} {2 〈 〈 X_{t}, b (X_{t}) 〉 〉 + trace (a (X_{t}))} \end{matrix}

\begin{matrix} + 2 [1 + α] α {(∥ X_{t} ∥^{2})}^{α} 〈 〈 X_{t}, a (X_{t}) X_{t} 〉 〉] d t \end{matrix}

\begin{matrix} + (1 + α) {(∥ X_{t} ∥^{2})}^{1 + α} 2 〈 〈 X_{t}, σ (X_{t}) d W_{t} 〉 〉, \end{matrix}

\begin{matrix} d \frac{1}{1 + ∥ X_{t} ∥^{3 + α}} & = - \frac{d ∥ X_{t} ∥^{3 + α}}{(1 + ∥ X_{t} ∥^{3 + α})^{2}} \end{matrix}

\begin{matrix} + \frac{2 (1 + α) ∥ X_{t} ∥^{1 + α}}{(1 + ∥ X_{t} ∥^{3 + α})^{3}} 〈 〈 X_{t}, a (X_{t}) X_{t} 〉 〉 d t, \end{matrix}

\begin{matrix} d V (X_{t}) & = \frac{d ∥ X_{t} ∥^{4 + 2 α}}{1 + ∥ X_{t} ∥^{3 + α}} + ∥ X_{t} ∥^{4 + 2 α} d \frac{1}{1 + ∥ X_{t} ∥^{3 + α}} \end{matrix}

\begin{matrix} + \frac{2 {(1 + α)}^{2} ∥ X_{t} ∥^{3 (1 + α)} 〈 〈 X_{t}, a (X_{t}) X_{t} 〉 〉}{(1 + ∥ X_{t} ∥^{3 + α})^{2}} d t . \end{matrix}

Thus, we have

d V (X_{t}) = L V (X_{t}) d t + d N_{t}

, where

\begin{matrix} L V (x) & = \frac{1}{1 + ∥ x ∥^{3 + α}} [(1 + α) ∥ x ∥^{2 (1 + α)} {2 〈 〈 x, b (x) x 〉 〉 + trace (a (x))} \end{matrix}

\begin{matrix} + 2 α (1 + α) ∥ x ∥^{2 α} 〈 〈 x, a (x) x 〉 〉] \end{matrix}

\begin{matrix} - \frac{∥ x ∥^{4 + 2 α}}{(1 + ∥ X_{t} ∥^{3 + α})^{2}} [\frac{1}{2} (1 + α) ∥ x ∥^{1 + α} {2 〈 〈 x, b (x) x 〉 〉 + trace (a (x))} \end{matrix}

\begin{matrix} + \frac{1}{2} (1 + α) (α - 1) ∥ x ∥^{α - 1} 〈 〈 x, a (x) x 〉 〉] \end{matrix}

\begin{matrix} + \frac{∥ x ∥^{4 + 2 α}}{(1 + ∥ X_{t} ∥^{3 + α})^{3}} 2 (1 + α) ∥ x ∥^{1 + α} 〈 〈 x, a (x) x 〉 〉 \end{matrix}

\begin{matrix} \leq (1 + α) 〈 〈 x, b (x) x 〉 〉 [\frac{2 ∥ x ∥^{2 (1 + α)}}{1 + ∥ x ∥^{3 + α}} - \frac{∥ x ∥^{4 + 2 α}}{(1 + ∥ x ∥^{3 + α})^{2}}] + o (∥ x ∥^{2 α}) \end{matrix}

\begin{matrix} = (1 + α) 〈 〈 x, b (x) x 〉 〉 \frac{2 ∥ x ∥^{2 (1 + α)} + 2 ∥ x ∥^{5 + 3 α} - ∥ x ∥^{4 + 2 α}}{(1 + ∥ x ∥^{3 + α})^{2}} + o (∥ x ∥^{2 α}) \end{matrix}

\begin{matrix} \leq - c ∥ x ∥^{2 α} + c \leq - c V^{\frac{2 α}{1 + α}} (x) + c \end{matrix}

and

N_{t} = \int_{0}^{t} 〈 〈 X_{s}, σ (X_{s}) d W_{s} 〉 〉 [\frac{2 (1 + α) ∥ X_{s} ∥^{2 (1 + α)}}{1 + ∥ X_{s} ∥^{3 + α}} - \frac{(1 + α) ∥ X_{s} ∥^{5 + 3 α}}{(1 + ∥ X_{s} ∥^{3 + α})^{2}}],

that is,

〈 N 〉_{t} \leq \int_{0}^{t} (c ∥ X_{s} ∥^{2 α} + c) d s \leq \int_{0}^{t} (c V^{\frac{2 α}{1 + α}} (X_{s}) + c) d s .

Thus, the assumptions of Theorem A.1 are fulfilled and, thereby, for sufficiently large

l

, we have

{lim}_{t \to \infty} ϱ (t) log P (\int_{0}^{t} V^{\frac{2 α}{1 + α}} (X_{s}) d s > t l^{δ}) = - \infty

and, it is left to notice that

\frac{2 α}{1 + α} > 1

for

α > 1

. □

4.2 The proof of Theorem 3.2

(A_{b, σ}^{'})

\begin{matrix} 〈 〈 x, b (x) 〉 〉 & = 〈 〈 x, (b (x) - b (0) 〉 〉 + 〈 〈 x, b (0) 〉 〉 \end{matrix}

\begin{matrix} \leq - 〈 〈 x, B_{0} x 〉 〉 + ∥ b (0) ∥ ∥ x ∥ \end{matrix}

\begin{matrix} \leq - ν ∥ x ∥^{2} + ∥ b (0) ∥ ∥ x ∥ \end{matrix}

that is, there exists

r > 0

such that

〈 〈 x, b (x) 〉 〉 \leq - r ∥ x ∥^{2} .

Hence,

(A_{b, σ}^{'}) \Rightarrow (A_{b}) (α = 1)

. However since, by

(A_{H}^{'})

∥ H (z) ∥ \leq c (1 + ∥ z ∥)

is admissible, Theorem 2 from Pardoux and Veretennikov, [21] , is no longer applicable.

At the same time, Theorem 1 from [21] states that

U

from 2.2 solves the Poisson equation

L U (z) = - H (z)

and satisfies the following properties: for some

m > 2

∥ U (x) ∥ \leq c (1 + ∥ x ∥^{m}) and ∥ \nabla U (x) ∥ \leq c (1 + ∥ x ∥^{m}) .

Nevertheless, regardless of that,

(A_{H}^{'})

provides

\begin{matrix} ∥ U (x) ∥ \leq c (1 + ∥ x ∥) and ∥ \nabla U (x) ∥ \leq c . \end{matrix}

(4.3)

Actually, let

X_{t}^{x}

denotes the solution of 1.2 subject to

X_{0} = x

. Since for any

x^{'}

and

x^{''}

, we have

U (x^{'}) - U (x^{''}) = \int_{0}^{\infty} E [H (X_{t}^{x^{'}}) - H (X_{t}^{x^{''}})] d t

, by

(A_{H}^{'})

, we have (

L

is the Lipschitz constant for

H

)

\begin{matrix} | U (x^{'}) - U (x^{''}) | & \leq L \int_{0}^{\infty} | E [X_{t}^{x^{'}} - X_{t}^{x^{''}}] | d t \end{matrix}

\begin{matrix} \leq L \int_{0}^{\infty} {(E ∥ X_{t}^{x^{'}} - X_{t}^{x^{''}} ∥^{2})}^{1 / 2} d t, \end{matrix}

where

d [X_{t}^{x^{'}} - X_{t}^{x^{''}}] = [b (X_{t}^{x^{'}}) - b (X_{t}^{x^{''}})] d t + [σ (X_{t}^{x^{'}}) - σ (X_{t}^{x^{''}})] d W_{t} .

With the help of Itô's formula, we find that

\begin{matrix} d ∥ X_{t}^{x^{'}} - X_{t}^{x^{''}} ∥_{t}^{2} & = 2 〈 〈 (X_{t}^{^{'}} - X_{t}^{x^{''}}), [b (X_{t}^{x^{'}}) - b (X_{t}^{x^{''}})] 〉 〉 d t \end{matrix}

\begin{matrix} + 2 〈 〈 (X_{t}^{^{'}} - X_{t}^{x^{''}}), [σ (X_{t}^{x^{'}}) - σ (X_{t}^{x^{''}})] d W_{t} 〉 〉 \end{matrix}

\begin{matrix} + trace [σ (X_{t}^{x^{'}}) - σ (X_{t}^{x^{''}})] [σ (X_{t}^{x^{'}}) - σ (X_{t}^{x^{''}})]^{*} . \end{matrix}

Hence,

v_{t} = E ∥ X_{t}^{x^{'}} - X_{t}^{x^{''}} ∥^{2}

is differentiable relative to

d t

and

{\dot{v}}_{t} = 2 E [〈 〈 [X_{t}^{x^{'}} - X_{t}^{x^{''}}], [b (X_{t}^{x^{'}} - b (X_{t}^{x^{''}})] 〉 〉 + trace [σ (X_{t}^{x^{'}}) - σ (X_{t}^{x^{''}})] [σ (X_{t}^{x^{'}}) - σ (X_{t}^{x^{''}})]^{*}] .

Then, by

(A_{b, σ}^{'})

, we have

{\dot{v}}_{t} \leq - ν v_{t},

i.e.,

v_{t} \leq ∥ x^{'} - x^{''} ∥^{2} e^{- t ν} .

The latter implies the Lipschitz continuity of

U

and, in turn, 4.3 .

We proceed the proof with the verification of (i) and (ii) from Theorem 2.1 .

(i): Due to 4.3 , suffice it show that

{lim}_{t \to \infty} ϱ (t) log P (∥ X_{t} ∥^{2} > ɛ t^{2 κ}) = - \infty

what is verified with the help of Theorem A.1 for

V (x) = ∥ x ∥^{2}

. With the help of Itô's formula, one can find that

L V (x) = 2 〈 〈 x, b (x) 〉 〉 + trace a (x) and N_{t} = \int_{0}^{t} 2 〈 〈 X_{s}, σ (X_{s}) d W_{s} 〉 〉

and next that

L V (x) \leq - c V (x) + c, 〈 N 〉_{t} \leq \int_{0}^{t} c V (X_{s}) d s

(ii): It is verified similarly to 4.2 for

α = 1

□

4.3 The proof of Theorem 3.3

Under

(A)

(A_{B})

, the Pardoux-Veretennikov concept is no longer valid. Nevertheless,

(A)

and

(A_{B})

provide the ergodicity of

X = (X_{t})_{t \geq 0}

with the unique zero mean Gaussian invariant measure characterized by a nonsingular covariance matrix

P

solving Lyapunov's equation, see 3.3 .

We prove the theorem in a few steps. Step 1. Invariant and transition densities.For

X_{0} = x

, the diffusion process

X_{t}

is Gaussian with the expectation

E X_{t} = e^{A t} x

and the covariance matrix

cov (X_{t}, X_{t}) = \int_{0}^{t} e^{(t - s) A^{*}} B B^{*} e^{(t - s) A} d s = : P_{t}

solving the differential equation

\begin{matrix} {\dot{P}}_{t} = A^{*} P_{t} + P_{t} A + B B^{*} \end{matrix}

(4.4)

subject to

P_{0} = 0

. It is well known, and is readily verified that, under

(A)

and

(A_{B})

, we have

P_{t} > 0

over

t > 0

and

{lim}_{t \to \infty} P_{t} = P (> 0) .

If in addition

B B^{*} > 0

, then, for

t

in a vicinity of zero,

\begin{matrix} | P_{t}^{- 1 / 2} | \leq \frac{c}{\sqrt{t}} . \end{matrix}

(4.5)

Since

P, P_{t} > 0

, the invariant density

p (y)

and the density of

P_{x}^{(t)} (d y)

relative to

d y

are defined as:

p (y) = \frac{1}{(2 π det P)^{d / 2}} e^{- \frac{1}{2} ∥ y ∥_{P^{- 1}}^{2}} p (x, t, y) = \frac{1}{(2 π det P_{t})^{d / 2}} exp (- \frac{1}{2} ∥ y - e^{t A} x ∥_{P_{t}^{- 1}}^{2})

Step 2.

U

existence.We prove that

U (x)

from 2.2 is well defined over

R^{d}

by showing

\begin{matrix} \int_{0}^{\infty} | E_{x} H (X_{t}) | d t < \infty . \end{matrix}

(4.6)

Assume

(A_{H}^{''})_{1)}

. Let

X_{t}^{μ}

X_{t}^{x}

denote the stationary version of

X_{t}

and

X_{t}

with

X_{0} = x

respectively. By 1.1 and the Lipschitz property of

H

(with the Lipschitz constant

L

), it holds

| E_{x} (X_{t}) | = | E [H (X_{t}^{x}) - H (X_{t}^{μ}) | \leq L E | X_{t}^{x} - X_{t}^{μ} |,

where, by 1.2 ,

\frac{d}{d t} [X_{t}^{x} - X_{t}^{μ}] = A [X_{t}^{x} - X_{t}^{μ}],

i.e.,

[X_{t}^{x} - X_{t}^{μ}] = e^{t A} [x - X_{0}^{μ}] .

Hence and by

(A)

, there exists a positive constant

λ

such that

| X_{t}^{x} - X_{t}^{μ} | \leq e^{- t λ} c (1 + ∥ x ∥ + ∥ X_{0}^{μ} ∥) .

The random vector

X_{0}^{μ}

is Gaussian, so that,

E ∥ X_{0}^{μ} ∥ = c .

Thus,

| E_{x} (X_{t}) | \leq e^{- t λ} c (1 + ∥ x ∥)

and 4.6 holds true.

Assume

(A_{H}^{''})_{2)}

. We may adapt the results of Meyn and Tweedie, [19] (see also Mattingly and Stuart, [17] and Mattingly Stuart and Higham, [18] ) for getting 4.6 .

However, taking into account the explicit formulae for

p (y)

and

p (x, t, y)

, the direct proof of 4.6 is given.

For a definiteness, let

| H | \leq K

. We apply an obvious inequality

| E_{x} H (X_{t}) | \leq K \int_{R^{d}} | p (x, t, y) - p (y) | d y (\leq 2 K) .

A changing of variables:

z = (y - e^{t A} z) P_{t}^{- 1 / 2}

and the identity

\begin{matrix} \frac{p (P_{t}^{1 / 2} z + e^{t A} x)}{p (x, t, P_{t}^{1 / 2} z + e^{t A} x)} = \sqrt{\frac{det P_{t}}{det P}} \end{matrix}

\begin{matrix} \times exp (- \frac{1}{2} [〈 〈 z, (P_{t} P^{- 1} - I) z 〉 〉 + 2 〈 〈 P^{- 1 / 2} z, e^{t A} x 〉 〉 + ∥ e^{t A} x ∥_{P_{t}^{- 1}}^{2}]) \end{matrix}

provide

\begin{matrix} \int_{R^{d}} | p (x, t, y) - p (y) | d y = \int_{R^{d}} | 1 - \frac{p (y)}{p (x, t, y)} | p (x, t, y) d y \end{matrix}

\begin{matrix} = \int_{R^{d}} | 1 - \frac{p (P_{t}^{1 / 2} z + e^{t A} x)}{p (x, t, P_{t}^{1 / 2} z + e^{t A} x)} | p (z) d z \end{matrix}

\begin{matrix} \leq | \sqrt{\frac{det P_{t}}{det P}} - 1 | + \sqrt{\frac{det P_{t}}{det P}} \int_{R^{d}} | exp (- \frac{1}{2} [〈 〈 z, (P_{t} P^{- 1} - I) z 〉 〉 \end{matrix}

\begin{matrix} + 2 〈 〈 P^{- 1 / 2} z, e^{t A} x 〉 〉 + ∥ e^{t A} x ∥_{P_{t}^{- 1}}^{2}]) - 1 | p (z) d z . \end{matrix}

Due to

(A)

e^{t A} x

converges to zero in

t \to \infty

exponentially fast in a sense that

| e^{t A} x | \leq c e^{- t λ} ∥ x ∥

for some generic

λ > 0

. Moreover,

| P_{t} P^{- 1} - I | \leq c e^{- t λ}

, owing to

P - P_{t}

solves the differential equation

{\dot{△}}_{t} = A^{*} △_{t} + △_{t} A

subject to

△_{0} = P

(see, 3.3 and 4.4 ) . The above-mentioned convergence implies also

| {(\frac{det P_{t}}{det P})}^{1 / 2} - 1 | \leq c e^{- t λ} .

Thus, there exists an appropriate positive continuous function

υ (x) (< \infty)

over

R^{d}

such that for

t \geq t_{0} > 0

\int_{R^{d}} | p (x, t, y) - p (y) | d y \leq c e^{- t λ} [1 + \int_{R^{d}} {∥ z ∥^{2} + ∥ x ∥^{2}} e^{c e^{- t λ} [∥ z ∥^{2} + ∥ x ∥^{2}]} p (z) d z] \leq c e^{- t λ} (1 + υ (∥ x ∥))

and, in turn, 4.6 holds true, owing to

\int_{0}^{\infty} | E_{x} H (X_{s}) | d s \leq 2 K t_{0} + \int_{t_{0}}^{\infty} | E_{x} H (X_{s}) | d s \leq 2 K + \frac{K c}{λ} c (1 + υ (∥ x ∥)) .

Step 3.

\nabla U

existence.Assume

(A_{H}^{''})_{1)}

and notice that

\begin{matrix} \int_{0}^{\infty} | \int_{R^{d}} \nabla_{x} H (P_{t} z + e^{t A} x) \frac{1}{(2 π)^{d / 2}} e^{- \frac{1}{2} ∥ z ∥^{2}} d z | d t \leq const. \end{matrix}

(4.7)

Since

U (x) = - \int_{0}^{\infty} \int_{R^{d}} H (P_{t} z + e^{t A} x) \frac{1}{(2 π)^{d / 2}} e^{- \frac{1}{2} ∥ z ∥^{2}} d z d t,

by virtue of of 4.7 we have

\nabla U (x) = - \int_{0}^{\infty} \int_{R^{d}} \nabla_{x} H (P_{t} z + e^{t A} x) \frac{1}{(2 π)^{d / 2}} e^{- \frac{1}{2} ∥ z ∥^{2}} d z d t .

In particular,

\nabla U

is bounded.

Assume

(A_{H}^{''})_{2)}

. Now, we prove that

\begin{matrix} \int_{0}^{\infty} | \int_{R^{d}} H (y) \nabla_{x} p (x, t, y) d y | d t \leq const. \end{matrix}

(4.8)

The use of

\nabla_{x} p (x, t, y) = - p (x, t, y) {(y - e^{t A} x)}^{*} P_{t}^{- 1} e^{t A}, t > 0,

provides

\begin{matrix} \int_{R^{d}} H (y) \nabla_{x} p (x, t, y) d y & = - E H (X_{t}^{x}) (X_{t}^{x} - E X_{t}^{x})^{*} P_{t}^{- 1} e^{t A} . \end{matrix}

\begin{matrix} = E [H (X_{t}^{x}) - E H (X_{t}^{x})] [X_{t}^{x} - E X_{t}^{x}]^{*} P_{t}^{- 1} e^{t A} . \end{matrix}

Consequently, taking into account the boundedness of

H

and 4.5 , by Cauchy-Schwarz's inequality we get (with a generic positive constant

λ

| \int_{R^{d}} H (y) \nabla_{x} p (x, t, y) d y | \leq c {(E ∥ X_{t}^{x} - E X_{t}^{x} ∥^{2})}^{1 / 2} | P_{t}^{- 1} | | e^{t} A | \leq c e^{- t λ} {(trace (P_{t}))}^{1 / 2} | P_{t}^{- 1} | \leq c \frac{e^{- t λ}}{\sqrt{t}}

Then, 4.8 holds and

\nabla U

is bounded. Step 4.

M_{t}

existence.Since an applicability of Itô's (Krylov-Itô's) formula to

U (X_{t})

is questionable, we show that

(M_{t}, F_{t}^{X})_{t \geq 0}

, with

M_{t} = U (X_{t}) - U (x) + \int_{0}^{t} H (X_{s}) d s,

is the continuous martingale,

\begin{matrix} 〈 M 〉_{t} = \int_{0}^{t} \nabla^{*} U (X_{s}) B B^{*} \nabla U (X_{s}) d s \end{matrix}

(4.9)

and

E ∥ M_{t} ∥^{2} < \infty

over

t \in R_{+}

; the latter is provided by the boundedness of

\nabla U

The use of a homogeneity in

t

of the Markov process

X_{t}

enables to claim that

U (X_{t})

admits the following presentation a.s.,

U (X_{t}) = \int_{t}^{\infty} E_{X_{t}} U (X_{s}) d s = \int_{t}^{\infty} E (H (X_{s}) | F_{t}^{X}) d s .

Then for any

t^{'} < t

, we have

\begin{matrix} M_{t} - M_{t^{'}} & = \int_{t^{'}}^{\infty} E (H (X_{s}) | F_{t}^{X}) d s - \int_{t^{'}}^{\infty} E (H (X_{s}) | F_{t^{'}}^{X}) d s \end{matrix}

\begin{matrix} + \int_{t^{'}}^{t} E (H (X_{s}) | F_{t}^{X}) d s - \int_{t^{'}}^{t} H (X_{s}) d s a.s. \end{matrix}

and the martingale property,

E (M_{t} | F_{t^{'}}^{X}) = M_{t^{'}}

a.s., becomes obvious.

Now, we establish 4.9 with the help of well known fact: for any

t > 0

〈 M 〉_{t}

coincides with the limit, in probability, in

k \to \infty

\sum_{1 \leq j \leq k} (M_{t_{j}^{k}} - M_{t_{j - 1}^{k}}) {(M_{t_{j}^{k}} - M_{t_{j - 1}^{k}})}^{*},

where

0 \equiv t_{0}^{k} < t_{1}^{k} < \dots < t_{t_{k}}^{k} \equiv t

is a condensing sequence of time values. We recall only that

M_{t_{j}^{k}} - M_{t_{j - 1}^{k}} = U (X_{t_{j}^{k}}) - U (X_{t_{j - 1}^{k}}) + O (t_{j}^{k} - t_{j - 1}^{k})

and

U (X_{t_{j}^{k}}) - U (X_{t_{j - 1}^{k}}) = \nabla^{*} U (X_{t_{j - 1}^{k}}) B [W_{t_{j}^{k}} - W_{t_{j - 1}^{k}}] + O (t_{j}^{k} - t_{j - 1}^{k}) .

Step 5. (i) verification.Due to the linear growth condition of

∥ U (x) ∥

, suffice it to show that

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (V (X_{t}) > t^{2} ɛ) = - \infty . \end{matrix}

(4.10)

for

V (x) = 〈 〈 x, Γ x 〉 〉

with an appropriate positive definite matrix

Γ

. In view of

(A)

, it is convenient to choose

Γ

solving the Lyapunov equation

A^{*} Γ + Γ A + I = 0 .

The function

V (x)

belongs to the range of definition for

L

with

\begin{matrix} L V (x) & = 〈 〈 x, (A^{*} Γ + Γ A) x 〉 〉 + trace (B Γ B^{*}) \end{matrix}

\begin{matrix} = - ∥ x ∥^{2} + trace (B Γ B^{*}) \leq - c V (x) + c \end{matrix}

while

V (X_{t}) - V (x) - \int_{0}^{t} L V (X_{s}) d s = \int_{0}^{t} 2 〈 〈 X_{s}, Γ B d W_{s} 〉 〉 = : N_{t}

is the martingale (relative to

(F_{t}^{X})

) with

〈 N 〉_{t} = \int_{0}^{t} 4 〈 〈 X_{s}, Γ^{2} X_{s} 〉 〉 d t \leq \int_{0}^{t} c V (X_{s}) d s .

Now, 4.10 is provided by Corollary 1 to Theorem A.1 .

Step 6. (ii) verification.Since

\nabla U

is bounded and continuous, (ii) holds true if

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (| \int_{0}^{t} h (X_{s}) d s | > t ɛ) = - \infty \end{matrix}

for any bounded and continuous

h : R^{d} \Rightarrow R

with

\int_{R^{d}} h (z) p (z) = 0

Assume for a moment that

h

satisfy

(A_{H}^{''})_{1)}

from Theorem 3.3 . Then, the function

u (x) = - \int_{0}^{\infty} E h (X_{t}) d t

is well defined and

(u (X_{t}), F_{t}^{X})_{t \geq 0}

is the semimartingale:

u (X_{t}) = u (x) - \int_{0}^{t} h (X_{s}) d s + m_{t},

where

(m_{t}, F_{t}^{X})_{t \geq 0}

is the continuous martingale with

〈 m 〉_{t} = \int_{0}^{t} \nabla^{*} u (X_{s}) B B^{*} \nabla U (X_{s}) d s

and

\nabla u (x)

is bounded and continuous.

Hence, suffice it to verify 4.11 with

\int_{0}^{t} h (X_{s}) d s

replaced by

u (X_{t})

and

m_{t}

separately.

First of all notice that the version of 4.11 with

m_{t}

is valid due to Theorem A.2 owing to

〈 m 〉_{t} \leq K t

, where

K \geq \nabla^{*} u (X_{t}) B B^{*} \nabla u (X_{t})

t

over

R_{+}

. Further, because of

\nabla u

is bounded and, then,

u

satisfies the linear growth condition, the version of 4.11 with

u (X_{t})

is reduced to 4.10 .

h

does not satisfy

(A_{H}^{''})_{1)}

, we apply the decomposition

h = h^{'} + h^{''}

borrowed from the proof of Theorem 3.1 ,

α = 1

. Then, the version of 4.11 with

h^{''}

is reduced to: for sufficiently large

l

{lim}_{t \to \infty} ϱ (t) log P (\int_{0}^{t} 〈 〈 X_{s}, Γ X_{s} 〉 〉 d s > t (l^{2} ɛ)) = - \infty,

and is verified with the help of Theorem A.1 for

V (x) = 〈 〈 x, Γ x 〉 〉

The verification of 4.11 with

h^{'}

differs from the corresponding part of proof for Theorem 3.1 ,

α = 1

. Let

l

, involved in the definition of

h^{'}

, and

ɛ > 0

be chosen.

Since

h^{'}

is compactly supported, there exists a polynomial

h_{ɛ}

such that

\begin{matrix} c_{ɛ} : = {sup}_{x} | h^{'} (x) - h_{ɛ} (x) | = o (ɛ) \end{matrix}

\begin{matrix} d_{ɛ} : = \int_{R^{d}} h_{ɛ} (z) p (z) d z = o (ɛ) . \end{matrix}

Because of

{\hat{h}}_{ɛ} = h_{ɛ} - d_{ɛ}

satisfies 1.1 and

(A_{H}^{''})_{1)}

, the validity of 4.11 with

{\hat{h}}_{ɛ}

is obvious. So, it is left to recall only that

{sup}_{x} | h^{'} (x) - {\hat{h}}_{ɛ} (x) | = o (ɛ)

. □

4.4 The proof of Theorem 3.4

Obviously,

H (x)

satisfies 1.1 .

We shall verify (i), (ii) from Theorem 2.1 . By virtue of 2.2 , the quadratic form of

H

is inherited by

U

. We examine the following

U (x) = 〈 〈 x, ϒ x 〉 〉 - υ

with a positive definite matrix

ϒ

and positive number

υ

. By Itô's formula we find that

d U (X_{t}) = {\underset{︸}{[〈 〈 X_{t}, [ϒ A + A^{*} ϒ] X_{t} 〉 〉 + trace (B^{*} ϒ B)]}}_{candidate to be - H (X_{t})} d t + {\underset{︸}{2 〈 〈 X_{t}, ϒ B d W_{t} 〉 〉}}_{= M_{t}} .

The realization of this project requires for

ϒ

to be a solution of Lyapunov's equation

ϒ A + A^{*} ϒ + Γ = 0

what, in particular, provides

trace (B^{*} ϒ B) = trace (Γ^{1 / 2} P Γ^{1 / 2}),

where

P

is the covariance of the invariant measure. With chosen

ϒ

, set

D = ϒ B P B^{*} ϒ

and notice that

〈 M 〉_{t} = \int_{0}^{t} 4 〈 〈 X_{s}, D X_{s} 〉 〉 d s

(i) is reduced to

{lim}_{t \to \infty} ϱ (t) log P (〈 〈 X_{t}, Γ X_{t} 〉 〉 > t^{κ} ɛ) = - \infty

which holds since for positive and sufficiently small

λ

the moment generating function

log E e^{λ 〈 〈 X_{t}, ϒ X_{t} 〉 〉}

is bounded over

t \in R_{+}

and, then, Chernoff 's inequality provides

\frac{1}{t^{2 κ - 1}} log P (〈 〈 X_{t}, ϒ X_{t} 〉 〉 > t^{κ} ɛ) \leq - λ t^{1 - κ} ɛ + \frac{log E e^{λ 〈 〈 X_{t}, ϒ X_{t} 〉 〉}}{t^{2 κ - 1}} - - - \to t \to \infty - \infty .

(ii) is valid if

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (| \int_{0}^{t} [〈 〈 X_{s}, ϒ B B^{*} ϒ X_{s} 〉 〉 - trace (D)] d s | > t ɛ) = - \infty . \end{matrix}

(4.11)

Let us denote

γ = ϒ B B^{*} ϒ

and

h (x) = 〈 〈 x, γ x 〉 〉 - trace (D) .

We repeat the previous arguments to find

u (x) = 〈 〈 x, r x 〉 〉 - r

with a positive definite matrix

r

and positive number

r

such that

m_{t} = u (X_{t}) - u (x) + \int_{0}^{t} h (X_{s}) d s

is a continuous martingale with

〈 m 〉_{t} = \int_{0}^{t} 〈 〈 X_{s}, q X_{s} 〉 〉 d s,

where

q

is a positive definite matrix. Now, we may replace 4.11 by

\begin{matrix} (1) {lim}_{t \to \infty} ϱ (t) log P (〈 〈 X_{t}, γ X_{t} 〉 〉 > t ɛ) = - \infty \end{matrix}

\begin{matrix} (2) {lim}_{t \to \infty} ϱ (t) log P (| m_{t} | > t ɛ) = - \infty . \end{matrix}

(1) is verified similarly to (i). (2) is verified with the help of Theorem A.2 by showing

{lim}_{t \to \infty} ϱ (t) log P (〈 m 〉_{t} > t n) = - \infty

for sufficiently large

n

what is nothing but

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (\int_{0}^{t} 〈 〈 X_{s}, q X_{s} 〉 〉 d s > t n) = - \infty . \end{matrix}

(4.12)

A version of 4.12 with

q

replaced by any positive definite matrix

G

provides 4.12 too. For computational convenience, we take

G

solving Lyapunov's equation

A^{*} G + G A + I = 0 .

The function

V (x) = 〈 〈 x, G x 〉 〉

belongs to the range of definition of

L

with

L V (x) = - 2 ∥ x ∥^{2} + trace (B P B^{*}) \leq - c V (x) + c

and

N_{t} = V (X_{t}) - V (x) - \int_{0}^{t} L V (X_{s}) d s = \int_{0}^{t} 2 〈 〈 X_{s}, B d W_{s} 〉 〉,

so that,

〈 N 〉_{t} \leq \int_{0}^{t} c V (X_{s}) d s .

Thus, the proof is completed by applying Theorem A.1 . □

5 Example of statistical application

Let

X_{t} (\in R)

be a diffusion process:

d X_{t} = - θ X_{t} d t + d W_{t},

subject to a fixed

X_{0}

. The parameter

θ \in (0, \infty)

is unknown and is evaluated with help of well known estimate

{\hat{θ}}_{t} = \frac{\int_{0}^{t} X_{s} d X_{s}}{\int_{0}^{t} X_{s}^{2} d s}, t > 0 .

It is well known that the CLT holds for the family

{(\sqrt{t} (θ - {\hat{θ}}_{t}))}_{t \to \infty}

with a limit:

zero mean Gaussian random variable with the variance

2 θ

In this section, we show that

θ - {\hat{θ}}_{t}

possesses an asymptotic (in

t \to \infty

) in the MDP scale,

\frac{1}{2} < κ < 1

, that is, the family

{(t^{1 - κ} (θ - {\hat{θ}}_{t}))}_{t \to \infty}

obeys

(ϱ, J)

-MDP with

J (Y) = \frac{Y^{2}}{4 θ}

. The use of some details from the proof of Theorem 3.4 enables to claim that

{(\frac{1}{t} \int_{0}^{t} [X_{s}^{2} - \frac{1}{2 θ}] d s)}_{t \to \infty}

is negligible in

ϱ

-MDP scale. Therefore, the family

{(t^{1 - κ} (θ - {\hat{θ}}_{t}))}_{t \to \infty}

shares the MDP with

{(\frac{1}{t^{κ}} \int_{0}^{t} 2 θ X_{s} d W_{s})}_{t \to \infty} .

Further, the announced MDP hold if (ii) from Theorem 2.1 is valid:

\begin{matrix} {lim}_{t \to \infty} log ϱ (t) P (| \int_{0}^{t} [4 θ^{2} X_{s}^{2} - Q] d s | > t ɛ) = - \infty . \end{matrix}

(5.1)

Obviously,

Q = 2 θ

and the validity of 5.1 is verified with the help of arguments used in the proof of Theorem 3.4 .

In particular, this MDP and the contraction Varadhan's principle, for sufficiently large

t

provide

\frac{1}{t^{2 κ - 1}} log P (t^{1 - κ} | θ - {\hat{θ}}_{t} | > δ) ≍ - \frac{δ^{2}}{4 θ} .

A Exponential negligibility of functionals and martingales

Let

X_{t}

be a diffusion process defined in 1.2 with

X_{0} = x

Assume

V (x) : R^{d} \to R_{+}

, with

{lim}_{∥ x ∥ \to \infty} V (x) = \infty

, belongs to the range of definition of

L

. Introduce a martingale relative to

(F)_{t \geq 0}

\begin{matrix} N_{t} = V (X_{t}) - V (x) - \int_{0}^{t} L V (X_{s}) d s . \end{matrix}

(A.1)

Theorem A.1. Assume 1)

L V \leq - c V^{ℓ} + c

\exists ℓ > 0

〈 N 〉_{t} \leq \int_{0}^{t} c (1 + V^{r} (X_{s})) d s

\exists r \leq ℓ

Then, for any

ɛ > 0

and sufficiently large number

n

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (V (X_{t}) > t^{2 κ} ɛ) = - \infty, \end{matrix}

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (\int_{0}^{t} V^{ℓ} (X_{s}) d s > t n) = - \infty . \end{matrix}

over

x \in R^{d}

Corollary 1.

{lim}_{t \to \infty} ϱ (t) log P (V (X_{t}) > t^{2} ɛ) = - \infty,

since

t^{2} > t^{2 κ}, t > 1 .

Remark 2. The statements of Theorem A.1 remain valid if constants

c, c, c

, involved in 1) and 2) depend on

ɛ

Theorem A.2. Let

M_{t} (\in R, M_{0} = 0)

be a continuous martingale.

Then, for any

ɛ > 0

{lim}_{t \to \infty} ϱ (t) log P (| M_{t} | > t ɛ) = - \infty

provided that, under sufficiently large number

n

depending on

ɛ

{lim}_{t \to \infty} ϱ (t) log P (〈 M 〉_{t} > t n) = - \infty .

The proof of Theorem A.1 . With

λ \in R

, and the (continuous) martingale

N_{t}

from A.1 , we introduce a positive random process

z_{t} (λ) = e^{λ N_{t} - 0.5 λ^{2} 〈 N 〉_{t}} .

It is well known and easily verified with the help of Itô's formula that

(z_{t} (λ), F_{t}^{X})_{t \geq 0}

is a positive local martingale. Moreover, by Problem 1.4.4, [16] , it is a supermartingale too. We shall use the supermartingale property:

E z_{t} (λ) \leq E z_{0} (λ) \equiv 1

over

t \in R_{+}

. Denote by

A_{1} = {V (X_{t}) > t^{2 κ} ɛ} and A_{2} = {\int_{0}^{t} V^{r} (X_{s}) d s > t n} .

The use of

E z_{t} (λ) \leq 1

provides

\begin{matrix} 1 \geq E I_{A_{i}} z_{t} (λ), i = 1, 2 \end{matrix}

(A.2)

Notice that A.2 remains valid with

z_{t} (λ)

replaced by its lower bound on

A_{i}

. We proceed the proof by finding appropriate deterministic (!) lower bounds. Write

λ N_{t} - 0.5 λ^{2} 〈 N 〉_{t} = λ (V (X_{t}) - V (x) - \int_{0}^{t} L V (X_{s}) d s) - 0.5 λ^{2} 〈 N 〉_{t} .

Thence, in view of 1) and 2), with

λ > 0

we get

\begin{matrix} λ N_{t} - 0.5 λ^{2} 〈 N 〉_{t} & \geq λ (V (X_{t}) - V (x) + \int_{0}^{t} [c V^{ℓ} (X_{s}) - c] d s) \end{matrix}

\begin{matrix} - 0.5 λ^{2} \int_{0}^{t} c (1 + V^{r} (X_{s})) d s . \end{matrix}

Taking into account

1 + V^{r} (X_{s}) \leq 2 + V^{ℓ} (X_{s})

, provided by

r \leq ℓ

, and choosing

λ^{\circ} = {argmax}_{λ > 0} [c λ - 0.5 c λ^{2}] = \frac{c}{c}

, we get

\begin{matrix} λ^{\circ} N_{t} - 0.5 (λ^{\circ})^{2} 〈 N 〉_{t} & \geq \frac{c}{c} [V (X_{t}) - V (x)] - t \frac{c}{c} [c + c] + \frac{c^{2}}{2 c} \int_{0}^{t} V^{ℓ} (X_{s}) d s \end{matrix}

\begin{matrix} \geq {\begin{matrix} \frac{c}{c} [t^{2 κ} ɛ - V (x)] - t \frac{c}{c} [c + c], & over A_{1}, \\ - \frac{c}{c} V (x) - t \frac{c}{c} [c + c] + \frac{c^{2}}{2 c} t n, & over A_{2} . \end{matrix} \end{matrix}

These lower bounds jointly with A.2 provide

\begin{matrix} ϱ (t) log P (A_{1}) \leq - \frac{c}{c} [t ɛ - \frac{V (x)}{t^{2 κ - 1}}] + t^{2 (1 - κ)} \frac{c}{c} [c + c], & over A_{1} \\ ϱ (t) log P (A_{2}) \leq \frac{c}{c} \frac{V (x)}{t^{2 κ - 1}} + t^{2 (1 - κ)} \frac{c}{c} [c + c] - t^{2 (1 - κ)} \frac{c^{2}}{2 c} n, & over A_{2} \end{matrix}} - - - \to t \to \infty - \infty .

□ The proof of Theorem A.2 . Notice that only

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (| M_{t} | > t ɛ, 〈 M 〉_{t} \leq t n) = - \infty \end{matrix}

(A.3)

is required to be proved. Moreover, it suffices to prove only

\begin{matrix} {lim}_{t \to \infty} ϱ (t) log P (M_{t} > t ɛ, 〈 M 〉_{t} \leq t n) = - \infty \end{matrix}

(A.4)

owing to a version with

- M_{t}

is verified similarly and both “

\pm M_{t}

” provide A.3 .

For A.4 verification, we use the inequality from A.2 with

λ > 0

and

N_{t}

〈 N 〉_{t}

replaced by

M_{t}

〈 M 〉_{t}

respectively and

A_{i}

replaced by

A = {M_{t} > t ɛ, 〈 M 〉_{t} \leq t n}

and notice that

log z_{t} (λ) = λ M_{t} - 0.5 λ^{2} 〈 M 〉_{t} {\underset{︸}{\geq}}_{over A} λ t ɛ - 0.5 λ^{2} t n \geq {min}_{λ > 0} (λ t ɛ - 0.5 λ^{2} t n) = t \frac{ɛ^{2}}{2 n} .

Then, owing to

1 \geq e^{t \frac{ɛ^{2}}{2 n}} E I_{A}

, we get

ϱ (t) log P (A) \leq - t^{2 (1 - κ)} \frac{ɛ^{2}}{2 n} \to - \infty .

□ References

Albert, A. (1972) Regression and the Moore-Penrose Pseudoinverse. Academic Press, New York and London.
Bayer, U., Freidlin, M.I. (1977) Theorems on large deviations and stability under random perturbations” DAN USSR. 235, 2, pp. 253-256.
Bhattacharya, R.N. (1992) On the functional central limit theorem and the law of the iterated logarithm for Markov processes, Z. Wharsch. verw. Geb. 60, pp. 185–201.
Delyon, B., Juditsky, A. and Liptser, R. (2005) Moderate deviation principle for ergodic Markov chain. Lipschitz summands Shiryev's Festschrift.
Dembo, A. (1996) Moderate deviations for martingales with bounded jumps,Elect. Comm. in Probab. 1, pp. 11-17.
Down, D., Meyn, S.P. and Tweedie, R.L. (1995) Exponential and uniform ergodicity of Markov processes. Ann. Probab. 23, no. 4, pp. 1671-1691.
Ethier, S.N., Kurtz, T.G. (1986), Markov processes. Characterization and convergence, Wiley Series in Probability and Mathematical Statistics, John Wiley & Sons, New York et al.
Gong, F. and Wu, L. (2000) Spectral gap of positive operators and applications, C. R. Acad. Sci., Ser. I, Math. 331(12), pp. 983-988.
Guillin, A. (2001) Moderate deviations of inhomogeneous functionals of Markov processes and application to averaging. Stoch. Proc. Appl. 92, pp. 287-313.
Guillin, A. (2003) Averaging of SDE with small diffusions: Moderate deviations. Ann. Prob., Vol. 31(1), pp. 413-443.
Inglot, T. and Kallenberg, C.M. (2000) Moderate Deviations of Minimum Contrast Estimators under Contamination. Preprint.
Kalman, R.E. (1960) Contribution in the theory of optimal control. Bol. Soc. Mat. Mex., 5, pp. 102-119.
Krylov, N.V. (1980) Controlled diffusion processes. Springer, (in Russian, Moscow, 1977).
Khasminskii, R.Z. (1980). Stochastic stability of differential equations. Sijthoff & Noordhoff.
Ladyzenskaja, O., Solonnikov, V., Ural'ceva, N. (1968) Linear and quasilinear equations of parabolic type. Translation Monographs, 23, AMS, Provvidence.
Liptser, R.Sh. and Shiryayev, A.N. (1989) Theory of Martingales. Kluwer Acad. Publ.
Mattingly, J. C. and Stuart, A. M., (2002) Geometric ergodicity of some hypo-elliptic diffusions for particle motions, Markov Process. Related Fields, vol. 8 no. 2 , pp. 199–214 (Inhomogeneous random systems (Cergy-Pontoise, 2001)
Mattingly, J. C., Stuart, A. M. and Higham, D. J. (2002) Ergodicity for SDEs and approximations: locally Lipschitz vector fields and degenerate noise, Stochastic Process. Appl., vol. 101 no., pp. 185–232.
Meyn, S.P.and Tweedie, R.L. (1993) Markov chains and stochastic stability. Springer-Verlag.
Papanicolaou, C.C., Stroock, D.W., Varahan, S.R.S. (1977) Martingale approach to some limit theorems. in: Conference on Statistical Mechanics, Dinamical Systems and Turbulence, M. Reed ed., Duke Univ. Math. Series, 3.
Pardoux, E., Veretennikov, A.Yu. (2001) On Poisson equation and diffusion approximation, 1. Ann. Prob. 29 (2001), n. 3, pp. 1061-1085.
Pardoux, E., Veretennikov, A.Yu. (2003) On Poisson equation and diffusion approximation, 2. Ann. Prob. 31 , n. 3, pp. 1166-1192.
Puhalskii, A.A. (1991) On functional principle of large deviations”. New trends in Probability and Statistics., Vilnius, Lithuania, VSP/Mokslas, pp. 198-218.
Puhalskii, A.A. (1994) The method of stochastic exponentials for large deviations. Stochast. Proc. Appl. 54, , pp. 45-70.
Puhalskii, A. (1999) Large deviations of semimartingales: a maxingale problem approach. II. Uniqueness for the maxingale problem. Applications. Stoch. Stoch. Rep., 68, pp. 65-143.
Puhalskii, A. (2001) Large Deviations and Idempotent Probability, Chapman & Hall/CRC Press.
Varadhan, S.R.S. (1984) Large Deviations and Applications. SIAM, Philadelphia.
Veretennikov, A.Yu. (1999), On polynomial mixing and convergence rate for stochastic difference and differential equations. Teoria veroyatnostej i ee primeneniya. 44, 2, pp. 312–327 (in Russian; English version: preprint 393 (1998), WIAS, Berlin).
Wu, L. (1995) Moderate deviations of dependent random variables related to CLT and LIL, Annals of Probability. 23, no. 1, pp. 420-445.
Wu, L. (2000) Uniformly integrable operators and large deviations for Markov processes, J. Funct. Anal., 172(2), pp. 301-376.
Wu, L. (2000) Some notes on large deviations of Markov processes, Acta Math. Sin., Engl. Ser., 16(3), pp. 369-394.
Wu, L. (2001) The principle of large deviations for empirical processes, J. Math., Wuhan Univ., 21(3), pp. 295-300.
Wu, L. (20001) Large and moderate deviations and exponential convergence for stochastic damping Hamiltonian systems. Stochastic Processes and their Applications. 91, pp. 205-238.

CEREMADE, Universite Paris Dauphine and TSI, Ecole nationale des Telecommunications E-mail address : guillin@ceremade.dauphine.fr Electrical Engineering Systems, Tel Aviv University, 69978 Ramat Aviv, Tel Aviv, Israel E-mail address : liptser@eng.tau.ac.il

December 25, 2004.

A. Guillin

R. Liptser