This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Hitting times, functional inequalities, Lyapunov conditions and uniform ergodicity.

 Patrick Cattiaux     Patrick CATTIAUX,
Institut de Mathématiques de Toulouse. CNRS UMR 5219.
Université Paul Sabatier,
118 route de Narbonne, F-31062 Toulouse cedex 09.
cattiaux@math.univ-toulouse.fr
 and   Arnaud Guillin ♢,♣ Arnaud GUILLIN,
Laboratoire de Mathématiques, CNRS UMR 6620, Université Blaise Pascal, avenue des Landais, F-63177 Aubière.
guillin@math.univ-bpclermont.fr
(Date: September 7, 2025)
Abstract.

The use of Lyapunov conditions for proving functional inequalities was initiated in [5]. It was shown in [4, 30] that there is an equivalence between a Poincaré inequality, the existence of some Lyapunov function and the exponential integrability of hitting times. In the present paper, we close the scheme of the interplay between Lyapunov conditions and functional inequalities by

  • showing that strong functional inequalities are equivalent to Lyapunov type conditions;

  • showing that these Lyapunov conditions are characterized by the finiteness of generalized exponential moments of hitting times.

We also give some complement concerning the link between Lyapunov conditions and integrability property of the invariant probability measure and as such transportation inequalities, and we show that some “unbounded Lyapunov conditions” can lead to uniform ergodicity, and coming down from infinity property.

Université de Toulouse

Université Blaise Pascal

Institut Universitaire de France


Key words : Lyapunov functions, hitting times, uniform ergodicity, F-Sobolev inequalities, Poincaré inequality, ultracontractivity.

MSC 2010 : 26D10, 39B62, 47D07, 60G10, 60J60

1. Introduction

Let DD be some smooth open domain in d\mathbb{R}^{d}. In this paper, we will mainly consider the differential operator defined for smooth functions fC(D)f\in C^{\infty}(D) by

Lf=12ij(σσ)ij(x)ij2fxij+ibi(x)ifxi,Lf=\frac{1}{2}\sum_{ij}(\sigma\,\sigma^{*})_{ij}(x)\,\frac{\partial^{2}_{ij}f}{\partial x_{ij}}+\sum_{i}b_{i}(x)\frac{\partial_{i}f}{\partial x_{i}}\,,

where σ\sigma is an d×m{\mathbb{R}}^{d\times m} smooth and bounded (for simplicity Cb(D¯)C_{b}^{\infty}(\bar{D})) matrix field and bb a C(D¯)C^{\infty}(\bar{D}) vector field.
We may see LL as the infinitesimal generator of a diffusion process associated to the stochastic differential equation (SDE)

dXt=σ(Xt)dBt+b(Xt)dt,X0=x,dX_{t}=\sigma(X_{t})dB_{t}+b(X_{t})dt\quad,\quad X_{0}=x\,,

where BtB_{t} is an usual m{\mathbb{R}}^{m}-Brownian motions when D=dD=\mathbb{R}^{d}, or to the reflected SDE

dXt=σ(Xt)dBt+b(Xt)dt+dNt,0t 1D(Xs)𝑑Ns=NtX0=x,dX_{t}=\sigma(X_{t})dB_{t}+b(X_{t})dt+dN_{t},\quad\int_{0}^{t}\,\mathbf{1}_{\partial D}(X_{s})dN_{s}=N_{t}\quad X_{0}=x\,,

if DD is some smooth subdomain.
The domain 𝒟(L)\mathcal{D}(L) of LL (viewed as a generator) is thus some extension of the set of smooth and compactly supported functions Cc(D¯)C_{c}^{\infty}(\bar{D}) such that the normal derivative fn\frac{\partial f}{\partial n} vanishes on D\partial D (if D\partial D is non void). This corresponds to normal reflection or to a Neumann condition on the boundary. We also define PtP_{t} the associated semi-group

Ptf(x)=𝔼x(f(Xt))P_{t}f(x)=\mathbb{E}_{x}(f(X_{t}))

which is defined for bounded functions ff.
In order to use classical results in PDE theory we will also assume that LL is uniformly elliptic, i.e.

σσaId\sigma\,\sigma^{*}\,\geq\,a\,Id

in the sense of quadratic forms for some a>0a>0, or more generally that LL is uniformly strongly hypo-elliptic in the sense of Bony (see [15]) and that the boundary is non characteristic. For details we refer to [17].

We will also assume (though it should be a consequence of some of our assumptions) that there exists a probability measure

μ(dx)=eV(x)dx\mu(dx)=e^{-V(x)}\,dx

which is an invariant measure for the process (or the semi-group) i.e. for all bounded and smooth function f𝒟(L)f\in\mathcal{D}(L),

Lf𝑑μ=0 or equivalently for all t,𝔼μ(f(Xt))=f𝑑μ.\int\,Lf\,d\mu=0\quad\textrm{ or equivalently for all $t$,}\quad{\mathbb{E}}_{\mu}(f(X_{t}))=\int f\,d\mu\,.

PtP_{t} then extends to a contraction semi-group on 𝕃p(μ)\mathbb{L}^{p}(\mu) for 1p+1\leq p\leq+\infty. We shall say that μ\mu is symmetric, or that PtP_{t} is μ\mu symmetric if for smooth ff and gg in the domain of LL,

fLg𝑑μ=Lfg𝑑μ.\int f\,Lg\,d\mu=\int\,Lf\,g\,d\mu\,.

The standard example of μ\mu-symmetric semi-group is obtained for σ=2Id\sigma=\sqrt{2}\,Id and b=Vb=-\nabla V (provided VV is smooth enough). In all cases our ellipticity assumptions imply that this measure is unique and ergodic.

Among the most fascinating recent developments at the border of analysis and probability theory, a lot of work has been devoted to the study of the relationship between

  • geometric properties of the measure μ\mu, for instance concentration properties,

  • functional inequalities (the study of weighted Sobolev or Orlicz-Sobolev spaces associated to LL and μ\mu) like the Poincaré (Wirtinger) inequality or the Gross logarithmic Sobolev inequality,

  • transportation inequalities like the T2T_{2} Talagrand’s inequality,

  • the rate of convergence to equilibrium for the semi-group PtP_{t} in various functional spaces, 𝕃2(μ)\mathbb{L}^{2}(\mu), Orlicz spaces related to μ\mu,

  • the rate of convergence of its dual PtP_{t}^{*} (i.e. the distribution of the process at time tt) in total variation or in Wasserstein distance,

  • the existence of Lyapunov functions,

  • and finally some properties of the stochastic process X.X_{.} in particular existence of general moments for hitting times of some subsets (for instance the control of how the process comes down from infinity in the ultracontractive situation).

We refer to the monographs of Davies [33], Ledoux [45], Wang [58] and Bakry-Gentil-Ledoux [7], the surveys by Gross [39] and Ledoux [44], the collective book [3] and the papers [2, 6, 9, 8, 10, 11, 14, 13, 12, 18, 22, 20, 43, 55] among many others, for the first four items.
For the last three items we refer to the monographs of Hasminskii [40] and Meyn-Tweedie [52] and the papers [16, 35, 36, 37, 53, 54, 57] among many others.

The link between both approaches was done in [5] for the first time, up to our knowledge. It was extended in [4, 24, 27, 28, 29]. One can see the (now outdated) survey [25].

To be a little bit more precise, let us recall the following result from [30] Theorem 2.3 (also see [42] for a similar statement)

Theorem 1.1.

Consider the following properties

  1. (HP1)

    There exists a Lyapunov function WW, i.e. a smooth function W:DW:D\to{\mathbb{R}}, s.t. Ww>0W\geq w>0, and there exist a constant λ>0\lambda>0 and an open connected bounded subset UU such that Wn=0\frac{\partial W}{\partial n}=0 on D\partial D and

    LWλW+𝟏U¯.LW\,\leq\,-\,\lambda\,W\,+\mathbf{1}_{\bar{U}}\,.
  2. (HP2)

    There exist an open connected bounded subset UU and a constant θ>0\theta>0 such that for all xx,

    𝔼x(eθTU)<+,{\mathbb{E}}_{x}\left(e^{\theta\,T_{U}}\right)<+\infty\,,

    where TUT_{U} denotes the hitting time of UU.

  3. (HP3)

    The process is geometrically ergodic, i.e. there exist constants β>0\beta>0 and C>0C>0 and a function W1W\geq 1 belonging to 𝕃1(μ){\mathbb{L}}^{1}(\mu) such that for all xx

    Pt(x,.)μTVCW(x)eβt,\|P_{t}(x,.)-\mu\|_{TV}\,\leq\,C\,W(x)\,e^{-\beta\,t}\,,

    where Pt(x,.)P_{t}(x,.) denotes the distribution of XtX_{t} (when X0=xX_{0}=x) and .TV\|.\|_{TV} denotes the total variation distance.

  4. (HP4)

    μ\mu satisfies a Poincaré inequality, i.e. there exists a constant CP(μ)C_{P}(\mu) such that for all smooth f𝒟(L)f\in\mathcal{D}(L),

    Varμ(f)CP(μ)(f),\textrm{Var}_{\mu}(f)\,\leq\,C_{P}(\mu)\,\mathcal{E}(f)\,,

    where

    (f)=Lffdμ=12|σf|2𝑑μ.\mathcal{E}(f)\,=\,\int\,-Lf\,f\,d\mu\,=\,\frac{1}{2}\,\int\,|\sigma\,\nabla f|^{2}\,d\mu\,.
  5. (HP5)

    There exists a constant λP(μ)\lambda_{P}(\mu) such that for all f𝕃2(μ)f\in{\mathbb{L}}^{2}(\mu),

    Varμ(Ptf)eλP(μ)tVarμ(f).\textrm{Var}_{\mu}(P_{t}f)\,\leq\,e^{-\,\lambda_{P}(\mu)\,t}\,\textrm{Var}_{\mu}(f)\,.

Then (HP5)(HP4)(HP5)\Leftrightarrow(HP4), (HP4)(HP2)(HP4)\Rightarrow(HP2), (HP2)(HP1)(HP2)\Leftrightarrow(HP1) and (HP1)(HP3)(HP1)\Rightarrow(HP3). Actually (HP4)(HP4) implies (HP2)(HP2) for all (non-empty) open connected and bounded subset UU.
If in addition μ\mu is symmetric, then

(HP1)(HP2)(HP3)(HP4)(HP5).(HP1)\Leftrightarrow(HP2)\Leftrightarrow(HP3)\Leftrightarrow(HP4)\Leftrightarrow(HP5)\,.

When μ\mu is not symmetric, examples are known (kinetic diffusions) where (HP1)(HP1) (hence (HP3)(HP3)) is satisfied but (HP4)(HP4) is not. In the case of kinetic diffusions, it is evident that (HP4)(HP4) cannot hold as the Dirichlet form is degenerate. It is however also possible to build ”monster” diffusion where the invariant probability measure has some polynomial tail but the diffusion (with identity diffusion matrix) may converge exponentially fast and thus a Lyapunov condition holds.

The first equivalence is well known (see [3]), the second one is a simple application of Ito’s calculus and PDE results (see [30] and [17]), the final implication is a consequence of the Meyn-Tweedie theory. The implication (HP4)(HP2)(HP4)\Rightarrow(HP2) is shown in [30] by using the deviation results for the occupation measure obtained in [23] (using a beautiful deviation result obtained in [59]). We shall see in the next section another much more direct approach. Finally, in the symmetric case the converse implications are obtained by using the method in [4].

Some extensions of this theorem to polynomial ergodicity are discussed in [30] in connexion with weak Poincaré inequalities. A deeper study of this situation is done in [50, 49, 48].

The questions we shall address in the present paper are not concerned with weakening but with reinforcing of the assumptions, that is, does it exist similar results as in Theorem 1.1 if we replace the Poincaré inequality by stronger inequalities, for instance FF-Sobolev inequalities in the spirit of [1, 9, 10] ?

Partial answers are known: an FF-Sobolev inequality is equivalent to exponential stabilization in some Orlicz space (see [56]) and in the symmetric situation, reinforced Lyapunov conditions imply super-Poincaré inequalities or FF-Sobolev inequalities (see [27, 28, 29]). Recently, Liu ([47, 46]) proposed some new ideas in order to directly link Lyapunov conditions on one hand, concentration properties or functional inequalities on the other hand. Though some aspects of his proofs are a little bit obscure for us, we shall follow his main idea in order to deduce a Lyapunov condition from a functional inequality, and then get equivalent results in terms of hitting times. This yields the following result written here for the logarithmic Sobolev inequality which is the best known FF-Sobolev

Theorem 1.2.

Assume that DD is not bounded. Consider the following properties

  1. (HLS1)

    There exists a Lyapunov function WW, i.e. there exists a smooth function W:DW:D\to{\mathbb{R}}, with Ww>0W\geq w>0, and there exist constants λ>0\lambda>0 and b>0b>0 such that Wn=0\frac{\partial W}{\partial n}=0 on D\partial D and

    LW(x)λ|x|2W(x)+b.LW(x)\,\leq\,-\,\lambda\,|x|^{2}\,W(x)\,+b\,.
  2. (HLS1’)

    There exists a Lyapunov function WW, i.e. there exists a smooth function W:DW:D\to{\mathbb{R}}, with Ww>0W\geq w>0, and there exist constants λ>0\lambda>0 and b>0b>0 such that Wn=0\frac{\partial W}{\partial n}=0 on D\partial D and

    LW(x)λV(x)W(x)+b.LW(x)\,\leq\,-\,\lambda\,V(x)\,W(x)\,+b\,.
  3. (HLS2)

    There exist an open connected bounded subset UU and a constant θ>0\theta>0 such that for all xx,

    𝔼x(exp(0TUθ|Xs|2𝑑s))<+,{\mathbb{E}}_{x}\left(\exp\left(\int_{0}^{T_{U}}\,\theta\,|X_{s}|^{2}\,ds\right)\right)<+\infty\,,

    where TUT_{U} denotes the hitting time of UU.

  4. (HLS2’)

    There exist an open connected bounded subset UU and a constant θ>0\theta>0 such that for all xx,

    𝔼x(exp(0TUθV(Xs)𝑑s))<+,{\mathbb{E}}_{x}\left(\exp\left(\int_{0}^{T_{U}}\,\theta\,V(X_{s})\,ds\right)\right)<+\infty\,,

    where TUT_{U} denotes the hitting time of UU.

  5. (HLS4)

    μ\mu satisfies a logarithmic-Sobolev inequality, i.e. there exists a constant CLS(μ)C_{LS}(\mu) such that for all smooth f𝒟(L)f\in\mathcal{D}(L),

    Entμ(f2):=f2ln(f2f2𝑑μ)𝑑μCLS(μ)(f),Ent_{\mu}(f^{2})\,:=\,\int\,f^{2}\,\ln\left(\frac{f^{2}}{\int\,f^{2}\,d\mu}\right)\,d\mu\,\leq\,C_{LS}(\mu)\,\mathcal{E}(f)\,,

    where

    (f)=Lffdμ=12|σf|2𝑑μ.\mathcal{E}(f)\,=\,\int\,-Lf\,f\,d\mu\,=\,\frac{1}{2}\,\int\,|\sigma\,\nabla f|^{2}\,d\mu\,.
  6. (HLS5)

    There exists a constant CE(μ)C_{E}(\mu) such that for all f2𝕃ln𝕃(μ)f^{2}\in{\mathbb{L}}\ln{\mathbb{L}}(\mu) s.t. f2𝑑μ=1\int f^{2}d\mu=1,

    Entμ(Pt(f2))eCE(μ)tEntμ(f2).Ent_{\mu}(P_{t}(f^{2}))\,\leq\,e^{-\,C_{E}(\mu)\,t}\,\textrm{Ent}_{\mu}(f^{2})\,.

Then

  • 1)

    (HLS5)(HLS4)(HLS5)\Leftrightarrow(HLS4), (HLS4)(HLS1)(HLS4)\Rightarrow(HLS1) and (HLS2)(HLS1)(HLS2)\Leftrightarrow(HLS1).

  • 1’)

    Assume that VV goes to infinity at infinity and that there exists some a>0a>0 such that μ(eaV)<+.\mu(e^{aV})<+\infty.
    Then (HLS5)(HLS4)(HLS5)\Leftrightarrow(HLS4), (HLS4)(HLS1)(HLS4)\Rightarrow(HLS1^{\prime}) and (HLS2)(HLS1)(HLS2^{\prime})\Leftrightarrow(HLS1^{\prime}).

Actually (HLS4)(HLS4) implies in both cases (HLS2)(HLS2) or (HLS2)(HLS2^{\prime}) for all open nice subset UU.

Assume in addition that μ\mu is symmetric and that σ.σ\sigma.\sigma^{*} is uniformly elliptic.

  • 2’)

    Assume that VV goes to infinity at infinity, that |V(x)|v>0|\nabla V(x)|\geq v>0 for |x||x| large enough and that there exists some a>0a>0 such that μ(eaV)<+.\mu(e^{aV})<+\infty.
    Then (HLS1)(HLS2)(HLS4)(HLS5)(HLS1^{\prime})\Leftrightarrow(HLS2^{\prime})\Leftrightarrow(HLS4)\Leftrightarrow(HLS5).

  • 2)

    Assume the curvature condition Ric+HessVC>Ric+HessV\geq-C>-\infty where Ricci and Hess are related to the riemanian metric defined by σ\sigma.
    Then (HLS1)(HLS2)(HLS4)(HLS5)(HLS1)\Leftrightarrow(HLS2)\Leftrightarrow(HLS4)\Leftrightarrow(HLS5).

Except (HLS5)(HLS4)(HLS5)\Leftrightarrow(HLS4) which is well known, we will prove this Theorem (and actually more general results) in section 3. Part of this result can be extended to general FF-Sobolev inequalities, this is done in section 4.

The next two sections are devoted to the rate of convergence of Pt(ν,.)P_{t}(\nu,.), the distribution at time tt of the process with initial distribution ν\nu, to the invariant distribution μ\mu for the total variation distance. In section 5 Theorem 5.4 we get the following: under some natural assumptions on μ\mu, almost any FF-Sobolev inequality combined with the Poincaré inequality provides an exponential convergence

Pt(ν,.)μTVC(ν)eβt,\|P_{t}(\nu,.)-\mu\|_{TV}\leq\,C(\nu)\,e^{-\,\beta\,t}\,,

for all ν\nu absolutely continuous w.r.t. μ\mu such that dνdμ\frac{d\nu}{d\mu} belongs to 𝕃p(μ)\mathbb{L}^{p}(\mu) for some p>1p>1. The remarkable fact here is that β\beta does not depend on the integrability property (the pp) of dν/dμd\nu/d\mu. If this and more general results were proved in [24], the proof given here is particularly simple and understandable.
In the next section 6 we study 𝕃\mathbb{L}^{\infty} properties of the Lyapunov functions, in relation with the property of “coming down from infinity” for the process. In particular we show that if the ultra boundedness property of the semi-group implies the “coming down from infinity” property for the process, the converse is not true. All these notions are particularly relevant in the study of quasi-stationnary distributions.

Finally in the last section we directly rely general Lyapunov condition to the existence of some exponential moments for the measure μ\mu, extending the results in [47].

2. Back to the Poincaré inequality.

As we said in the introduction, we shall give here a new direct proof of (HP4)(HP1)(HP4)\Rightarrow(HP1) in Theorem 1.1.

Theorem 2.1.

Assume that μ\mu satisfies a Poincaré inequality with constant CP(μ)C_{P}(\mu). Then for all open subset AA, there exists a smooth Lyapunov function W𝒟(L)W\in\mathcal{D}(L) i.e. a smooth function satisfying Ww>0W\geq w>0 on AcA^{c} and LWcWLW\leq-c\,W on AcA^{c} with

c=μ(A)min(14CP(μ),18).c=\mu(A)\min\left(\frac{1}{4C_{P}(\mu)}\,,\,\frac{1}{8}\right)\,.

It is easily seen that if AA is smooth an bounded, we can modify WW to get (HP1)(HP1) (with a λ\lambda smaller than the cc in the theorem).

Proof.

Let start with a simple lemma.

Lemma 2.2.

Assume that μ\mu satisfies a Poincaré inequality with constant CP(μ)C_{P}(\mu). Then for all subset AA such that μ(A)>0\mu(A)>0 and all ff in H1(μ)=𝒟()H^{1}(\mu)=\mathcal{D}(\mathcal{E}) it holds

f2𝑑μ2CP(μ)μ(A)(f)+4μ(A)Af2𝑑μ.\int\,f^{2}\,d\mu\leq\,\frac{2C_{P}(\mu)}{\mu(A)}\,\mathcal{E}(f)\,+\,\frac{4}{\mu(A)}\,\int_{A}\,f^{2}\,d\mu\,.
Proof.

Using Poincaré inequality and the elementary

(a+b)2(1+λ)a2+(1+1λ)b2 for all λ>0(a+b)^{2}\leq(1+\lambda)a^{2}+\left(1+\frac{1}{\lambda}\right)\,b^{2}\quad\textrm{ for all $\lambda>0$}

we can write

f2𝑑μ\displaystyle\int\,f^{2}\,d\mu \displaystyle\leq CP(μ)(f)+μ2(f)\displaystyle C_{P}(\mu)\,\mathcal{E}(f)\,+\,\mu^{2}(f)
\displaystyle\leq CP(μ)(f)+μ2(f𝟏A+f𝟏Ac)\displaystyle C_{P}(\mu)\,\mathcal{E}(f)\,+\mu^{2}(f\mathbf{1}_{A}+f\mathbf{1}_{A^{c}})
\displaystyle\leq CP(μ)(f)+(1+λ)μ2(f𝟏A)+1+λλμ2(f𝟏Ac)\displaystyle C_{P}(\mu)\,\mathcal{E}(f)\,+(1+\lambda)\,\mu^{2}(f\mathbf{1}_{A})+\frac{1+\lambda}{\lambda}\,\mu^{2}(f\mathbf{1}_{A^{c}})
\displaystyle\leq CP(μ)(f)+(1+λ)μ(A)Af2𝑑μ+1+λλμ(f2)μ(Ac)\displaystyle C_{P}(\mu)\,\mathcal{E}(f)\,+(1+\lambda)\,\mu(A)\,\int_{A}\,f^{2}\,d\mu+\frac{1+\lambda}{\lambda}\,\mu(f^{2})\mu(A^{c})

where we have used Cauchy-Schwartz in the final inequality. Hence provided

μ(A)(1+λ)1> 0,\mu(A)(1+\lambda)-1\,>\,0\,,

we have obtained

f2𝑑μλCP(μ)μ(A)(1+λ)1(f)+λ(1+λ)μ(A)μ(A)(1+λ)1Af2𝑑μ.\int\,f^{2}\,d\mu\leq\frac{\lambda\,C_{P}(\mu)}{\mu(A)(1+\lambda)-1}\,\mathcal{E}(f)\,+\frac{\lambda(1+\lambda)\,\mu(A)}{\mu(A)(1+\lambda)-1}\,\int_{A}\,f^{2}\,d\mu\,.

The result follows by choosing λ=2(1μ(A))/μ(A)\lambda=2(1-\mu(A))/\mu(A). ∎

Now define

ϕ(x)=c+ 1A(x),\phi(x)=-c+\,\mathbf{1}_{A}(x)\,, (2.3)

and introduce for all smooth u𝒟(L)u\in\mathcal{D}(L),

Hu=Lu+ϕu.Hu=-Lu+\phi\,u\,.

On one hand, it holds

μ(u.Hu)(u)+Au2dμ.\mu(u.Hu)\leq\mathcal{E}(u)+\,\int_{A}\,u^{2}\,d\mu\,.

On the other hand, applying the previous lemma we have

μ(u.Hu)\displaystyle\mu(u.Hu) =\displaystyle= (u)+Au2𝑑μcμ(u2)\displaystyle\mathcal{E}(u)+\,\int_{A}\,u^{2}\,d\mu-c\mu(u^{2})
\displaystyle\geq (u)+Au2𝑑μc(2CP(μ)μ(A)(u)+4μ(A)Au2𝑑μ)\displaystyle\mathcal{E}(u)+\,\int_{A}\,u^{2}\,d\mu-c\,\left(\frac{2C_{P}(\mu)}{\mu(A)}\,\mathcal{E}(u)\,+\,\frac{4}{\mu(A)}\,\int_{A}\,u^{2}\,d\mu\right)
\displaystyle\geq 12((u)+Au2𝑑μ)\displaystyle\frac{1}{2}\left(\mathcal{E}(u)+\,\int_{A}\,u^{2}\,d\mu\right)

if we choose

c=μ(A)min(14CP(μ),18).c=\mu(A)\min\left(\frac{1}{4C_{P}(\mu)}\,,\,\frac{1}{8}\right)\,.

Now we will linearize μ(u.Hu)\mu(u.Hu). If v𝒟(L)v\in\mathcal{D}(L) and uH1(μ)u\in H^{1}(\mu), (u,v)=μ(u.Hv)\mathcal{H}(u,v)=\mu(u.Hv) is well defined, and using an integration by parts (or the Green-Rieman formula since the normal derivative of vv at the boundary vanishes) can be written as a (non necessarily symmetric) bilinear form on H1(μ)H^{1}(\mu). It is easily seen that this bilinear form \mathcal{H} is continuous on H1(μ)H^{1}(\mu) equipped with the (usual) hilbertian norm ((u)+u2𝑑μ)12\left(\mathcal{E}(u)+\int\,u^{2}\,d\mu\right)^{\frac{1}{2}}, hence equipped with the hilbertian norm ((u)+Au2𝑑μ)12\left(\mathcal{E}(u)+\int_{A}\,u^{2}\,d\mu\right)^{\frac{1}{2}} which is equivalent according to Lemma 2.2. But according to what precedes, \mathcal{H} is also coercive for this norm.
Hence, we may apply the Lax-Milgram theorem which tells us that for any gH1(μ)g\in H^{1}(\mu) there exists some vH1(μ)v\in H^{1}(\mu) such that for all uu, (u,v)=u,g\mathcal{H}(u,v)=\langle u,g\rangle. We will use this result with g1g\equiv 1.
First of all, the previous relation with uu compactly supported in DD shows that Hv=gHv=g in 𝒟(D)\mathcal{D}^{\prime}(D), so that thanks to ellipticity (or hypo-ellipticity), vC(D)v\in C^{\infty}(D) and satisfies Hv=gHv=g in the usual sense in DD. When D=dD=\mathbb{R}^{d} this is enough. Otherwise, since the boundary is non characteristic, vv admits sectional traces on D\partial D of any order (see [17] Theorem 4.6) and using the results in [17] section 4, one can see that vC(D¯)v\in C^{\infty}(\bar{D}) and satisfies nv=0\partial_{n}v=0 on D\partial D. Since Hv𝕃2(μ)Hv\in\mathbb{L}^{2}(\mu), it follows that v𝒟(H)v\in\mathcal{D}(H).
As a routine, defining v=min(v,0)v_{-}=\min(v,0), one can check using integration by parts or the Green-Rieman formula, that μ(v.Lv)=(v)\mu(v_{-}.Lv)=\mathcal{E}(v_{-}) so that, using the previous lower bound we obtain,

μ(v.g)\displaystyle\mu(v_{-}.g) =\displaystyle= (v,v)=μ(v.Hv)=μ(v.Lv)+μ(ϕ.v.v)=(v)μ(ϕ.v2)\displaystyle\mathcal{H}(v_{-},v)=\mathcal{\mu}(v_{-}.Hv)\,=\,\mathcal{\mu}(v_{-}.Lv)\,+\,\mu(\phi.v.v_{-})=-\,\mathcal{E}(v_{-})\,-\,\mu(\phi.v^{2}_{-})
\displaystyle\leq 12((v)+Av2𝑑μ)\displaystyle\,-\,\frac{1}{2}\,\left(\mathcal{E}(v_{-})+\,\int_{A}\,v_{-}^{2}\,d\mu\right)

and v=0v_{-}=0 μ\mu since g>0g>0. So v0v\geq 0 almost surely.
One should now use the maximum principle but we prefer use Ito’s formula. Assume that AA is open and bounded. Since ϕ=c\phi=-c on AcA^{c}, we get for any xAcx\in A^{c},

v(x)\displaystyle v(x) =\displaystyle= 𝔼x(v(XtTA)+𝔼x(0t 1sTA(Hvϕv)(Xs)ds)\displaystyle{\mathbb{E}}_{x}(v(X_{t\wedge T_{A}})+{\mathbb{E}}_{x}\left(\int_{0}^{t}\,\mathbf{1}_{s\leq T_{A}}\,(Hv-\phi\,v)(X_{s})\,ds\right)
\displaystyle\geq 𝔼x(tTA)> 0\displaystyle\,{\mathbb{E}}_{x}(t\wedge T_{A})>\,0\,

since TA>0T_{A}>0 if xAcx\in A^{c}. Replacing AA by Aε={y,d(y,A)<ε}A_{\varepsilon}=\{y,d(y,A)<\varepsilon\} we may finally let tt go to infinity, use the fact that vv is bounded from below by a positive constant v(A,ε)v(A,\varepsilon) on Aε\partial A_{\varepsilon} which is compact and obtain that v(x)v(A,ε)v(x)\geq v(A,\varepsilon) for all xAεcx\in A_{\varepsilon}^{c} using the previous inequality (which actually furnishes exactly the minimum principle). For a general open set AA just take the intersection with a large ball to get a bounded subset. ∎

Remark 2.4.

As soon as we know the existence of a Lyapunov function satisfying (HP1)(HP1) it immediately follows using Ito’s formula with the function (t,y)eλtW(y)(t,y)\mapsto e^{\lambda t}\,W(y) that for all xx,

WU(x)=𝔼x(eλTU)<+,W_{U}(x)={\mathbb{E}}_{x}(e^{\lambda\,T_{U}})<+\infty\,,

and conversely, if the exponential moment is finite, WUW_{U} is a Lyapunov function (see [30]).
Also notice that the proof of Theorem 2.1 furnishes a Lyapunov function in H1(μ)H^{1}(\mu), hence in 𝒟(L)\mathcal{D}(L) since Lv=(𝟏Ac)v1Lv=(\mathbf{1}_{A}-c)v-1 implies Lv𝕃2(μ)Lv\in\mathbb{L}^{2}(\mu). Hence the previous WUW_{U} belongs to 𝒟(L)\mathcal{D}(L). Conversely if WUW_{U} is finite, LWU=λWULW_{U}=-\lambda W_{U} in UcU^{c}. Replacing λ\lambda by λ/2\lambda/2 if necessary, we may assume that WUW_{U} is in 𝕃2(μ)\mathbb{L}^{2}(\mu) so that LWULW_{U} is square integrable too (at least in UcU^{c}). If UU is relatively compact it is easy to see that one can modify UU and WUW_{U} to get a smooth function everywhere that belongs to 𝒟(L)\mathcal{D}(L). \diamondsuit

3. The logarithmic Sobolev inequality.

We start with an analogue of Theorem 2.1.

Proposition 3.1.

Assume that μ\mu satisfies the logarithmic Sobolev inequality (HLS4)(HLS4). Let hh be a non-negative continuous function such that b=2μ(eh)<+b=2\mu(e^{h})<+\infty. For ε>0\varepsilon>0, define Uε(h)={(1ε)h>b}U_{\varepsilon}(h)=\{(1-\varepsilon)h>b\}.
Then there exists a Lyapunov function W𝒟(L)W\in\mathcal{D}(L) such that W(x)wε>0W(x)\geq w_{\varepsilon}>0 on Uε(h)U_{\varepsilon}(h) and

(HLh)LWε2CLShW on Uε(h).(HLh)\quad LW\leq-\frac{\varepsilon}{2\,C_{LS}}\,h\,W\quad\textrm{ on $U_{\varepsilon}(h)$.}
Proof.

We follow and modify the proof in [46]. Assume that hh is a non-negative function such that μ(eh)<+\mu(e^{h})<+\infty. We follow the proof of Theorem 2.1 and define, for 2ρ1CLS2\,\rho\leq\frac{1}{C_{LS}} so that 2ρEntμ(u2)(u)2\rho\,Ent_{\mu}(u^{2})\leq\mathcal{E}(u),

ϕ(x)=ρ(h(x)+b),\phi(x)=\rho\,(-\,h(x)+\,b)\,, (3.2)

with b=2μ(eh)b=2\mu(e^{h}) and introduce for all smooth u𝒟(L)u\in\mathcal{D}(L),

Hu=Lu+ϕu.Hu=-Lu+\phi\,u\,.

On one hand, it holds

μ(u.Hu)(u)+ρbμ(u2).\mu(u.Hu)\leq\mathcal{E}(u)+\,\rho\,b\,\mu(u^{2})\,.

On the other hand, applying this time Young’s inequality and LSI, we get for a smooth u𝒟(L)u\in\mathcal{D}(L),

μ(u.Hu)\displaystyle\mu(u.Hu) =\displaystyle= (u)+ρbμ(u2)ρμ(hu2)\displaystyle\mathcal{E}(u)+\,\rho\,b\,\mu(u^{2})\,-\,\rho\,\mu(hu^{2})
\displaystyle\geq (u)+ρbμ(u2)ρμ(u2)μ(ehu2μ(u2)+u2μ(u2)ln(u2μ(u2)))\displaystyle\mathcal{E}(u)+\,\rho\,b\,\mu(u^{2})-\,\rho\,\mu(u^{2})\,\mu\left(e^{h}-\frac{u^{2}}{\mu(u^{2})}+\frac{u^{2}}{\mu(u^{2})}\,\ln\left(\frac{u^{2}}{\mu(u^{2})}\right)\right)
\displaystyle\geq (u)+ρbμ(u2)ρb2μ(u2)+ρμ(u2)ρEnt(u2)\displaystyle\mathcal{E}(u)+\,\rho\,b\,\mu(u^{2})-\,\rho\,\frac{b}{2}\,\mu(u^{2})\,+\,\rho\,\mu(u^{2})\,-\,\rho\,Ent(u^{2})
\displaystyle\geq 12((u)+ρbμ(u2)).\displaystyle\frac{1}{2}\,\left(\mathcal{E}(u)+\,\rho\,b\,\mu(u^{2})\right)\,.

We can then follow the proof of Theorem 2.1, and thus apply again the Lax-Milgram theorem to get the existence of a non-negative smooth function vH1(μ)v\in H^{1}(\mu) satisfying

Lv=ϕv 1= 1ρ(bh)v.Lv=\phi\,v\,-\,1\,=\,-\,1-\,\rho\,(b-h)\,v\,.

If in addition

h(x)>b,xUch(x)>b\quad,\forall x\in U^{c}

we obtain that vv is bounded from below by a positive constant in UεcU_{\varepsilon}^{c}. The proof is completed. ∎

This result is in particular interesting when DD is not bounded and hh goes to infinity at infinity. Two cases are mainly relevant, due to the converse statements we will prove below

Corollary 3.3.

Assume that μ\mu satisfies the logarithmic Sobolev inequality (HLS4)(HLS4) and that DD is not bounded. Then

  • 1)

    for all x0Dx_{0}\in D, there exists a Lyapunov function W𝒟(L)W\in\mathcal{D}(L) with W(x)w>0W(x)\geq w>0 for all xDx\in D satisfying

    LW(x)λd2(x,x0)W(x)+b,LW(x)\leq-\lambda\,d^{2}(x,x_{0})\,W(x)\,+\,b\,,

    for some λ\lambda and bb strictly positive;

  • 2)

    if in addition VV goes to infinity at infinity and eaV𝕃1(μ)e^{aV}\in\mathbb{L}^{1}(\mu) for some a>0a>0, there exists a Lyapunov function W𝒟(L)W\in\mathcal{D}(L) with W(x)w>0W(x)\geq w>0 for all xDx\in D satisfying

    LW(x)λV(x)W(x)+b,LW(x)\leq-\lambda\,V(x)\,W(x)\,+\,b\,,

    for some λ\lambda and bb strictly positive.

We may replace bb by b𝟏Ab\mathbf{1}_{A} for some well chosen bounded subset AA of DD.

Proof.

In case 2), just apply the previous proposition with h=a|V|h=a|V| and modify WW in the corresponding level set AA of VV. For case 1), recall that the logarithmic-Sobolev inequality implies that there exists some c>0c>0 such that μ(ecd2(.,x0))<+\mu(e^{cd^{2}(.,x_{0})})<+\infty and conclude as before with h=cd2(.,x0)h=cd^{2}(.,x_{0}). ∎

Introducing the process

dHt=h(Xt)dtdH_{t}=h(X_{t})\,dt

and applying Ito’s formula to (H,x)eε2CLSH(H,x)\mapsto e^{\frac{\varepsilon}{2\,C_{LS}}\,H} it is easy to show that (HLh)(HLh) implies for all xx,

Wh,ε(x)=𝔼x(exp(0TUε(h)ε2CLSh(Xs)𝑑s))<+.W_{h,\varepsilon}(x)\,=\,\mathbb{E}_{x}\left(\exp\left(\int_{0}^{T_{U_{\varepsilon(h)}}}\,\frac{\varepsilon}{2\,C_{LS}}\,h(X_{s})\,ds\right)\right)<+\infty\,. (3.4)

Conversely, if Wh,ε(x)W_{h,\varepsilon}(x) is finite for all xx, using the arguments in [17] one can prove that it satisfies (HLh)(HLh) with an equality instead of an inequality. Notice that once again we may apply the arguments in Remark 2.4.

To complete the proof of Theorem 1.2, it remains to look at the converse statements in the symmetric situation.
The first case is the case h(x)=cd2(x,x0)h(x)=cd^{2}(x,x_{0}). Statement 2) in Theorem 1.2 under the curvature assumption is proved in [28] using transportation inequalities. An alternative method of proof was recently proposed in [46] (with some points to be corrected). For the second case h=aVh=aV we will use the results in [27] based on super-Poincaré inequalities.

Proposition 3.5.

Assume that μ\mu is symmetric and that σ.σ\sigma.\sigma^{*} is uniformly elliptic. Assume in addition that VV goes to infinity at infinity, that |V(x)|v>0|\nabla V(x)|\geq v>0 for |x||x| large enough and that eaV𝕃1(μ)e^{aV}\in\mathbb{L}^{1}(\mu) for some a>0a>0.
If there exists a Lyapunov function WW with W(x)w>0W(x)\geq w>0 for all xDx\in D, Wn=0\frac{\partial W}{\partial n}=0 on D\partial D and satisfying

LW(x)λV(x)W(x)+b,LW(x)\leq-\lambda\,V(x)\,W(x)\,+\,b\,,

for some λ\lambda and bb strictly positive, then μ\mu satisfies a logarithmic-Sobolev inequality.

Proof.

We follow the method in [27] Theorem 2.1 (itself inspired by [4]). Let Ar={Vr}A_{r}=\{V\leq r\}. For r0r_{0} large enough and some λ<λ\lambda^{\prime}<\lambda we have

LW(x)λV(x)W(x)+b 1Ar0,LW(x)\leq-\lambda^{\prime}\,V(x)\,W(x)\,+\,b\,\mathbf{1}_{A_{r_{0}}}\,,

so that we may assume that

LWW(x)λV(x) for xArc and all r large enough.\frac{LW}{W}(x)\leq-\,\lambda\,V(x)\quad\textrm{ for $x\in A_{r}^{c}$ and all $r$ large enough.}

Denote by M=sup(V)M=\sup(-V). We have for ss0s\leq s_{0} and r>r0r>r_{0},

f2𝑑μ\displaystyle\int\,f^{2}\,d\mu =\displaystyle= Arf2𝑑μ+Arcf2𝑑μ\displaystyle\int_{A_{r}}f^{2}\,d\mu\,+\,\int_{A^{c}_{r}}f^{2}\,d\mu
\displaystyle\leq eM(1+bλr0)Arf2𝑑x+1λrλV(x)f2𝑑μ,\displaystyle e^{M}\left(1+\frac{b}{\lambda r_{0}}\right)\,\int_{A_{r}}f^{2}\,dx\,+\,\frac{1}{\lambda r}\int\lambda V(x)\,f^{2}d\mu,
\displaystyle\leq eM(1+bλr0)Arf2𝑑x+1λrf2(LWW)𝑑μ\displaystyle e^{M}\left(1+\frac{b}{\lambda r_{0}}\right)\,\int_{A_{r}}f^{2}\,dx\,+\,\frac{1}{\lambda\,r}\,\int f^{2}\,\left(\frac{-LW}{W}\right)\,d\mu
\displaystyle\leq eM(1+bλr0)(sAr|f|2dx+Csd/2(Ar|f|dx)2)+1λr|σ.f|2dμ.\displaystyle e^{M}\left(1+\frac{b}{\lambda r_{0}}\right)\,\left(s\,\int_{A_{r}}|\nabla f|^{2}\,dx\,+\,\frac{C}{s^{d/2}}\,\left(\int_{A_{r}}|f|\,dx\right)^{2}\right)\,+\,\frac{1}{\lambda\,r}\,\int|\sigma.\nabla f|^{2}\,\,d\mu\,.

The first part of the last bound is obtained by using (3.1.4) in [27] (it is here that we are using the assumption on |V||\nabla V|), while the second bound is obtained using integration by parts or the Green-Rieman formula (see [27] (2.2)). Using uniform ellipticity we thus obtain, denoting c=eM(1+bλr0)c=e^{M}\left(1+\frac{b}{\lambda r_{0}}\right)

μ(f2)(sca+1λr)|σ.f|2dμ+Csd/2ce2r(|f|dμ)2.\mu(f^{2})\,\leq\,\left(\frac{s\,c}{a}\,+\,\frac{1}{\lambda\,r}\right)\,\int|\sigma.\nabla f|^{2}\,\,d\mu\,+\,C\,s^{-d/2}\,c\,e^{2r}\,\left(\int\,|f|\,d\mu\right)^{2}\,. (3.6)

Hence choosing r=c/sr=c^{\prime}/s we get the following super-Poincaré inequality for small ss,

μ(f2)s|σ.f|2dμ+Cec~/s(|f|dμ)2,\mu(f^{2})\,\leq\,s\,\int|\sigma.\nabla f|^{2}\,\,d\mu\,+\,C^{\prime}\,e^{\tilde{c}/s}\,\left(\int\,|f|\,d\mu\right)^{2}\,,

which is known to be equivalent to a defective logarithmic Sobolev inequality (see the introduction of [27]). But the Lyapunov condition being stronger than (HP1)(HP1), we know that μ\mu satisfies a Poincaré inequality, hence using Rothaus lemma, that it satisfies a (tight) log-Sobolev inequality.

Remark 3.7.

With some additional effort one should replace (3.1.4) in [27] directly by a similar statement with σ.f\sigma.\nabla f instead of f\nabla f, even in the strongly hypo-elliptic case, replacing the arguments in [27] by the Jerison and Sanchez-Calle estimates for such operators, up to a modification of the power sd/2s^{-d/2} replaced by sms^{-m} where mm depends on the dimension of the graded Lie algebra. We do not want to go further into details here, that is why we choosed to only consider the uniformly elliptic situation. \diamondsuit


4. F-Sobolev inequalities.

We will extend the results of the previous section to general FF-Sobolev inequalities introduced by Aida ([1]) and studied in [58, 9, 10]. Actually we will not be as complete as for the log-Sobolev inequality, because for general functions FF instead of the logarithm, results are much more intricate. In particular the reader will find in [56] convergence results in Orlicz spaces (replacing (HLS5)(HLS5) we shall not give here.

We are mainly interested here with the following version of (defective) FF-Sobolev inequalities for a nice FF defined on +\mathbb{R}^{+}:

(HFS4defect)f2F(f2f2𝑑μ)CF(μ)(f)+DFμ(f2).(HFS4defect)\quad\int\,f^{2}\,F\left(\frac{f^{2}}{\int f^{2}\,d\mu}\right)\,\leq\,C_{F}(\mu)\,\mathcal{E}(f)\,+\,D_{F}\,\mu(f^{2})\,.

When DF=0D_{F}=0 one say that the inequality is tight and simply denote it by (HFS4)(HFS4). The relationship between FF-Sobolev and super-Poincaré inequalities is due to Wang ([58] Theorem 3.3.1 and Theorem 3.3.3). Recall some basic facts on these inequalities

Proposition 4.1.

We have:

  • (see [58].)   A super-Poincaré inequality

    μ(f2)s(f)+β(s)(μ(|f|))2\mu(f^{2})\leq s\,\mathcal{E}(f)+\beta(s)\,(\mu(|f|))^{2}

    implies (HFS4defect)(HFS4defect) for

    F(u)=1u0uξ(t/2)dt,ξ(t)=supa>0(1aβ(a)ta).F(u)=\frac{1}{u}\,\int_{0}^{u}\,\xi(t/2)\,dt\quad,\quad\xi(t)=\sup_{a>0}\left(\frac{1}{a}\,-\,\frac{\beta(a)}{ta}\right)\,.
  • (see [58].)   If (HFS4defect)(HFS4defect) holds true for FF such that .+1uF(u)𝑑u<+\int_{.}^{+\infty}\,\frac{1}{uF(u)}\,du<+\infty, then the semi-group PtP_{t} is ultra-bounded, i.e. for all t>0t>0 there exists C(t)C(t) such that,

    PtfC(t)f𝕃1(μ)\|P_{t}f\|_{\infty}\,\leq C(t)\,\|f\|_{\mathbb{L}^{1}(\mu)}

    so that if in addition a Poincaré inequality holds,

    Ptfμ(f)MeCtfμ(f)𝕃2(μ)\|P_{t}f-\mu(f)\|_{\infty}\,\leq M\,e^{-Ct}\,\|f-\mu(f)\|_{\mathbb{L}^{2}(\mu)}

    for some C>0C>0 (one can replace 𝕃2\mathbb{L}^{2} by any 𝕃p\mathbb{L}^{p} for p>1p>1 just changing CC using interpolation results, see e.g. [24, 26]).

  • (see [9] lemma 8.)   If F(1)=0F(1)=0, FF is C2C^{2} in a neighborhood of 11 and 2F(1)+F′′(1)=c>02F^{\prime}(1)+F^{\prime\prime}(1)=c>0, (HFS4)(HFS4) implies the Poincaré inequality (HP4)(HP4) with CP(μ)=1/(2c)C_{P}(\mu)=1/(2c).

  • (see [9] Remark 22.)   If F0F\geq 0 and F(u)c>0F(u)\geq c>0 for u2u\geq 2, then (HFS4)(HFS4) implies the Poincaré inequality (HP4)(HP4).

  • (see [9] lemma 9.)   If FF is concave, non-decreasing, growths to infinity and satisfies F(1)=0F(1)=0 and uF(u)MuF^{\prime}(u)\leq M, then (HFSdefect)(HFSdefect) and the Poincaré inequality imply (HFS4)(HFS4).

We thus have two situations: either FlogF\leq\log (interpolating between Poincaré and log-Sobolev), in which case (with additional structural conditions on FF) FF-Sobolev inequalities are equivalent to an exponential convergence in some Orlicz space ([56]), we still have some Rothaus(-Orlicz) lemma allowing us to tight a defective FF-Sobolev inequality and a lot of additional properties connected with Orlicz hyperboundedness, concentration and isoperimetry (see [9, 10]), or .+1uF(u)𝑑u<+\int_{.}^{+\infty}\,\frac{1}{uF(u)}\,du<+\infty in which case exponential convergence holds in 𝕃\mathbb{L}^{\infty}, with a very small gap between both classes of FF.

We will now prove the analogue of Proposition 3.1. To this end we need to introduce some convexity notions.

Definition 4.2.

Assume that uuF(u)=G(u)u\mapsto uF(u)=G(u) is convex. We define GG^{*} as the Fenchel-Legendre dual function of GG i.e. G(u)=supt>0(utG(t))G^{*}(u)=\sup_{t>0}\,(ut-G(t)).
For instance if F(u)F(u) behaves like ln+β(u)\ln_{+}^{\beta}(u) at infinity for some β>0\beta>0, then G(t)G^{*}(t) behaves like βt(β1)/βet1β\beta\,t^{(\beta-1)/\beta}\,e^{t^{\frac{1}{\beta}}} at infinity (see [9] subsection 7.1 to see the correct FF to be chosen).

Proposition 4.3.

Assume that μ\mu satisfies the F-Sobolev inequality (HFS4defect)(HFS4defect) and that G(s)=sF(s)G(s)=sF(s) is convex. Let hh be a non-negative continuous function such that b=2(DF+μ(G(h)))<+b=2(D_{F}+\mu(G^{*}(h)))<+\infty. For ε>0\varepsilon>0, define Uε(h)={(1ε)h>b}U_{\varepsilon}(h)=\{(1-\varepsilon)h>b\}.
Then there exists a Lyapunov function W𝒟(L)W\in\mathcal{D}(L) such that W(x)wε>0W(x)\geq w_{\varepsilon}>0 on Uε(h)U_{\varepsilon}(h) and

LWε2CFhW on Uε(h).LW\leq-\frac{\varepsilon}{2\,C_{F}}\,h\,W\quad\textrm{ on $U_{\varepsilon}(h)$.}
Proof.

The proof mimic the one of proposition 3.1 with the following modifications: take ϕ=ρ(h+b)\phi=\rho(-h+b) with b=2(DF+μ(G(h)))b=2(D_{F}+\mu(G^{*}(h))) and ρCF=12\rho\,C_{F}=\frac{1}{2}, use Young’s inequality stG(s)+G(t)st\leq G(s)+G^{*}(t). ∎

We thus obtain, for a particular choice of FF:

Theorem 4.4.

Assume that DD is not bounded, that VV goes to infinity at infinity and that eaV𝕃1(μ)e^{aV}\in\mathbb{L}^{1}(\mu) for some a>0a>0. Consider the following properties

  • (HFS1)

    There exists a Lyapunov function WW, i.e. there exists a smooth function W:DW:D\to\mathbb{R} with Ww>0W\geq w>0, and there exist constants λ>0\lambda>0 and b>0b>0 such that Wn=0\frac{\partial W}{\partial n}=0 on D\partial D and

    LW(x)λ|V|β(x)W(x)+b.LW(x)\leq-\,\lambda\,|V|^{\beta}(x)\,W(x)\,+\,b\,.
  • (HFS2)

    There exist an open connected bounded subset UU and a constant θ>0\theta>0 such that for all xx,

    𝔼x(exp(0TUθ|V|β(Xs)𝑑s))<+,{\mathbb{E}}_{x}\left(\exp\left(\int_{0}^{T_{U}}\,\theta\,|V|^{\beta}(X_{s})\,ds\right)\right)<+\infty\,,

    where TUT_{U} denotes the hitting time of UU.

  • (HFSβ)(HFS\beta)

    μ\mu satisfies (HFS4defect)(HFS4defect) with F(s)=ln+β(s)F(s)=\ln_{+}^{\beta}(s).

Then (HFSβ)(HFS1)(HFS\beta)\Rightarrow(HFS1) and (HFS1)(HFS2)(HFS1)\Leftrightarrow(HFS2).

If in addition μ\mu is symmetric, σ.σ\sigma.\sigma^{*} is uniformly elliptic and |V(x)|v>0|\nabla V(x)|\geq v>0 for |x||x| large enough, then

(HFSβ)(HFS1)(HFS2).(HFS\beta)\Leftrightarrow(HFS1)\Leftrightarrow(HFS2)\,.

For β1\beta\leq 1 we may replace (HFSβ)(HFS\beta) by its tight version.

Proof.

For the first part we use the previous proposition with h=a|V|βh=a|V|^{\beta} for some aa small enough. In the symmetric situation, we mimic the proof of proposition 3.5 yielding a super-Poincaré inequality with β(s)=ec/s1β\beta(s)=e^{c/s^{\frac{1}{\beta}}} hence the corresponding defective FF-Sobolev inequality using Proposition 4.1. But (HFS1)(HFS1) implies (HP1)(HP1) hence a Poincaré inequality and we can use the final statement of Proposition 4.1 to get a tight version when β1\beta\leq 1. ∎

We know in particular (see [9]) that for V(x)=|x|αV(x)=|x|^{\alpha}, α1\alpha\geq 1, μ\mu satisfies (HFSβ)(HFS\beta) with

β=2(11α).\beta=2\left(1-\frac{1}{\alpha}\right)\,.

The main reason for writing Theorem 4.4 only in the case Fln+βF\sim\ln_{+}^{\beta} is the converse part (HFS1)(HFSβ)(HFS1)\Rightarrow(HFS\beta) for which the argument is easy since we have an explicit expression of GG^{*}. Of course Proposition 4.3 contains much more general situations. Here is one which will be useful in the sequel

Theorem 4.5.

Assume that DD is not bounded, that μ\mu satisfies both a Poincaré inequality and the F-Sobolev inequality (HFS4defect)(HFS4defect), that G(s)=sF(s)G(s)=sF(s) is convex, non decreasing at infinity and that (G)1(G^{*})^{-1} (the inverse function of GG^{*}) growths to infinity at infinity.
Then for all x0Dx_{0}\in D there exist aa and θ\theta two positive constants, such that, defining

h(x)=(G)1(ead(x,x0)),h(x)=(G^{*})^{-1}(e^{ad(x,x_{0})})\,,

we have for all xx and all non empty, open and bounded subset UU,

Wθ,U,h(x)=𝔼x(exp(0TUθh(Xs)𝑑s))<+,W_{\theta,U,h}(x)={\mathbb{E}}_{x}\left(\exp\left(\int_{0}^{T_{U}}\,\theta\,h(X_{s})\,ds\right)\right)<+\infty\,,

where TUT_{U} denotes the hitting time of UU. Actually Wθ,U,h𝕃1(μ)W_{\theta,U,h}\in\mathbb{L}^{1}(\mu).

Proof.

Since μ\mu satisfies a Poincaré inequality, it is known that there exists a>0a>0 such that

μ(G(h))=μ(ead(x,x0))<+.\mu(G^{*}(h))=\mu(e^{ad(x,x_{0})})<+\infty\,.

In addition hh goes to infinity at infinity so that its level sets are compact. It remains to apply Proposition 4.3 to get the Lyapunov function Wθ,U,hW_{\theta,U,h} and consequently the result. As usual since LWθ,U,h+θhWθ,U,h0LW_{\theta,U,h}+\theta hW_{\theta,U,h}\leq 0 outside of a compact set, we get that hWθ,U,hh\,W_{\theta,U,h} is integrable and since hh goes to infinity that Wθ,U,hW_{\theta,U,h} is integrable too. ∎


5. 𝕃p\mathbb{L}^{p} geometric ergodicity and functional inequalities.

Come back to the geometric ergodicity property (HP3)(HP3). If we replace the initial distribution δx\delta_{x} by some initial probability distribution ν\nu, we have

Pt(ν,.)μTVCeβt,\|P_{t}(\nu,.)-\mu\|_{TV}\leq\,C\,e^{-\,\beta\,t}\,,

provided

(LWν)W𝕃1(ν).(LW\nu)\qquad\qquad W\in\mathbb{L}^{1}(\nu)\,. (5.1)

If ν\nu is absolutely continuous w.r.t. μ\mu and dνdμ𝕃p(μ)\frac{d\nu}{d\mu}\in\mathbb{L}^{p}(\mu) a sufficient condition is thus

(LWq)W𝕃q(μ)(LWq)\qquad\qquad W\in\mathbb{L}^{q}(\mu)

for 1p+1q=1\frac{1}{p}+\frac{1}{q}=1. It is thus interesting to study the property (LWq)(LWq).

As shown in [30], once (HP1)(HP1) is satisfied, (LW1)(LW1) is satisfied too. It follows that for some θ>0\theta>0, Wθ(x)=𝔼x(eθTU)W_{\theta}(x)=\mathbb{E}_{x}(e^{\theta T_{U}}) is finite, hence satisfies LWθ=θWθLW_{\theta}=-\theta W_{\theta} on UcU^{c}, so that enlarging a little bit UU (say UεU_{\varepsilon}) we can modify the previous WθW_{\theta} in UεU_{\varepsilon} in order to get a new Lyapunov function, still denoted WθW_{\theta} for simplicity satisfying

LWθθWθ+b𝟏U¯.LW_{\theta}\leq\,-\theta W_{\theta}+b\mathbf{1}_{\bar{U}}\,.

Hence Wθ𝕃1(μ)W_{\theta}\in\mathbb{L}^{1}(\mu). It follows that for every p1p\geq 1, defining Wθ,p(x)=𝔼x(eθpTU)W_{\theta,p}(x)=\mathbb{E}_{x}(e^{\frac{\theta}{p}T_{U}}), we have first that Wθ,p𝕃p(μ)W_{\theta,p}\in\mathbb{L}^{p}(\mu), second that (after similar modifications) Wθ,pW_{\theta,p} is a Lyapunov function with θ\theta replaced by θ/p\theta/p. Since U¯\bar{U} is compact, these modifications do not modify the integrability properties of Wθ,pW_{\theta,p}. Hence we have obtained

Proposition 5.2.

If (HP1)(HP1) is satisfied, for all p>1p>1, one can find another Lyapunov function (associated to a different λ\lambda and a different UU) Wp𝕃p(μ)W_{p}\in\mathbb{L}^{p}(\mu). Hence there exists some βp\beta_{p} such that as soon as dνdμ𝕃q(μ)\frac{d\nu}{d\mu}\in\mathbb{L}^{q}(\mu) with 1p+1q=1\frac{1}{p}+\frac{1}{q}=1,

Pt(ν,.)μTVC(ν)eβpt.\|P_{t}(\nu,.)-\mu\|_{TV}\leq\,C(\nu)\,e^{-\,\beta_{p}\,t}\,.

In the symmetric case the situation is better understood. Indeed, (HP1)(HP1) implies the geometric ergodicity in 𝕃2(μ)\mathbb{L}^{2}(\mu) (HP5)(HP5), so that using Riesz-Thorin interpolation theorem in an appropriate way (see [26]) we have that provided dνdμ𝕃q(μ)\frac{d\nu}{d\mu}\in\mathbb{L}^{q}(\mu) for some 1<q21<q\leq 2,

Pt(dνdμ)1𝕃q(μ)Kqe(q1)qλP(μ)tdνdμ1𝕃q(μ).\left\|P_{t}\left(\frac{d\nu}{d\mu}\right)-1\right\|_{\mathbb{L}^{q}(\mu)}\leq K_{q}\,e^{-\,\frac{(q-1)}{q}\,\lambda_{P}(\mu)\,t}\,\left\|\frac{d\nu}{d\mu}-1\right\|_{\mathbb{L}^{q}(\mu)}\,.

In this situation we thus have geometric convergence for a stronger topology.

But the discussion preceding Proposition 5.2 furnishes a stronger result. Indeed (HP1)(HP1) yields (HP3)(HP3)

H(x)=Pt(x,.)μTVCW(x)eβt,H(x)=\|P_{t}(x,.)-\mu\|_{TV}\leq\,C\,W(x)\,e^{-\,\beta\,t}\,, (5.3)

so that H(x)H(x) converges to 0 for all xx at a geometric rate. But we may replace WW by WpW_{p} and β\beta by βp\beta_{p} and get that actually HH converges to 0 in all 𝕃p(μ)\mathbb{L}^{p}(\mu) for 1p<+1\leq p<+\infty, with a geometric rate depending on pp.

Assume from now on that μ\mu satisfies some FF-Sobolev inequality, for some smooth FF. Does it improve the previous results ? In this situation, the rate of convergence to equilibrium in total variation distance was studied in [24]. The results we proved in the previous sections allow us to give a new and substantially simpler proof of some results contained in [24]. Here is a result in this direction:

Theorem 5.4.

Under the assumptions of Theorem 4.5, there exists λ>0\lambda>0 such that for all non-empty, open and bounded set UU,

W(.)=𝔼.(eλTU)1q<+𝕃q(μ).W(.)=\mathbb{E}_{.}\left(e^{\lambda\,T_{U}}\right)\,\in\,\bigcap_{1\leq q<+\infty}\,\mathbb{L}^{q}(\mu)\,.

It thus follows that there exists some β>0\beta>0 such that for all ν\nu absolutely continuous w.r.t. μ\mu such that dνdμ\frac{d\nu}{d\mu} belongs to 𝕃p(μ)\mathbb{L}^{p}(\mu) for some p>1p>1,

Pt(ν,.)μTVC(ν)eβt.\|P_{t}(\nu,.)-\mu\|_{TV}\leq\,C(\nu)\,e^{-\,\beta\,t}\,.

Actually HH defined in (5.3) converges to 0 in all the 𝕃p\mathbb{L}^{p}’s with a rate eβte^{-\beta\,t}.

Proof.

Let hh as in Theorem 4.5. The level sets HR={hR}H_{R}=\{h\leq R\} of hh are smooth (since FF is smooth) compact sets. Denote by TRT_{R} the hitting time of HRH_{R}. For RR large enough, HRH_{R} contains UU. Let λ>0\lambda>0 be such that W(x)=𝔼x(eλTU)<+W(x)=\mathbb{E}_{x}(e^{\lambda T_{U}})<+\infty for all xx. Such a λ\lambda exists since μ\mu satisfies a Poincaré inequality. If yHR¯y\in\bar{H_{R}}, W(y)K(R)<+W(y)\leq K(R)<+\infty using the regularity of WW. If xHRx\notin H_{R}, we have

𝔼x(eλTU)=𝔼x(eλTR𝔼XTR(eλTU))K(R)𝔼x(eλTR).\mathbb{E}_{x}\left(e^{\lambda T_{U}}\right)=\mathbb{E}_{x}\left(e^{\lambda\,T_{R}}\,\mathbb{E}_{X_{T_{R}}}\left(e^{\lambda\,T_{U}}\right)\right)\leq K(R)\,\mathbb{E}_{x}\left(e^{\lambda\,T_{R}}\right)\,.

But now for θR>qλ\theta R>q\,\lambda,

Wq(x)\displaystyle W^{q}(x) \displaystyle\leq Kq(R)𝔼xq(eλTR)Kq(R)𝔼x(eqλTR)\displaystyle K^{q}(R)\,\mathbb{E}^{q}_{x}\left(e^{\lambda\,T_{R}}\right)\leq K^{q}(R)\,\mathbb{E}_{x}\left(e^{q\,\lambda\,T_{R}}\right)
\displaystyle\leq Kq(R)𝔼x(e0TRθh(Xs)𝑑s)=Kq(R)Wθ,U,h(x).\displaystyle K^{q}(R)\,\mathbb{E}_{x}\left(e^{\int_{0}^{T_{R}}\,\theta\,h(X_{s})\,ds}\right)\,=K^{q}(R)\,W_{\theta,U,h}(x)\,.

The first part of the Theorem follows from Theorem 4.5. The second part is immediate. ∎

Remark 5.5.

If dνdμ𝕃1(μ)\frac{d\nu}{d\mu}\in\mathbb{L}^{1}(\mu), La Vallée-Poussin theorem implies that there is some Young function ϕ(u)=uψ(u)\phi(u)=u\psi(u) with ψ\psi growing to infinity, such that dνdμ\frac{d\nu}{d\mu} belongs to the Orlicz space 𝕃ϕ\mathbb{L}_{\phi}. Hence the second part of the Theorem will follow from the first one and the Hölder-Orlicz inequality, once W𝕃ϕ(μ)W\in\mathbb{L}_{\phi^{*}}(\mu). But for the previous proof to work we need something like ϕ(eλu)Ceϕ(λ)u\phi^{*}(e^{\lambda u})\leq C\,e^{\phi**(\lambda)\,u} which is not true for ϕ\phi^{*} growing faster than a power function. That is why the result is only stated for ν\nu with a density belonging to some 𝕃p\mathbb{L}^{p} space. \diamondsuit

This result is nor new nor surprising. For instance, when a logarithmic Sobolev inequality holds true, Theorem 2.13 in [24] shows that there is exponential convergence to the equilibrium in total variation distance as soon as ν\nu belongs to the space 𝕃ln𝕃\mathbb{L}\,\ln\mathbb{L}. It is then a simple consequence of Pinsker’s inequality and the entropic convergence to equimibrium. More generally, the courageous reader will find in the jungle of section 3 of [24] similar results for general FF-Sobolev inequalities. It should be interesting to recover these results by using Lyapunov functions.

Actually we should describe the problem as follows: we know that reinforcing functional inequalities from Poincaré to FF-Sobolev, reinforces the Lyapunov condition and conversely in the symmetric case. Does a reinforced functional inequality reinforce the integrability of the Lyapunov function and conversely in the symmetric case ? In particular can we characterize (at least in the symmetric case) a functional inequality through integrability properties of (some) Lyapunov function ?

Example 5.6.

Look at the simple symmetric case L=ΔV.L=\Delta-\nabla V.\nabla in the whole d\mathbb{R}^{d}. Then it is easily seen that the following holds: there is an equivalence between

  • x.V(x)α|x|2x.\nabla V(x)\geq\alpha\,|x|^{2} for large |x||x|,

  • We(x)=eα|x|2/2W_{e}(x)=e^{\alpha\,|x|^{2}/2} is a Lyapunov function, i.e. satisfies (HP1)(HP1),

  • W2(x)=|x|2W_{2}(x)=|x|^{2} also satisfies (HP1)(HP1).

The first item is of course equivalent to the fact that VV is uniformly convex. We see that the behavior of various Lyapunov functions can be very different. They also imply various integrability properties and functional inequalities. Notice however that thanks to what we said previously we can directly make the following reasoning: if VV is uniformly convex, W2W_{2} is a Lyapunov function and admits some exponential moment so that convergence to the equilibrium in total variation distance holds as soon as ν\nu belongs to the space 𝕃ln𝕃\mathbb{L}\,\ln\mathbb{L} as expected. Indeed, we have by (HP1)(HP1) that for some δ\delta and some rr (denoting BrB_{r} the euclidean ball of radius rr, and xBrcx\in B_{r}^{c}

𝔼x(eδTBr)W2(x).\mathbb{E}_{x}(e^{\delta T_{B_{r}}})\leq W_{2}(x).

Note however that in fact WeW_{e} satisfies a stronger Lyapunov condition, i.e.

LWe(x)λ|x|2We(x)+b1BrLW_{e}(x)\leq-\lambda|x|^{2}\,W_{e}(x)+b1_{B_{r}}

so that by the previous result on logarithmic , we have a stroonger integrability property linked to hitting times:

𝔼x(eδ0TBrXs2)We(x).\mathbb{E}_{x}\left(e^{\delta\int_{0}^{T_{B_{r}}}X_{s}^{2}}\right)\leq W_{e}(x).

However, it does not seem possible to pass from this last control to the previous one.
Hence, finally, these results do not give any precise idea of the dependence in xx of 𝔼x(eλTU)\mathbb{E}_{x}(e^{\lambda T_{U}}) for a given bounded UU. Actually this is a very difficult problem for which results are only known in the gaussian case (i.e. for the Ornstein-Uhlenbeck process). \diamondsuit

In the next section we shall look more into details at the case where one can find a bounded Lyapunov function.

6. Coming down from infinity, uniform geometric ergodicity and Lyapunov functions.

If dν/dμd\nu/d\mu only belongs to 𝕃1\mathbb{L}^{1}, it is interesting to look at a bounded Lyapunov function. What precedes allows us to state

Proposition 6.1.

The following statements are equivalent:

  • there exists λ>0\lambda>0 such that

    supxD¯𝔼x(eλTU)<+\sup_{x\in\bar{D}}\,\mathbb{E}_{x}\left(e^{\lambda T_{U}}\right)\,<\,+\infty

    for one (or all) non empty, open and bounded set UU,

  • there exists a bounded Lyapunov function satisfying (HP1)(HP1),

  • the process is uniformly geometrically ergodic, i.e. there exist β>0\beta>0 and C>0C>0 such that

    supxD¯Pt(x,.)μTVCeβt.\sup_{x\in\bar{D}}\,\|P_{t}(x,.)-\mu\|_{TV}\leq\,C\,e^{-\,\beta\,t}\,.

In this case of course, for any initial probability measure ν\nu,

Pt(ν,.)μTVCeβt.\|P_{t}(\nu,.)-\mu\|_{TV}\leq\,C\,e^{-\,\beta\,t}\,.

There exists a stronger form of uniform exponential integrability, the notion of “coming down from infinity” which is used by people who are studying quasi-stationary distributions or more precisely Yaglom limits (see e.g. the recent book [32]). We shall use the following definition

Definition 6.2.

We say that the process comes down from infinity if for all a>0a>0 there exists some open, bounded subset UaU_{a} such that supx𝔼x(eaTUa)<+\sup_{x}\,\mathbb{E}_{x}\left(e^{a\,T_{U_{a}}}\right)\,<\,+\infty.

In one dimension, this property was related to the uniqueness of quasi-stationary distributions (QSD) and to the fact that \infty is an entrance boundary, in [19]. Uniqueness of a (QSD) also follows from the ultraboundedness property of the semi-group, even in higher dimension (see e.g. [31]). In [30] Proposition 5.3, we claimed that ultraboundedness is actually equivalent to coming down from infinity for one dimensional diffusion processes with generator ΔV.\Delta-\nabla V.\nabla satisfying some extra condition. D. Loukianova pointed out to us that the proof of this proposition in fact needs slightly more stringent assumptions (the function zF(z)/zz\mapsto F(z)/z therein is not necessarily non-increasing) for this equivalence to hold true.

Nevertheless, part of this result is true and we shall give a direct and simple proof.

Indeed, assume that WW satisfies (HP1)(HP1). We have seen that we can always assume that W𝕃2(μ)W\in\mathbb{L}^{2}(\mu) and then W𝒟(L))W\in\mathcal{D}(L)). It follows that

L(PtW)=Pt(LW)λPtW+Pt(𝟏U¯).L(P_{t}W)=P_{t}(LW)\leq-\lambda\,P_{t}W+P_{t}(\mathbf{1}_{\bar{U}})\,. (6.3)

Now assume that Rd(x,U¯)2RR\leq d(x,\bar{U})\leq 2R. We have

Pt(𝟏U¯)(x)\displaystyle P_{t}(\mathbf{1}_{\bar{U}})(x) =\displaystyle= 𝔼x(𝟏XtU¯)\displaystyle\mathbb{E}_{x}\left(\mathbf{1}_{X_{t}\in\bar{U}}\right)
\displaystyle\leq 𝔼x(𝟏TU¯<t)supd(y,U¯)Ry(TU¯<t)\displaystyle\mathbb{E}_{x}\left(\mathbf{1}_{T_{\bar{U}}<t}\right)\leq\sup_{d(y,\bar{U})\geq R}\,\mathbb{Q}_{y}(T_{\bar{U}}<t)

where QyQ_{y} denotes the law of the process Y.Y_{.} with the same generator LL but reflected on d(z,U¯)=2Rd(z,\bar{U})=2R (which can be assumed to be smooth) and starting from yy. Indeed if the process X.X_{.} hits U¯\bar{U} before to leave {d(z,U¯)2R}\{d(z,\bar{U})\leq 2R\}, it coincides with Y.Y_{.} (with the same starting point) up to TU¯T_{\bar{U}}. If not, X.X_{.} leaves {d(z,U¯)2R}\{d(z,\bar{U})\leq 2R\} before TU¯T_{\bar{U}}, but in order to hit U¯\bar{U} it has to come back to {d(z,U¯)2R}\{d(z,\bar{U})\leq 2R\} first, so that using the Markov property we may apply the same argument as before this time starting from XTRX_{T_{R}} where TRT_{R} is the hitting time of {d(z,U¯)2R}\{d(z,\bar{U})\leq 2R\}. That is why the final upper bound contains the supremum over yy.
Now, since all coefficients are smooth, they are bounded with bounded derivatives of any order in {d(z,U¯)2R}\{d(z,\bar{U})\leq 2R\} which is compact. It is then well known that

supd(y,U¯)Ry(TU¯<t)CecR/t\sup_{d(y,\bar{U})\geq R}\mathbb{Q}_{y}(T_{\bar{U}}<t)\leq C\,e^{-cR/t}

for some constants CC and cc only depending on these bounds. Hence

Pt(𝟏U¯)(x)CecR/t,P_{t}(\mathbf{1}_{\bar{U}})(x)\leq C\,e^{-cR/t}\,,

as soon as Rd(x,U¯)2RR\leq d(x,\bar{U})\leq 2R. If d(x,U¯)>2Rd(x,\bar{U})>2R, TU>TRT_{U}>T_{R}, and we may apply again the Markov property to get the same upper bound.
Pick some R>0R>0 once for all and choose t>0t>0 in such a way that

CecR/t<12λw.Ce^{-cR/t}<\frac{1}{2}\,\lambda\,w\,.

We thus have

L(PtW)(x)\displaystyle L(P_{t}W)(x) \displaystyle\leq λPtW(x)+𝟏d(x,U¯)R+CecR/t 1d(x,U¯)R\displaystyle-\lambda\,P_{t}W(x)+\mathbf{1}_{d(x,\bar{U})\leq R}+Ce^{-cR/t}\,\mathbf{1}_{d(x,\bar{U})\geq R}
\displaystyle\leq λ2PtW(x)+𝟏d(x,U¯)R,\displaystyle-\frac{\lambda}{2}\,P_{t}W(x)+\mathbf{1}_{d(x,\bar{U})\leq R}\,,

so that PtWP_{t}W is a new Lyapunov function with λ/2\lambda/2 and {d(x,U¯)R}\{d(x,\bar{U})\leq R\} in place of λ\lambda and U¯\bar{U} (of course PtWP_{t}W belongs to 𝒟(L)\mathcal{D}(L) and satisfies PtWw>0P_{t}W\geq w>0).
We deduce immediately

Theorem 6.4.

Assume that (HP1)(HP1) is satisfied (for example μ\mu satisfies a Poincaré inequality) and that the semi-group P.P_{.} is ultra-bounded, i.e. that PtP_{t} maps continuously 𝕃1\mathbb{L}^{1} into 𝕃\mathbb{L}^{\infty} for any t>0t>0. Then the process comes down from infinity.

Proof.

(HP1)(HP1) together with ultra-boundedness imply that the semi-group is hyper-contractive so that μ\mu satisfies a logarithmic Sobolev inequality according to Gross theorem. Hence, we may apply Corollary 3.3 and find a Lyapunov function WW such that

LWλd2(.,x0)W+b 1U¯LW\leq-\lambda\,d^{2}(.,x_{0})\,W+b\,\mathbf{1}_{\bar{U}}

for some bounded open subset UU. Hence for all large enough a>0a>0,

LW2aW+ba 1UaLW\leq-2a\,W\,+b_{a}\,\mathbf{1}_{U_{a}}

with Ua={x;λd2(x,x0)2a}U_{a}=\{x\,;\,\lambda\,d^{2}(x,x_{0})\,\leq 2a\} and ba=supxUaLWb_{a}=\sup_{x\in U_{a}}\,LW, aa being large enough for U¯Ua\bar{U}\subset U_{a}. According to the previous discussion there exists a new Lyapunov function Wa=PtWW_{a}=P_{t}W for some adequat tt which is bounded and satisfies LWa(x)aWa(x)LW_{a}(x)\leq-aW_{a}(x) for xVa={d(x,Ua)>R}x\in V_{a}=\{d(x,U_{a})>R\}, so that

supx𝔼x(eaTa)<+ for Ta the hitting time of Vac.\sup_{x}\,\mathbb{E}_{x}\left(e^{aT_{a}}\right)<+\infty\quad\textrm{ for $T_{a}$ the hitting time of $V_{a}^{c}$.}

Remark 6.5.

The fact that PtWP_{t}W is still a Lyapunov function is interesting by itself, but except in the ultra-bounded situation, it does not furnish new results. For instance if we want to get a Lyapunov function that belongs to all the 𝕃p\mathbb{L}^{p} spaces, we have to assume that P.P_{.} is immediately hyper bounded, which is stronger than log-Sobolev, while Theorem 5.4 gives the result under a simple FF-Sobolev condition. Working a little bit more, one can extend the previous result to discrete valued Markov process, which are ultracontractive (birth-death processes,…). \diamondsuit

The previous proof indicates a way to prove “coming down from infinity” using Lyapunov functions, more precisely nested Lyapunov conditions

Definition 6.6.

We shall say that a Super-Lyapunov condition is satisfied if there exist a sequence Wk1W_{k}\geq 1, a sequence of increasing bounded sets BkB_{k} growing to n{\mathbb{R}}^{n}, an increasing sequence λk>0\lambda_{k}>0 growing to infinity and a sequence bk>0b_{k}>0 such that

LWkλkWk+bk𝟏Bk.LW_{k}\leq-\lambda_{k}\,W_{k}+b_{k}\mathbf{1}_{B_{k}}.

For instance in the situation of Theorem 4.5, (SLC)(SLC) is satisfied with the same Lyapunov function WW, i.e. W=WkW=W_{k}, similarly as what we have done (in the case of log-Sobolev) in the previous proof. Of course if all the WkW_{k} are bounded, (SLC)(SLC) is equivalent to “coming down from infinity”.

Assume that (SLC)(SLC) is satisfied. Denote by TkT_{k} the hitting time of BkB_{k} and choose some δ>0\delta>0. As we did in the proof of Theorem 5.4, for kk0k\geq k_{0} and λk0>δ\lambda_{k_{0}}>\delta we have for xBkx\notin B_{k},

𝔼x(eδTk0)\displaystyle\mathbb{E}_{x}\left(e^{\delta T_{k_{0}}}\right) =\displaystyle= 𝔼x(eδTk𝔼XTk(eδTk0))𝔼x(eδTk)supyBk𝔼y(eδTk0)\displaystyle\mathbb{E}_{x}\left(e^{\delta\,T_{k}}\,\mathbb{E}_{X_{T_{k}}}\left(e^{\delta\,T_{k_{0}}}\right)\right)\leq\,\mathbb{E}_{x}\left(e^{\delta\,T_{k}}\right)\sup_{y\in\partial B_{k}}\,\mathbb{E}_{y}\left(e^{\delta T_{k_{0}}}\right)
\displaystyle\leq (𝔼x(eλkTk))δ/λksupyBk𝔼y(eδTk0)\displaystyle\left(\mathbb{E}_{x}\left(e^{\lambda_{k}\,T_{k}}\right)\right)^{\delta/\lambda_{k}}\,\sup_{y\in\partial B_{k}}\,\mathbb{E}_{y}\left(e^{\delta T_{k_{0}}}\right)
\displaystyle\leq (Wk(x))δ/λksupyBk𝔼y(eδTk0).\displaystyle(W_{k}(x))^{\delta/\lambda_{k}}\,\sup_{y\in\partial B_{k}}\,\mathbb{E}_{y}\left(e^{\delta T_{k_{0}}}\right)\,.

Define

wk=supyBk+1BkWk(y).w_{k}=\sup_{y\in B_{k+1}-B_{k}}\,W_{k}(y)\,.

Proceeding by induction we have for all xBk0+1x\notin B_{k_{0}+1}

𝔼x(eλTk0)Ck0k=k0+(wk)δλk.\mathbb{E}_{x}\left(e^{\lambda T_{k_{0}}}\right)\leq C_{k_{0}}\prod_{k=k_{0}}^{+\infty}(w_{k})^{\frac{\delta}{\lambda_{k}}}. (6.7)

As we see, if the sequence λk\lambda_{k} goes to infinity, δ\delta has no role in the convergence of the previous infinite product, as well as the value of k0k_{0}. We thus have

Theorem 6.8.

If (SLC)(SLC) is satisfied, and for some k0k_{0},

k=k0+lnwkλk<+,\sum_{k=k_{0}}^{+\infty}\,\frac{\ln w_{k}}{\lambda_{k}}<+\infty\,,

the process comes down from infinity. Hence (SLC)(SLC) is satisfied with another sequence (λk,Wk,Bk)(\lambda_{k},W_{k},B_{k}) where all the WkW_{k} are bounded.

Example 6.9.

In one dimension consider (for β>0\beta>0) the potential

Vβ(x)=(1+x2)lnβ(1+x2).V_{\beta}(x)=(1+x^{2})\,\ln^{\beta}(1+x^{2})\,.

For W(x)=ex2/2W(x)=e^{x^{2}/2} it holds for |x||x| large enough,

LβW(x)(1ε)x2lnβ(1+x2)W(x).L_{\beta}W(x)\leq-\,(1-\varepsilon)\,x^{2}\,\ln^{\beta}(1+x^{2})\,W(x)\,.

Hence if Bk=B(0,Rk)B_{k}=B(0,R_{k}) we have ln(wk)12Rk2\ln(w_{k})\leq\frac{1}{2}\,R_{k}^{2}, while for |x|>Rk|x|>R_{k}, we may choose λk=cRk2lnβ(Rk)\lambda_{k}=c\,R_{k}^{2}\,\ln^{\beta}(R_{k}). It follows that the process comes down from infinity for any β>0\beta>0 by choosing for instance Rk=expk2/βR_{k}=\exp k^{2/\beta} in Theorem 6.8.
It is known however that the semi-group is ultra-bounded if and only if β>1\beta>1 (see e.g [41]). Hence we have some examples of processes coming down from infinity for which the semi-group is not ultra-bounded (but is immediately hyper-contractive as shown in [41]).


7. Integrability and Lyapunov conditions.

In this section we will study the interplay between Lyapunov conditions and integrability conditions. So it is not a restriction to look at L=ΔV.L=\Delta-\nabla V.\nabla such that μ\mu is symmetric (or reversible).

Let us recall that integrability conditions are related to transportation inequalities. Indeed, if μ\mu satisfies a Gaussian integrability condition, i.e.

δ>0,x0eδd(x,x0)2𝑑μ<\exists\delta>0,x_{0}\qquad\int e^{\delta d(x,x_{0})^{2}}d\mu<\infty

then

(T1)ν𝒫1,W1(ν,μ)2CH(ν|μ)(T1)\quad\forall\nu\in{\mathcal{P}}_{1},\qquad W_{1}(\nu,\mu)\leq\sqrt{2CH(\nu|\mu)}

for some explicit CC (see [34]), where W1W_{1} and HH denote respectively the 11-Wasserstein distance and the relative entropy (or Kullback-Leibler information). The converse, (T1)(T1) implies gaussian integrability is also true and due to K. Marton [51]. Gozlan [38] has generalized the approach to various type of (T1)(T1) type transportation inequalities, getting that integrability property

δ>0,x0eα(δd(x,x0))𝑑μ<\exists\delta>0,x_{0}\qquad\int e^{\alpha(\delta d(x,x_{0}))}d\mu<\infty

for α\alpha convex (quadratic near 0) is equivalent to α(W1(ν,μ)CH(ν|μ)\alpha(W_{1}(\nu,\mu)\leq C\,H(\nu|\mu) for very ν\nu. The links between functional inequalities and integrability properties are usually consequences of concentration properties of lipschitzian function (as Herbst argument for logarithmic Sobolev inequality). A more direct approach is taken in [22], where in particular an equivalence between a Poincaré inequality and a weak form of a (T2)(T2) inequality (involving W2W_{2}) is shown. Note however that we won’t require here some local inequality, so that we cannot hope to get better functional inequalities than (T1)(T1) like.

We shall now investigate the relationship between the existence of Lyapunov functions and integrability properties. In the sequel we will thus assume the following

Definition 7.1.

Let ϕ\phi be a (strictly) positive function. We say that (ϕLyap)(\phi-Lyap) is satisfied if there exist W>0W>0, b0b\geq 0 and a bounded open subset CC such that

LWϕ2W+b 1C.LW\leq-\,\phi^{2}\,W+b\,\mathbf{1}_{C}\,.

If we define U=ln(W)U=\ln(W) it is immediately seen that

LU+|U|2ϕ2+bminC¯W 1C.LU\,+\,|\nabla U|^{2}\,\leq\,-\phi^{2}+\,\frac{b}{\min_{\bar{C}}W}\,\mathbf{1}_{C}\,. (7.2)

Using this inequality it is shown in [28] lemma 3.4 that for any smooth hh,

h2ϕ2𝑑μ(h)+bminC¯WCh2𝑑μ.\int h^{2}\,\phi^{2}d\mu\,\leq\,\mathcal{E}(h)\,+\,\frac{b}{\min_{\bar{C}}W}\,\int_{C}\,h^{2}\,d\mu\,. (7.3)

Of course, if Ch𝑑μ=0\int_{C}\,h\,d\mu=0, and provided CC is smooth enough (which is not a restriction), we may apply the Holley-Stroock perturbation argument and get

h2ϕ2𝑑μ(1+bminC¯WeOscC¯VCP(C,dx))(h)\int h^{2}\,\phi^{2}d\mu\,\leq\,\left(1+\frac{b}{\min_{\bar{C}}W}\,e^{Osc_{\bar{C}}V}\,C_{P}(C,dx)\right)\mathcal{E}(h)

where OscU(V)Osc_{U}(V) denotes the oscillation of VV on the subset UU and CP(U,dx)C_{P}(U,dx) the Poincaré constant of the uniform measure on UU. If ϕ\phi goes to infinity at infinity, the previous inequality is thus stronger than the Poincaré inequality (as expected). It is well known that the Poincaré inequality implies the exponential integrability of cd(.,x0)c\,d(.,x_{0}) for some small enough positive cc, as the logarithmic Sobolev inequality implies the exponential integrability of cd2(.,x0)c\,d^{2}(.,x_{0}) for some small enough positive cc (we previously recalled this result).

In [47], Yuan Liu proved that (ϕLyap)(\phi-Lyap) with ϕ(x)=ad(x,x0)\phi(x)=a\,d(x,x_{0}) implies the exponential integrability of cd2(.,x0)c\,d^{2}(.,x_{0}) for c<ac<a. This result is not surprising since in this case (ϕLyap)(\phi-Lyap) is exactly (HLS1)(HLS1) which is equivalent to the logarithmic-Sobolev inequality, at least if the curvature is bounded from below according to Theorem 1.2, hence implies as a consequence the quoted exponential integrability.

We shall here follow and generalize Liu’s argument in order to answer the following question: what are sufficient conditions on ψ\psi for cψ2c\,\psi^{2} to be exponentially integrable when (ϕLyap)(\phi-Lyap) is satisfied ?

A particularly interesting example is the case when ϕ(x)=adp(x,x0)\phi(x)=a\,d^{p}(x,x_{0}). Indeed, for p=0p=0 we know that a Poincaré inequality is satisfied and for p=1p=1 a log-Sobolev inequality is satisfied under a curvature assumption. The use of curvature is very specific and strongly connected to p=1p=1. Hence for 0<p<10<p<1 we do not know whether (ϕLyap)(\phi-Lyap) implies some natural functional inequality or not, even for bounded from below curvature. We shall see however that cdp+1(.,x0)c\,d^{p+1}(.,x_{0}) is exponentially integrable for some small enough positive cc and all 0p0\leq p, so that a generalized transportation inequality holds.

For ψ0\psi\geq 0, introduce

βn:=ψ2n𝑑μ.\beta_{n}:=\int\psi^{2n}d\mu\,.

We will use (7.3) to initiate a recurrence on βn\beta_{n} using the notation b¯=bminC¯W\bar{b}=\frac{b}{\min_{\bar{C}}W},

βn\displaystyle\beta_{n} =\displaystyle= ψ2nϕ2ϕ2𝑑μ\displaystyle\int\frac{\psi^{2n}}{\phi^{2}}\phi^{2}d\mu
\displaystyle\leq |(ψnϕ)|2𝑑μ+b¯Cψ2nϕ2𝑑μ.\displaystyle\int\left|\nabla\left(\frac{\psi^{n}}{\phi}\right)\right|^{2}d\mu+\bar{b}\int_{C}\frac{\psi^{2n}}{\phi^{2}}d\mu.

Let us focus on the first term

|(ψnϕ)|2𝑑μ\displaystyle\int\left|\nabla\left(\frac{\psi^{n}}{\phi}\right)\right|^{2}d\mu =\displaystyle= |nψn1ψϕψnϕϕ2|2𝑑μ\displaystyle\int\left|n\psi^{n-1}\frac{\nabla\psi}{\phi}-\psi^{n}\frac{\nabla\phi}{\phi^{2}}\right|^{2}\,d\mu
=\displaystyle= n2ψ2|ψ|2ϕ2ψ2(n2)+2nψϕ.ψϕ3ψ2(n1)𝑑μ+|ϕ|2ϕ4ψ2n𝑑μ.\displaystyle n^{2}\int\frac{\psi^{2}|\nabla\psi|^{2}}{\phi^{2}}\,\psi^{2(n-2)}+2n\int\frac{\psi\nabla\phi.\nabla\psi}{\phi^{3}}\psi^{2(n-1)}d\mu+\int\frac{|\nabla\phi|^{2}}{\phi^{4}}\psi^{2n}d\mu.

Let us assume then that there exists α,β,γ>0\alpha,\beta,\gamma>0 and 0<δ<10<\delta<1, (at least outside a compact KK, and if so choose ψ\psi to be 0 on KK) such that

ψ2|ψ|2ϕ2α,|ψϕ.ψϕ3|β,|ϕ|2ϕ4δ,supCψ2ϕ2γ.\frac{\psi^{2}|\nabla\psi|^{2}}{\phi^{2}}\leq\alpha,\quad\left|\frac{\psi\nabla\phi.\nabla\psi}{\phi^{3}}\right|\leq\beta,\quad\frac{|\nabla\phi|^{2}}{\phi^{4}}\leq\delta,\quad\sup_{C}\frac{\psi^{2}}{\phi^{2}}\leq\gamma\,. (7.4)

Under these assumptions, we get that

βnα1δn2βn2+2nβ+γb¯1δβn1.\beta_{n}\leq\frac{\alpha}{1-\delta}n^{2}\beta_{n-2}+\frac{2n\beta+\gamma\bar{b}}{1-\delta}\beta_{n-1}\,.

Combined with a direct consequence of Cauchy-Schwarz inequality, we obtain

βnβn+1βn1(α1δ(n+1)2βn1+2(n+1)β+γb¯1δβn)12βn112.\beta_{n}\leq\sqrt{\beta_{n+1}\beta_{n-1}}\leq\left(\frac{\alpha}{1-\delta}(n+1)^{2}\beta_{n-1}+\frac{2(n+1)\beta+\gamma\bar{b}}{1-\delta}\beta_{n}\right)^{\frac{1}{2}}\beta_{n-1}^{\frac{1}{2}}.

We then easily deduce that

βn12(2(n+1)β+γb¯1δ+(2(n+1)β+γb¯1δ)2+4α1δ(n+1)2)βn1anβn1\beta_{n}\leq\frac{1}{2}\left(\frac{2(n+1)\beta+\gamma\bar{b}}{1-\delta}+\sqrt{\left(\frac{2(n+1)\beta+\gamma\bar{b}}{1-\delta}\right)^{2}+4\frac{\alpha}{1-\delta}(n+1)^{2}}\right)\beta_{n-1}\leq an\,\beta_{n-1}

for

a>12(2β1δ+4β2(1δ)2+4α1δ)a>\frac{1}{2}\left(\frac{2\beta}{1-\delta}+\sqrt{\frac{4\beta^{2}}{(1-\delta)^{2}}+4\frac{\alpha}{1-\delta}}\right)

and nn large enough. It then follows that for some cc

βncann!\beta_{n}\leq ca^{n}n!

and thus we have that for a<a1a^{\prime}<a^{-1}

eaψ2𝑑μc1aa.\int e^{a^{\prime}\psi^{2}}d\mu\leq\frac{c}{1-a^{\prime}a}.

Now come back to (7.4). I f we assume that ϕ\phi is bounded from below by some positive constant in CC, the condition on γ\gamma is satisfied as soon as ϕ\phi and ψ\psi are, say, continuous. The condition on δ\delta says that 1/ϕ1/\phi is a contraction outside some compact set. Assuming this condition we see that both conditions on α\alpha and β\beta are the same, i.e. |(ψ2)|/ϕ|\nabla(\psi^{2})|/\phi is bounded. Thus

Theorem 7.5.

Assume that (ϕLyap)(\phi-Lyap) is satisfied for some function ϕ\phi such that ϕ\phi is bounded from below by a positive constant on CC and 1/ϕ1/\phi is η\eta-Lipschitz for some η<1\eta<1. Then for all function ψ2\psi^{2} such that |(ψ2)|/ϕ|\nabla(\psi^{2})|/\phi is bounded, there exists a>0a^{\prime}>0 such that eaψ2𝑑μ<+\int e^{a^{\prime}\psi^{2}}d\mu<+\infty.

Example 7.6.
  1. (1)

    For a constant ϕ\phi we recover the fact that any Lipschitz function has an exponential moment, hence the usual concentration result once a Poincaré inequality is satisfied. For ϕ(x)=ad2(x,x0)\phi(x)=ad^{2}(x,x_{0}), we recover the gaussian nature of the tails once a log-Sobolev inequality is satisfied.

  2. (2)

    If we take ϕ(x)=adp(x,x0)\phi(x)=ad^{p}(x,x_{0}) for p0p\geq 0 we obtain exponential integrability of functions gg such that |g|C|x|p|\nabla g|\leq C\,|x|^{p}, hence for instance for g(x)=dp+1(x,x0)g(x)=d^{p+1}(x,x_{0}). As for the Poincaré or the log-Sobolev case, this exponential integrability is sharp, since for μ(dx)=Zpe|x|p+1\mu(dx)=Z_{p}\,e^{-|x|^{p+1}}, (ϕLyap)(\phi-Lyap) is satisfied with ϕ(x)=c|x|p\phi(x)=c|x|^{p} and W(x)=ea|x|pW(x)=e^{a|x|^{p}} for a small enough aa.

  3. (3)

    One may of course also consider ϕ2(x)=ad1(x,x0)\phi^{2}(x)=ad^{-1}(x,x_{0}), if a>1a>1, which can be obtained for Cauchy type measures (see [21] for details). In this case we may take ψ2\psi^{2} behaving as log(|x|)\log(|x|) at infinity recovering polynomial integrability (also quite sharply). Note also that in [21, Th. 5.1], it is shown how a converse Poincaré inequality (obtained by a (ϕLyap)(\phi-Lyap) condition and a local Poincaré inequality) plus an involved integrability condition implies a weak Poincaré inequality. This integrability condition can be checked using Th. 7.5 for example for Cauchy type measure (via tedious computations), so that only a Lyapunov condition and local inequality are also sufficient for weak Poincaré inequality.

Of course there is no converse statement for Theorem 7.5, since for instance exponential integrability for the distance cannot imply a Poincaré inequality (disconnected domains for example).

Finally recall that that (ϕLyap)(\phi-Lyap) is equivalent to the following

𝔼x(e0TCϕ2(Xs)𝑑s)<+,\mathbb{E}_{x}\left(e^{\int_{0}^{T_{C}}\,\phi^{2}(X_{s})\,ds}\right)<+\infty\,,

where in our situation

Xt=X0+2Wt0tV(Xs)𝑑s.X_{t}=X_{0}+\sqrt{2}\,W_{t}-\int_{0}^{t}\,\nabla V(X_{s})\,ds\,.

It would be particularly interesting to show that weak Poincaré inequality implies back Lyapunov condition as done in section 2 for Poincaré inequality.

References

  • [1] S. Aida. Uniform positivity improving property, Sobolev inequalities and spectral gaps. J. Funct. Anal., 158:152–185, 1998.
  • [2] S. Aida, T. Masuda, and I. Shigekawa. Logarithmic Sobolev inequalities and exponential integrability. J. Funct. Anal., 126(1):83–101, 1994.
  • [3] C. Ané, S. Blachère, D. Chafaï, P. Fougères, I. Gentil, F. Malrieu, C. Roberto, and G. Scheffer. Sur les inégalités de Sobolev logarithmiques, volume 10 of Panoramas et Synthèses. Société Mathématique de France, Paris, 2000.
  • [4] D. Bakry, F. Barthe, P. Cattiaux, and A. Guillin. A simple proof of the Poincaré inequality for a large class of probability measures. Electronic Comm. in Probab., 13:60–66, 2008.
  • [5] D. Bakry, P. Cattiaux, and A. Guillin. Rate of convergence for ergodic continuous Markov processes : Lyapunov versus Poincaré. J. Func. Anal., 254:727–759, 2008.
  • [6] D. Bakry and M. Émery. Diffusions hypercontractives. In Séminaire de probabilités, XIX, 1983/84, volume 1123 of Lecture Notes in Math., pages 177–206. Springer, Berlin, 1985.
  • [7] D. Bakry, I. Gentil, and M. Ledoux. Analysis and Geometry of Markov Diffusion Operators, volume 348 of Grundlehren der mathematischen Wissenschaften. Springer, 2014.
  • [8] F. Barthe, P. Cattiaux, and C. Roberto. Concentration for independent random variables with heavy tails. AMRX, 2005(2):39–60, 2005.
  • [9] F. Barthe, P. Cattiaux, and C. Roberto. Interpolated inequalities between exponential and Gaussian, Orlicz hypercontractivity and isoperimetry. Rev. Mat. Iber., 22(3):993–1066, 2006.
  • [10] F. Barthe, P. Cattiaux, and C. Roberto. Isoperimetry between exponential and Gaussian. Electronic J. Prob., 12:1212–1237, 2007.
  • [11] F. Barthe and A. V. Kolesnikov. Mass transport and variants of the logarithmic Sobolev inequality. J. Geom. Anal., 18(4):921–979, 2008.
  • [12] S. G. Bobkov, I. Gentil, and M. Ledoux. Hypercontractivity of Hamilton-Jacobi equations. J. Math. Pu. Appli., 80(7):669–696, 2001.
  • [13] S. G. Bobkov and F. Götze. Exponential integrability and transportation cost related to logarithmic Sobolev inequalities. J. Funct. Anal., 163(1):1–28, 1999.
  • [14] S.G. Bobkov and M. Ledoux. Poincaré’s inequalities and Talagrand s concentration phenomenon for the exponential distribution. Prob. Theor. Rel. Fields, 107:383–400, 1997.
  • [15] J. M. Bony. Principe du maximum, inégalité de Harnack et unicité du problème de Cauchy pour les opérateurs elliptiques dégénérés. Ann. Inst. Fourier, 19(1):277–304, 1969.
  • [16] R. Carmona and A. Klein. Exponential moments for hitting times of uniformly ergodic Markov processes. Ann. Prob., 11:648–655, 1983.
  • [17] P. Cattiaux. Stochastic calculus and degenerate boundary value problems. Ann. Inst. Fourier, 42(3):541–624, 1992.
  • [18] P. Cattiaux. A pathwise approach of some classical inequalities. Potential Analysis, 20:361–394, 2004.
  • [19] P. Cattiaux, P. Collet, A. Lambert, S. Martinez, S. Méléard, and J. San Martin. Quasi-stationarity distributions and diffusion models in population dynamics. Ann. Probab., 37:1926–1969, 2009.
  • [20] P. Cattiaux, I. Gentil, and A. Guillin. Weak logarithmic Sobolev inequalities and entropic convergence. Probab. Theory Related Fields, 139(3-4):563–603, 2007.
  • [21] P. Cattiaux, N. Gozlan, A. Guillin, and C. Roberto. Functional inequalities for heavy tailed distributions and applications to isoperimetry. Elec. J. Prob., 15:346–385, 2010.
  • [22] P. Cattiaux and A. Guillin. On quadratic transportation cost inequalities. J. Math. Pures Appl., 88(4):341–361, 2006.
  • [23] P. Cattiaux and A. Guillin. Deviation bounds for additive functionals of Markov processes. ESAIM Probability and Statistics, 12:12–29, 2008.
  • [24] P. Cattiaux and A. Guillin. Trends to equilibrium in total variation distance. Ann. Inst. Henri Poincaré Probab. Stat., 45(1):117–145, 2009.
  • [25] P. Cattiaux and A. Guillin. Functional inequalities via Lyapunov conditions. In Optimal transportation. Theory and applications., volume 413 of London Mathematical Society Lecture Notes Series, pages 274–287. Cambridge Univ. Press,, 2014.
  • [26] P. Cattiaux, A. Guillin, and C. Roberto. Poincaré inequality and the lpl^{p} convergence of semi-groups. Elec. Comm. in Prob., 15:270–280, 2010.
  • [27] P. Cattiaux, A. Guillin, F.Y. Wang, and L. Wu. Lyapunov conditions for super Poincaré inequalities. J. Funct. Anal., 256(6):1821–1841, 2009.
  • [28] P. Cattiaux, A. Guillin, and L. Wu. A remark on Talagrand’s transportation inequality and logarithmic Sobolev inequality. Probab. Theory Relat. Fields, 148:285–304, 2010.
  • [29] P. Cattiaux, A. Guillin, and L. Wu. Some remarks on weighted logarithmic Sobolev inequalities. Indiana University Mathematics Journal, 60(6):1885–1904, 2011.
  • [30] P. Cattiaux, A. Guillin, and P. A. Zitt. Poincaré inequalities and hitting times. Ann. Inst. Henri Poincaré. Prob. Stat., 49(1):95–118, 2013.
  • [31] P. Cattiaux and S. Méléard. Competitive or weak cooperative stochastic Lotka Volterra systems conditioned to non extinction. J. Math. Biol., 60:797–829, 2010.
  • [32] P. Collet, S. Martinez, and J. San Martin. Quasi-Stationary Distributions. Probability and its applications. Springer, 2013.
  • [33] E. B. Davies. Heat kernels and spectral theory. Cambridge University Press, 1989.
  • [34] H. Djellout, A. Guillin, and L. Wu. Transportation cost-information inequalities for random dynamical systems and diffusions. Ann. Prob., 32(3):2702–2732, 2004.
  • [35] R. Douc, G. Fort, and A. Guillin. Subgeometric rates of convergence of ff-ergodic strong Markov processes. Stochastic Process. Appl., 119(3):897–923, 2009.
  • [36] N. Down, S. P. Meyn, and R. L. Tweedie. Exponential and uniform ergodicity of Markov processes. Ann. Prob., 23(4):1671–1691, 1995.
  • [37] P. W. Glynn and S. P. Meyn. A Liapounov bound for solutions of the Poisson equation. Ann. Probab., 24(2):916–931, 1996.
  • [38] Nathael Gozlan. Integral criteria for transportation-cost inequalities. Electron. Comm. Probab., 11:64–77 (electronic), 2006.
  • [39] L. Gross. Logarithmic Sobolev inequalities and contractivity properties of semi-groups. in Dirichlet forms. Dell’Antonio and Mosco eds. Lect. Notes Math., 1563:54–88, 1993.
  • [40] R. Z. Has’minskii. Stochastic stability of differential equations., volume 7 of Monographs and Textbooks on Mechanics of Solids and Fluids: Mechanics and Analysis,. Sijthoff and Noordhoff, 1980.
  • [41] O. Kavian, G. Kerkyacharian, and B. Roynette. Quelques remarques sur l’ultracontractivité. J. Func. Anal., 11:155–196, 1993.
  • [42] A. M. Kulik. Poincaré inequality and exponential integrability of the hitting times of a Markov process. Theory Stoch. Process., 17(2):71–80, 2011.
  • [43] R. Latała and K. Oleszkiewicz. Between Sobolev and Poincaré. in geometric aspects of Functional Analysis. Lect. Notes Math., 1745:147–168, 2000.
  • [44] M. Ledoux. Concentration of measure and logarithmic Sobolev inequalities. In Séminaire de Probabilités XXXIII, volume 1709 of Lecture Notes in Math., pages 120–216. Springer, Berlin, 1999.
  • [45] M. Ledoux. The concentration of measure phenomenon, volume 89 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2001.
  • [46] Y. Liu. A link between the log-Sobolev inequality and Lyapunov condition. ArXiv Math.PR 1410-6080., 2014.
  • [47] Y. Liu. Gaussian integrability of distance function under the Lyapunov condition. Electron. Commun. Probab., 20(9), 2015.
  • [48] E. Löcherbach, O. Loukianov, and D. Loukianova. Polynomial bounds in the ergodic theorem for one-dimensional diffusions and integrability of hitting times. Ann. Inst. Henri Poincar Probab. Stat., 47(2):425–449, 2011.
  • [49] E. Löcherbach, O. Loukianov, and D. Loukianova. Spectral condition, hitting times and Nash inequality. Ann. Inst. Henri Poincar Probab. Stat., 50(4):1213–1230, 2014.
  • [50] E. Löcherbach and D. Loukianova. Polynomial deviation bounds for recurrent Harris processes having general state space. ESAIM Probab. Stat., 17:195–218, 2013.
  • [51] K. Marton. Bounding d¯\overline{d}-distance by informational divergence: a method to prove measure concentration. Ann. Probab., 24(2):857–866, 1996.
  • [52] S. P. Meyn and R. L. Tweedie. Markov chains and stochastic stability. Communications and Control Engineering Series. Springer-Verlag London Ltd., London, 1993.
  • [53] S. P. Meyn and R. L. Tweedie. Stability of markovian processes II: continuous-time processes and sampled chains. Adv. Appl. Proba., 25:487–517, 1993.
  • [54] S. P. Meyn and R. L. Tweedie. Stability of markovian processes III: Foster-Lyapunov criteria for continuous-time processes. Adv. Appl. Proba., 25:518–548, 1993.
  • [55] F. Otto and C. Villani. Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality. J. Funct. Anal., 173(2):361–400, 2000.
  • [56] C. Roberto and B. Zegarlinski. Orlicz-Sobolev inequalities for sub-Gaussian measures and ergodicity of Markov semi-groups. J. Func. Anal., 243(1):28–66, 2007.
  • [57] A. Yu. Veretennikov. On polynomial mixing bounds for stochastic differential equations. Stochastic Process. Appl., 70(1):115–127, 1997.
  • [58] F. Y. Wang. Functional inequalities, Markov processes and Spectral theory. Science Press, Beijing, 2005.
  • [59] L. Wu. Large and moderate deviations and exponential convergence for stochastic damping Hamiltonian systems. Stochastic Process. Appl., 91(2):205–238, 2001.