We explain briefly how, as in the classical logarithmic Sobolev inequality of Gross, The last application proposed is the concentration property for probability measure satisfying inequality  11 . We obtain Hoeffding's type inequality, assume that a measure $\mu$  on $\mathbb{R}$  satisfies inequality  11 and let $f$  be a Lipschitz function on $\mathbb{R}$  with $\parallel f\parallel Lip⩽1$  . Then we get, for some constants $A,B,D⩾0$  independent of the dimension $n$  ,

$$\begin{array}{c}\mathbb{P}\left(\frac{1}{n}\left|{\sum }_{k=1}^{n}f\left(X_k\right)-\mu \left(f\right)\right|>\lambda \right)⩽\{\begin{array}{cc}2exp\left(-nA\Phi \left(B\lambda \right)\right)& \text{if}\lambda ⩾D,\\ 2exp\left(-nA{\lambda }^2\right)& \text{if}0⩽\lambda ⩽D,\end{array}\end{array}$$

(14)

or equivalently,

$$\begin{array}{c}\mathbb{P}\left(\frac{1}{\sqrt{n}}\left|{\sum }_{k=1}^{n}f\left(X_k\right)-\mu \left(f\right)\right|>\lambda \right)⩽\{\begin{array}{cc}2exp\left(-nA\Phi \left(B\frac{\lambda }{\sqrt{n}}\right)\right)& \text{if}\lambda ⩾D\sqrt{n},\\ 2exp\left(-A{\lambda }^2\right)& \text{if}0⩽\lambda ⩽D\sqrt{n}.\end{array}\end{array}$$

(15)

Inequality  15 is interesting because for large enough $n$  we find the Gaussian concentration, this is natural due to the convergence of $\frac{1}{\sqrt{n}}\left({\sum }_{k=1}^{n}f\left(X_k\right)-\mu \left(f\right)\right)$  to the Gaussian. This result is not a new one, Talagrand explains it in [Tal95] , see also [Led01] for a large review on this topic.

Note to finish the introduction that Barthe, Cattiaux and Roberto [BCR05] are studing the same sort of log-concave measure. They prove also functional inequalities with an other point of view, namely Beckner type inequalities or $\Phi$  -Sobolev inequalities, in particular one of their results is concentration inequalities for the same measure ${\mu }_{\Phi }$  . Let us also mention that the first author in [Gen05] , via Prekopa-Leindler inequality, recovers partly our large entropy result.

2 Proof of logarithmic Sobolev inequality (Theorem  1.2 )

Before explaining the proof of Theorem  1.2 we give a lemma for classical properties satisfied by the function $\Phi$  .

Lemma 2.1Assume that $\Phi$  satisfies assumption $(\mathbf{H})$  then there exists $C⩾0$  such that for large enough $x⩾0$  ,

$$\begin{array}{c}{x}^2⩽C{\Phi }^{*}\left(x\right),\end{array}$$

(16)

$$\begin{array}{c}\varepsilon \Phi \left({{\Phi }^{\prime }}^{-1}\left(x\right)\right)⩽{\Phi }^{*}\left(x\right)⩽\left(1-\varepsilon \right)\Phi \left({{\Phi }^{\prime }}^{-1}\left(x\right)\right),\end{array}$$

(17)

$$\begin{array}{c}\frac{1}{C}{{\Phi }^{\prime }}^{-1}\left(x\right)⩽\frac{{\Phi }^{*}\left(x\right)}{x}⩽C{{\Phi }^{\prime }}^{-1}\left(x\right).\end{array}$$

(18)

The proof of Lemma  2.1 is an easy consequence of the property $(\mathbf{H})$  .

For this we will note by smooth function a locally absolutely continuous function on $\mathbb{R}$  . This is the regularity needed for the use of Hardy inequality in our case.

2.1 Large entropy

The proof of $LSI$  for large entropy is based on the next lemma, we give a $LSI$  saturate on the left.

Lemma 2.2Let $h$  defined as follows

$$\begin{array}{c}h\left(x\right)=\{\begin{array}{cc}1& \text{if}\left|x\right|<M\\ \frac{x^2}{\Phi \left(x\right)}& \text{if}\left|x\right|⩾M.\end{array}\end{array}$$

(19)

Then there exists $C_h⩾0$  such that for every smooth function $g$  we have

$$\begin{array}{c}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(g^2\right)⩽C_h\int {g^{\prime }}^{2}hd{\mu }_{\Phi }.\end{array}$$

(20)

Proof $⊳$  We use Theorem 3 of [BR03] which is a refinement of the criterion of a Bobkov-Götze theorem (see Theorem 5.3 of [BG99] ).

The constant $C_h$  satisfies $max(b_{-},b_{+})⩽C_h⩽max(B_{-},B_{+})$  where $$b_{+}={sup}_{x⩾0}{\mu }_{\Phi }\left(\right[x,+\infty \left[\right)log\left(1+\frac{1}{2{\mu }_{\Phi }\left(\right[x,+\infty \left[\right)}\right){\int }_0^x{Z}_{\Phi }\frac{e^{\Phi \left(t\right)}}{h\left(t\right)}dt,$$  $$b_{-}={sup}_{x⩽0}{\mu }_{\Phi }\left(\right]-\infty ,x\left]\right)log\left(1+\frac{1}{2{\mu }_{\Phi }\left(\right]-\infty ,x\left]\right)}\right){\int }_x^0{Z}_{\Phi }\frac{e^{\Phi \left(t\right)}}{h\left(t\right)}dt,$$  $$B_{+}={sup}_{x⩾0}{\mu }_{\Phi }\left(\right[x,+\infty \left[\right)log\left(1+\frac{e^2}{{\mu }_{\Phi }\left(\right[x,+\infty \left[\right)}\right){\int }_0^x{Z}_{\Phi }\frac{e^{\Phi \left(t\right)}}{h\left(t\right)}dt,$$  $$B_{-}={sup}_{x⩽0}{\mu }_{\Phi }\left(\right]-\infty ,x\left]\right)log\left(1+\frac{e^2}{{\mu }_{\Phi }\left(\right[-\infty ,x\left[\right)}\right){\int }_x^0{Z}_{\Phi }\frac{e^{\Phi \left(t\right)}}{h\left(t\right)}dt.$$  An easy approximation proves that for large positive $x$ 

$$\begin{array}{c}{\mu }_{\Phi }\left(\right[x,\infty \left[\right)={\int }_x^{\infty }\frac{1}{Z_{\Phi }}{e}^{-\Phi \left(t\right)}dt{\sim }_{\infty }\frac{1}{Z_{\Phi }{\Phi }^{\prime }\left(x\right)}{e}^{-\Phi \left(x\right)},\end{array}$$

(21)

and $${\int }_0^x{Z}_{\Phi }\frac{e^{\Phi \left(t\right)}}{h\left(t\right)}dt{\sim }_{\infty }\frac{Z_{\Phi }}{h\left(x\right){\Phi }^{\prime }\left(x\right)}{e}^{\Phi \left(x\right)},$$  and one may prove similar behaviors for negative $x$  .

Then, there is $K$  such that for $x⩾M$  , $${\mu }_{\Phi }\left(\right[x,+\infty \left[\right)log\left(1+\frac{1}{2{\mu }_{\Phi }\left(\right[x,+\infty \left[\right)}\right){\int }_0^x{Z}_{\Phi }\frac{e^{\Phi \left(t\right)}}{h\left(t\right)}dt⩽K\frac{\Phi \left(x\right)}{{\Phi }^{\prime }\left(x{)}^{2}h\right(x)}=K{\left(\frac{\Phi \left(x\right)}{x{\Phi }^{\prime }\left(x\right)}\right)}^2.$$  The right hand term is bounded by the assumption $(\mathbf{H})$  .

A simple calculation then yields that constants $b_{+}$  , $b_{-}$  , $B_{+}$  and $B_{-}$  are finite and the lemma is proved. $⊲$ 

Remark 2.3Note that this lemma can be proved in a more general case, when $\Phi$  does not satisfy hypothesis $(\mathbf{H})$  . In [BL00] the authors prove this result for the symmetric exponential measure.

Proposition 2.4There exists $A,B,D,A^{\prime }⩾0$  such that for any functions $f⩾0$  satisfying $$\int f^{2}d{\mu }_{\Phi }=1\text{and}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)⩾1$$  we have

$$\begin{array}{c}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)⩽A^{\prime }\mathbf{V}\mathbf{a}\mathbf{r}{\mu }_{\Phi }\left(f\right)+A{\int }_{f⩾2}{H}_{\Phi }\left(\frac{f^{\prime }}{f}\right)d{\mu }_{\Phi },\end{array}$$

(22)

where $$H_{\Phi }\left(x\right)=\{\begin{array}{cc}{\Phi }^{*}\left(Bx\right)& \text{if}\left|x\right|⩾D,\\ x^2& \text{if}\left|x\right|⩽D.\end{array}$$ 

As we will see in the proof, $A^{\prime }$  does not depend on the function $\Phi$  .

Proof of Proposition  2.4  $⊳$  Let $f⩾0$  satisfying $\int f^{2}d{\mu }_{\Phi }=1$  .

A careful study of the function $$x\to -x^{2}log{x}^2+5(x-1{)}^2+x^2-1+(x-2{)}_{+}^{2}log(x-2{)}_{+}^2$$  proves that for every $x⩾0$  $$x^{2}log{x}^2⩽5(x-1{)}^2+x^2-1+(x-2{)}_{+}^{2}log(x-2{)}_{+}^2.$$  We know that $\int (f-1{)}^{2}d{\mu }_{\Phi }⩽2\mathbf{V}\mathbf{a}\mathbf{r}{\mu }_{\Phi }\left(f\right)$  , recalling that $\int f^{2}d{\mu }_{\Phi }=1$  and $f⩾0$  ,

$$\begin{array}{ccc}\int f^{2}log{f}^{2}d{\mu }_{\Phi }& ⩽& 5\int (f-1{)}^{2}d{\mu }_{\Phi }+\int (f^2-1)d{\mu }_{\Phi }+\int (f-2{)}_{+}^{2}log(f-2{)}_{+}^{2}d{\mu }_{\Phi }\end{array}$$

$$\begin{array}{ccc}& ⩽& 10\mathbf{V}\mathbf{a}\mathbf{r}{\mu }_{\Phi }\left(f\right)+\int (f-2{)}_{+}^{2}log(f-2{)}_{+}^{2}d{\mu }_{\Phi }.\end{array}$$

Since $\int f^{2}d{\mu }_{\Phi }=1$  , one can easily prove that $$\int (f-2{)}_{+}^{2}d{\mu }_{\Phi }⩽1,$$  then $\int (f-2{)}_{+}^{2}log(f-2{)}_{+}^{2}d{\mu }_{\Phi }⩽\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left((f-2{)}_{+}^2\right),$  and

$$\begin{array}{ccc}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)⩽10\mathbf{V}\mathbf{a}\mathbf{r}{\mu }_{\Phi }\left(f\right)+\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left((f-2{)}_{+}^2\right).& & \end{array}$$

Hardy's inequality of Lemma  2.2 with $g=(f-2{)}_{+}$  gives

$$\begin{array}{c}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left((f-2{)}_{+}^2\right)⩽C_h\int (f-2{)}_{{+}^{\prime }}^{2}hd{\mu }_{\Phi }=C_h{\int }_{f⩾2}{f^{\prime }}^{2}hd{\mu }_{\Phi }.\end{array}$$

(23)

Due to the assumption $(\mathbf{H})$  , the function $h\left(x\right)=x^2/\Phi \left(x\right),$  is increasing on $[M,\infty [$  and $${lim}_{x\to \infty }h\left(x\right)=\infty .$$  We can assume that $\Phi \left(M\right)>0$  . We note $m=h\left(M\right)>0$  Let us define the function $\tau$  as follow

$$\begin{array}{c}\tau \left(x\right)=\{\begin{array}{cc}x\Phi \left(h^{-1}\left(m\right)\right)/\left(8C_{h}m\right)& \text{if}0⩽x⩽m\\ \Phi \left(h^{-1}\left(x\right)\right)/\left(8C_h\right)& \text{if}x⩾m\end{array}\end{array}$$

(24)

For all $x⩾M$  , we have $\tau \left(h\right(x\left)\right)=\Phi \left(x\right)/\left(8C_h\right)$  and then, an easy calculus gives that $\tau$  is increasing on $[0,\infty [$  .

Let $u>0$  ,

$$C_h{\int }_{f⩾2}{f^{\prime }}^{2}hd{\mu }_{\Phi }=C_h{\int }_{f⩾2}u{\left(\frac{f^{\prime }}{f}\right)}^2\frac{h}{u}{f}^{2}d{\mu }_{\Phi }⩽C_h{\int }_{f⩾2}{\tau }^{*}\left\{u{\left(\frac{f^{\prime }}{f}\right)}^2\right\}{f}^{2}d{\mu }_{\Phi }+{\int }_{f⩾2}{C}_h\tau \left(\frac{h}{u}\right)f^{2}d{\mu }_{\Phi }$$ 

For every function $f$  such that $\int f^{2}d{\mu }_{\Phi }=1$  and for every measurable function $g$  such that $\int f^{2}gd{\mu }_{\Phi }$  exists we get $$\int f^{2}gd\mu ⩽\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)+log\int e^{g}d{\mu }_{\Phi }.$$  Indeed, this inequality is also true for all function $g⩾0$  even if the above integrals are infinite.

This inequality is also true for all function $g⩾0$  even integrals are infinite.

We apply the previous inequality with $g=4C_h\tau \left(h/u\right)$  and we obtain $${\int }_{f⩾2}{C}_h\tau \left(\frac{h}{u}\right)f^{2}d{\mu }_{\Phi }⩽\frac{1}{4}\int 4C_h\tau \left(\frac{h}{u}\right)f^{2}d{\mu }_{\Phi }⩽\frac{1}{4}\left(\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)+log\int e^{4C_h\tau \left(\frac{h}{u}\right)}d{\mu }_{\Phi }\right).$$  If $u=1$  we have, by construction, $\int e^{4C_h\tau \left(\frac{h}{u}\right)}d{\mu }_{\Phi }<\infty$  , then we get $${lim}_{u\to \infty }\int e^{4C_h\tau \left(\frac{h}{u}\right)}d{\mu }_{\Phi }=1.$$  Then, by the bounded convergence theorem, there exists $u_0$  such that $\int e^{4C_h\tau \left(\frac{h}{u_0}\right)}d{\mu }_{\Phi }⩽e$  .

Thus we have $$\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left((f-2{{)}_{+}}^2\right)⩽C_h{\int }_{f⩾2}{\tau }^{*}\left\{u_0{\left(\frac{f^{\prime }}{f}\right)}^2\right\}{f}^{2}d{\mu }_{\Phi }+\frac{1}{4}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)+\frac{1}{4}$$  $\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)⩾1$  , implies $$\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)⩽20\mathbf{V}\mathbf{a}\mathbf{r}{\mu }_{\Phi }\left(f\right)+2C_h{\int }_{f⩾2}{\tau }^{*}\left\{u_0{\left(\frac{f^{\prime }}{f}\right)}^2\right\}{f}^{2}d{\mu }_{\Phi }.$$  Then Lemma  2.5 gives the proof of inequality  22 . $⊲$ 

Lemma 2.5There exist constants $A,B,C,D⩾0$  such that $$\forall x⩾0,{\tau }^{*}\left(x^2\right)⩽\{\begin{array}{cc}A{\Phi }^{*}\left(Cx\right)& \text{if}x⩾D,\\ B{x}^2& \text{if}x⩽D.\end{array}$$ 

Proof $⊳$  Let $x>0$  , $${\tau }^{*}\left(x\right)={sup}_{y⩾0}\left\{xy-\tau \left(y\right)\right\}.$$  Let $m=h\left(M\right)>0$  , then

$$\begin{array}{cc}{\tau }^{*}\left(x\right)& =max\left\{{sup}_{y\in [0,m[}\left\{xy-\tau \left(y\right)\right\},{sup}_{y⩾m}\left\{xy-\tau \left(y\right)\right\}\right\},\end{array}$$

$$\begin{array}{cc}& ⩽{sup}_{y\in [0,m[}\left\{xy-\tau \left(y\right)\right\}+{sup}_{y⩾m}\left\{xy-\tau \left(y\right)\right\}.\end{array}$$

We have ${sup}_{y\in [0,m[}\left\{xy-\tau \left(y\right)\right\}⩽xm$  , because $\tau$  is positive. Then the definition of $\tau$  implies that $${sup}_{y⩾m}\left\{xy-\tau \left(y\right)\right\}={sup}_{y⩾M}\left\{x\frac{y^2}{\Phi \left(y\right)}-\frac{\Phi \left(y\right)}{8C_h}\right\}.$$  Let define ${\psi }_x\left(y\right)=x{y}^2/\Phi \left(y\right)-\Phi \left(y\right)/\left(8C_h\right)$  for $y⩾M$  . We have $${\psi }_x^{\prime }\left(y\right)=xy\frac{2\Phi \left(y\right)-y{\Phi }^{\prime }\left(y\right)}{{\Phi }^2\left(y\right)}-\frac{{\Phi }^{\prime }\left(y\right)}{8C_h}.$$  Due to the property $(\mathbf{H})$  , there is $D>0$  such that $$\forall x⩾D,{sup}_{y⩾m}\left\{xy-\tau \left(y\right)\right\}=x\frac{y_x^2}{\Phi \left(y_x\right)}-\frac{\Phi \left(y_x\right)}{8C_h},$$  where $y_x⩾M$  satisfies $$x=\frac{1}{8C_h}\frac{{\Phi }^{\prime }\left(y_x\right){\Phi }^2\left(y_x\right)}{y_x\left(2\Phi \left(y_x\right)-y_x{\Phi }^{\prime }\left(y_x\right)\right)}.$$  The assumption $(\mathbf{H})$  implies that $$\varepsilon y_x{\Phi }^{\prime }\left(y_x\right)⩽2\Phi \left(y_x\right)-y_x{\Phi }^{\prime }\left(y_x\right)⩽\frac{1-\varepsilon }{1+\varepsilon }{y}_x{\Phi }^{\prime }\left(y_x\right),$$  then

$$\begin{array}{c}\frac{1}{8C_h(1-\varepsilon )(2-\varepsilon )}{{\Phi }^{\prime }}^2\left(y_x\right)⩽x⩽\frac{1}{8C_h\varepsilon (1+\varepsilon )}{{\Phi }^{\prime }}^2\left(y_x\right).\end{array}$$

(25)

We get with the assumption $(\mathbf{H})$  ,

$$\begin{array}{cc}\forall x⩾D,{sup}_{y⩾m}\left\{xy-\tau \left(y\right)\right\}& ⩽\frac{1}{8C_h\varepsilon (1+\varepsilon )}{{\Phi }^{\prime }}^2\left(y_x\right)\frac{y_x^2}{\Phi \left(y_x\right)}-\frac{1}{8C_y}\Phi \left(y_x\right)\end{array}$$

$$\begin{array}{cc}& ⩽\frac{(2-\varepsilon {)}^2}{8C_h\varepsilon (1+\varepsilon )}\Phi \left(y_x\right).\end{array}$$

Equation  25 gives, $$y_x⩽{{\Phi }^{\prime }}^{-1}\left(\sqrt{Cx}\right)$$  where $C>0$  . Then we get

$$\begin{array}{cc}\forall x⩾D,{sup}_{y⩾m}\left\{xy-\tau \left(y\right)\right\}& ⩽\frac{(2-\varepsilon {)}^2}{8C_h\varepsilon (1+\varepsilon )}\Phi \left({{\Phi }^{\prime }}^{-1}\left(\sqrt{Cx}\right)\right).\end{array}$$

We obtain, using inequality  17 of Lemma  2.1 , $$\forall x⩾D,{sup}_{y⩾m}\left\{xy-\tau \left(y\right)\right\}⩽\frac{1}{8C_h{\varepsilon }^2(1+\varepsilon )}{\Phi }^{*}\left(\sqrt{Cx}\right).$$  then, $$\forall x⩾D,{\tau }^{*}\left(x\right)⩽xm+K{\Phi }^{*}\left(\sqrt{Cx}\right).$$  Using inequality  16 of Lemma  2.1 we get $$\forall x⩾D,{\tau }^{*}\left(x\right)⩽K^{\prime }{\Phi }^{*}\left(\sqrt{Cx}\right),$$  for some $K^{\prime }⩾0$  .

On the other hand, the function $\tau$  is non-negative and satisfy $\tau \left(0\right)=0$  then ${\tau }^{*}\left(0\right)=0$  . ${\tau }^{*}$  is also a convex function, then there exists $m^{\prime }$  such that $$\forall x\in [0,D],{\tau }^{*}\left(x\right)⩽x{m}^{\prime },$$  which proves the lemma. $⊲$ 

Corollary 2.6For any smooth function $f>0$  on $\mathbb{R}$  satisfying $$\int f^{2}d{\mu }_{\Phi }=1,\text{and}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)⩾1,$$  we have $$\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)⩽C\int H_{\Phi }\left(\frac{f^{\prime }}{f}\right)f^{2}d{\mu }_{\Phi },$$  where $$H_{\Phi }\left(x\right)=\{\begin{array}{cc}{\Phi }^{*}\left(Bx\right)& \text{if}\left|x\right|⩾D\\ x^2& \text{if}\left|x\right|⩽D,\end{array}$$  and $B,D⩾0$  .

Proof $⊳$  Due to the property $(\mathbf{H})$  the measure ${\mu }_{\Phi }$  satisfies a Spectral Gap inequality, $$\mathbf{V}\mathbf{a}\mathbf{r}{\mu }_{\Phi }\left(f\right)⩽C_{SG}\int {f^{\prime }}^{2}d{\mu }_{\Phi },$$  with $C_{SG}⩾0$  . We apply inequality  22 to get the result. $⊲$ 

2.2 Small entropy

Lemma 2.7Let $\lambda >0$  and define the function $\psi$  by $$\psi \left(x\right)={\left\{({\Phi }^{*}{)}^{-1}\left(\lambda logx\right)\right\}}^2.$$  Then for all $\lambda >0$  there exists $A_{\lambda }>0$  such that the function $\psi$  is well defined, positive, increasing, concave on $[A_{\lambda },\infty [$  and satisfies $\psi \left(A_{\lambda }\right)⩾1$  .

Proof $⊳$  Let $\lambda >0$  be fixed. Classical property of the Legendre-Frenchel transform implies that ${\Phi }^{*}$  is convex. Due to the property ( $\mathbf{H}$  ), ${\left({\Phi }^{*}\right)}^{-1}(\lambda logx)$  is well defined for $x⩾M_1$  with $M_1>0$  . Then we get on $[M_1,\infty [$  , $${\psi }^{\prime }\left(x\right)=2g^{\prime }(\lambda logx)g(\lambda logx)\frac{\lambda }{x},$$  and $${\psi }^{\prime \prime }\left(x\right)=2g(\lambda logx)\frac{{\lambda }^2}{x^2}\left(g^{\prime \prime }(\lambda logx)-\frac{g^{\prime }(\lambda logx)}{\lambda }+\frac{{g^{\prime }}^2(\lambda logx)}{g(\lambda logx)}\right),$$  where, for simplicity, we have noted $g={\left({\Phi }^{*}\right)}^{-1}$  .

For $x$  large enough $g$  is non-negative and increasing and then $\psi$  is increasing on $[M_2,\infty [$  , with $M_2⩾0$  .

An easy estimation gives that as $x$  goes to infinity,

$$\begin{array}{c}\frac{g^{\prime }\left(x\right)}{g\left(x\right)}=o_{\infty }\left(1\right),\end{array}$$

(26)

then since ${\left({\Phi }^{*}\right)}^{-1}$  is concave, for all large enough $x$  , ${\psi }^{\prime \prime }\left(x\right)⩽0$  . Then one can find $A_{\lambda }>0$  such that properties on the Lemma  2.7 are true. $⊲$ 

The proof of $LSI$  for small entropy is based on the next lemma, we give a $LSI$  saturate on the right.

Lemma 2.8There exists $\lambda >0$  which depends on the function $\Phi$  such that if we note by $A_{\lambda }$  the constant of Lemma  2.7 we get for all $g$  defined on $[T,\infty [$  with $T\in [T_1,T_2]$  for some fixed $T_1,T_2$  , and verifying that $$g\left(T\right)=\sqrt{A_{\lambda }},g⩾\sqrt{A_{\lambda }}\text{and}{\int }_T^{\infty }{g}^{2}d{\mu }_{\alpha }⩽2A_{\lambda }+2.$$  Then we get

$$\begin{array}{c}{\int }_T^{\infty }(g-\sqrt{A_{\lambda }}{)}_{+}^2\psi (g^2){\mu }_{\Phi }⩽C_1{\int }_{[T,\infty [}{g^{\prime }}^{2}d{\mu }_{\Phi },\end{array}$$

(27)

where $\psi$  is defined on Lemma  2.7 .

The constant $C_1$  depend on $\Phi$  and $\lambda$  but does not depend on the value of $T\in [T_1,T_2]$  .

Proof $⊳$  Let use Hardy's inequality as explained in the introduction. We have $g\left(T\right)=A_{\lambda }$  . We apply inequality  12 on $[T,\infty [$  with the function $(g-\sqrt{A_{\lambda }}{)}_{+}$  and the following measures $$d\mu =\psi \left(g^2\right)d{\mu }_{\Phi }\text{and}\nu ={\mu }_{\Phi }.$$  Then the constant $C$  in inequality  27 is finite if and only if $$B={sup}_{x⩾T}{\int }_T^x{e}^{\Phi \left(t\right)}dt{\int }_x^{\infty }\psi \left(g^2\right)d{\mu }_{\Phi },$$  is finite.

By Lemma  2.7 , $\psi$  is concave on $[A_{\lambda },\infty [$  then by Jensen inequality, for all $x⩾T$  we get $${\int }_x^{\infty }\psi \left(g^2\right)d{\mu }_{\Phi }⩽{\mu }_{\Phi }\left([x,\infty [\right)\psi \left(\frac{{\int }_x^{\infty }{g}^{2}d{\mu }_{\Phi }}{{\mu }_{\Phi }\left([x,\infty [\right)}\right).$$  Then we have

$$\begin{array}{ccc}B⩽{sup}_{x>T_1}\left\{{\int }_{T_1}^x{e}^{\Phi \left(t\right)}dt{\mu }_{\Phi }\left([x,\infty [\right)\psi \left(\frac{{\int }_x^{\infty }{g}^{2}d{\mu }_{\Phi }}{{\mu }_{\Phi }\left([x,\infty [\right)}\right)\right\}& & \end{array}$$

(28)

Due to the property $(\mathbf{H})$  there exists $K>1$  such that

$$\begin{array}{c}{\Phi }^{\prime }\left(x\right)e^{\Phi \left(x\right)}⩽e^{K\Phi \left(x\right)},\end{array}$$

(29)

and $${\int }_{T_{{}_1}}^x{e}^{\Phi \left(t\right)}dt⩽\frac{K{e}^{\Phi \left(x\right)}}{{\Phi }^{\prime }\left(x\right)},{\int }_x^{\infty }{e}^{-\Phi \left(t\right)}dt⩽\frac{e^{-\Phi \left(x\right)}}{{\Phi }^{\prime }\left(x\right)},$$  for large enough $x$  . By  29 we get also for large enough $x$  that $$e^{-K\Phi \left(x\right)}⩽{\int }_x^{\infty }{e}^{-\Phi \left(t\right)}dt.$$  Then for large enough $x$  , uniformly in the previous $g$  , one have $${\int }_{T_{{}_1}}^x{e}^{\Phi \left(t\right)}dt{\mu }_{\Phi }\left([x,\infty [\right)\psi \left(\frac{{\int }_x^{\infty }{g}^{2}d{\mu }_{\Phi }}{{\mu }_{\Phi }\left([x,\infty [\right)}\right)⩽\frac{K}{{\left({\Phi }^{\prime }\left(x\right)\right)}^2}\psi \left(\frac{{\int }_x^{\infty }{g}^{2}d{\mu }_{\Phi }}{K}{e}^{K\Phi \left(x\right)}\right).$$  For $x$  large enough, $$\frac{{\int }_x^{\infty }{g}^{2}d{\mu }_{\Phi }}{K}⩽1.$$  Then, by definition of $\psi$  , for large enough $x$  , $${\int }_{T_{{}_1}}^x{e}^{\Phi \left(t\right)}dt{\mu }_{\Phi }\left([x,\infty [\right)\psi \left(\frac{{\int }_x^{\infty }{g}^{2}d{\mu }_{\Phi }}{{\mu }_{\Phi }\left([x,\infty [\right)}\right)⩽K{\left(\frac{{{\Phi }^{*}}^{-1}\left(\lambda K\Phi \left(x\right)\right)}{{\Phi }^{\prime }\left(x\right)}\right)}^2.$$  There is also $C_{\epsilon }$  such that, for $x$  large enough $${{\Phi }^{*}}^{-1}\left(x\right)⩽{\Phi }^{\prime }\left({\Phi }^{-1}\left(C_{\epsilon }x\right)\right),$$  as one can see from equation  17 .Then one can choose $\lambda =1/\left(K{C}_{\epsilon }\right)$  and the lemma is proved. Note that $\lambda$  depends only on the function $\Phi$  .

The constant $B$  on  28 is bounded by $K$  which does'nt depend on $T$  on $[T_1,T_2]$  . $⊲$ 

Proposition 2.9There exists $A,A^{\prime },B,D>0$  such that for any functions $f⩾0$  satisfying $$\int f^{2}d{\mu }_{\alpha }=1\text{and}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\alpha }\left(f^2\right)⩽1$$  we have $$\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)⩽A\mathbf{V}\mathbf{a}\mathbf{r}{\mu }_{\Phi }\left(f\right)+A^{\prime }{\int }_{f^2⩾A_{\lambda }}H\left(\frac{f^{\prime }}{f}\right)d{\mu }_{\Phi },$$  where $$H\left(x\right)=\{\begin{array}{cc}{\Phi }^{*}\left(Bx\right)& \text{if}\left|x\right|⩾D,\\ x^2& \text{if}\left|x\right|⩽D.\end{array}$$ 

Proof $⊳$  Let $f⩾0$  satisfying $\int f^{2}d{\mu }_{\alpha }=1$  .

We can assume that $A_{\lambda }⩾2$  . A careful study of the function $$x\to -x^{2}log{x}^2+A(x-1{)}^2+x^2-1+(x-\sqrt{A_{\lambda }}{)}_{+}^{2}log(x-\sqrt{A_{\lambda }}{)}_{+}^2$$  proves that there exists $A$  such that for every $x\in {\mathbb{R}}^{+}$  $$x^{2}log{x}^2⩽A(x-1{)}^2+x^2-1+(x-\sqrt{A_{\lambda }}{)}_{+}^{2}log(x-\sqrt{A_{\lambda }}{)}_{+}^2.$$  Then we get

$$\begin{array}{ccc}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\alpha }\left(f^2\right)=\int f^{2}log{f}^{2}d{\mu }_{\alpha }⩽A\mathbf{V}\mathbf{a}\mathbf{r}{\mu }_{\alpha }\left(f\right)+\int {\left(f-\sqrt{A_{\lambda }}\right)}_{+}^{2}log{f}^{2}d{\mu }_{\alpha },& & \end{array}$$

(30)

where $\sqrt{A_{\lambda }}$  is defined as in Lemmas  2.7 and  2.8 .

Fix $\lambda$  as in Lemma  2.8 . We define the function $K$  on $[A_{\lambda },\infty [$  by $$K\left(x\right)=\sqrt{\frac{log{x}^2}{\psi \left(x^2\right)}},$$  where $A_{\lambda }$  is defined on Lemma  2.8 .

Let now define $T_1<T_2$  such that $${\mu }_{\Phi }\left(]\infty ,T_1]\right)=\frac{3}{8},{\mu }_{\Phi }\left([T_1,T_2]\right)=\frac{1}{4}\text{and}{\mu }_{\Phi }\left([T_2,+\infty [\right)=\frac{3}{8}.$$  Since $\int f^{2}d{\mu }_{\Phi }=1$  there exists $T\in [T_1,T_2]$  such that $f\left(T\right)⩽A_{\lambda }$  .

Let us define $g$  on $[T_1,\infty ]$  as follow $$g=\sqrt{A_{\lambda }}+{\left(f-\sqrt{A_{\lambda }}\right)}_{+}K\left(f\right)\text{on}[T,\infty [.$$  Function $g$  satisfies $g\left(T\right)=\sqrt{A_{\lambda }}$  and $g\left(x\right)⩾\sqrt{A_{\lambda }}$  for all $x⩾T$  .

Then we have

$$\begin{array}{ccc}\begin{array}{cc}{\int }_T^{\infty }{g}^{2}d{\mu }_{\Phi }& ⩽{\int }_{T_1}^{\infty }{g}^{2}d{\mu }_{\Phi }\\ & ⩽2A_{\lambda }+2{\int }_{[T_1,\infty [\cap \left\{f^2⩾A_{\lambda }\right\}}{f}^2{K}^2\left(f\right)d{\mu }_{\Phi }\\ & ⩽2A_{\lambda }+2{\int }_{[T_1,\infty [}{f}^{2}log\left(f^2\right)d{\mu }_{\Phi }\\ & ⩽2A_{\lambda }+2,\end{array}& & \end{array}$$

(31)

where we are using the growth of $\psi$  on $[A_{\lambda },\infty [$  and $\psi \left(A_{\lambda }\right)⩾1$  .

Assumptions on Lemma  2.8 are satisfied, we obtain by inequality  27  $${\int }_T^{\infty }(g-\sqrt{A_{\lambda }}{)}_{+}^2\psi (g^2)d{\mu }_{\Phi }⩽C_1{\int }_{[T,\infty [\cap \left\{f^2⩾A_{\lambda }\right\}}{g^{\prime }}^{2}d{\mu }_{\Phi }.$$  Let us compare the various terms now.

Due to the property ( $\mathbf{H}$  ), $K$  is lower bounded on $[\sqrt{A_{\lambda }},\infty [$  by $\alpha ⩾1$  (maybe for $A_{\lambda }$  larger), then we get firstly $$\sqrt{A_{\lambda }}+{\left(f-\sqrt{A_{\lambda }}\right)}_{+}K\left(f\right)⩾\sqrt{A_{\lambda }}+{\left(f-\sqrt{A_{\lambda }}\right)}_{+}\alpha ⩾f\text{on}\left\{f^2⩾A_{\lambda }\right\}.$$  Then

$${\left(g-\sqrt{A_{\lambda }}\right)}_{+}^2\psi \left(g^2\right)={\left(f-\sqrt{A_{\lambda }}\right)}_{+}^{2}K(f{)}^2\psi {\left(\sqrt{A_{\lambda }}+{\left(f-\sqrt{A_{\lambda }}\right)}_{+}K\left(f\right)\right)}^2⩾{\left(f-\sqrt{A_{\lambda }}\right)}_{+}^{2}K(f{)}^2\psi \left(f^2\right)={\left(f-\sqrt{A_{\lambda }}\right)}_{+}^{2}log{f}^2,$$ 

by the definition of $K$  , then we obtain

$$\begin{array}{c}{\int }_T^{\infty }(f-\sqrt{A_{\lambda }}{)}_{+}^{2}log{f}^{2}d{\mu }_{\Phi }⩽{\int }_T^{\infty }(g-\sqrt{A_{\lambda }}{)}_{+}^2\psi \left(g^2\right)d{\mu }_{\Phi }.\end{array}$$

(32)

Secondly we have on $\left\{f⩾\sqrt{A_{\lambda }}\right\}$ 

$$\begin{array}{ccc}\begin{array}{cc}{g}^{\prime }& =f^{\prime }K\left(f\right)+{\left(f-\sqrt{A_{\lambda }}\right)}_{+}{f}^{\prime }{K}^{\prime }\left(f\right)\\ & =f^{\prime }K\left(f\right)\left(1+{\left(f-\sqrt{A_{\lambda }}\right)}_{+}\frac{K^{\prime }\left(f\right)}{K\left(f\right)}\right)\end{array}& & \end{array}$$

But we have for $x⩾\sqrt{A_{\lambda }}$ 

$$\begin{array}{ccc}\begin{array}{cc}\left|1+(x-\sqrt{A_{\lambda }})\frac{K^{\prime }\left(x\right)}{K\left(x\right)}\right|& ⩽1+x\left|\frac{K^{\prime }\left(x\right)}{K\left(x\right)}\right|\\ & ⩽1+\frac{1}{2logx}+\left|\frac{\lambda }{x}\frac{g^{\prime }(\lambda 2logx)}{g(\lambda 2logx)}\right|,\end{array}& & \end{array}$$

where $g\left(x\right)={{\Phi }^{*}}^{-1}\left(x\right)$  . Using Lemma  2.7 and the estimation  26 we obtain that there exists $C>0$  such that for all $x⩾\sqrt{A_{\lambda }}$  , $$\left|1+(x-\sqrt{A_{\lambda }})\frac{K^{\prime }\left(x\right)}{K\left(x\right)}\right|⩽C$$  We get then $${g^{\prime }}^2⩽C{f^{\prime }}^2{K}^2\left(f\right)\text{on}\left\{f^2⩾A_{\lambda }\right\},$$  for some $C<\infty$  and then

$$\begin{array}{c}{\int }_{[T,\infty [\cap \left\{f^2⩾A_{\lambda }\right\}}{g^{\prime }}^{2}d{\mu }_{\Phi }⩽C{\int }_{[T,\infty [\cap \left\{f^2⩾A_{\lambda }\right\}}{f^{\prime }}^2{K}^2\left(f\right)d{\mu }_{\Phi }.\end{array}$$

(33)

By equation  32 and  33 we obtain $${\int }_T^{\infty }(f-\sqrt{A_{\lambda }}{)}_{+}^{2}log{f}^{2}d{\mu }_{\Phi }⩽C{\int }_{[T,\infty [\cap \left\{f^2⩾A_{\lambda }\right\}}{f^{\prime }}^2{K}^2(f)d{\mu }_{\Phi }.$$  Let $u_0>0$  ,

$${\int }_T^{\infty }(f-\sqrt{A_{\lambda }}{)}_{+}^{2}log{f}^{2}d{\mu }_{\Phi }⩽C{\int }_{[T,\infty [\cap \left\{f^2⩾A_{\lambda }\right\}}{\tau }_2^{*}\left(u_0{\left(\frac{f^{\prime }}{f}\right)}^2\right)f^{2}d{\mu }_{\Phi }+{\int }_{[T,\infty [\cap \left\{f^2⩾A_{\lambda }\right\}}{\tau }_2\left(\frac{K^2\left(f\right)}{u_0}\right)f^{2}d{\mu }_{\Phi },$$ 

where the function ${\tau }_2$  is defined as in equation  24 by

$$\begin{array}{c}{\tau }_2\left(x\right)=\{\begin{array}{cc}x\Phi \left(h^{-1}\left(m\right)\right)\frac{1-\epsilon }{2\lambda m}& \text{if}0⩽x<m\\ \Phi \left(h^{-1}\left(x\right)\right)\frac{1-\epsilon }{2\lambda }& \text{if}x⩾m,\end{array}\end{array}$$

(34)

where $h$  is defined on equation  19 and $m$  on equation  24 . The function ${\tau }_2$  is equal to $\tau$  up to a constant factor.

Using Lemma  2.10 we get

$${\int }_T^{\infty }(f-\sqrt{A_{\lambda }}{)}_{+}^{2}log{f}^{2}d{\mu }_{\Phi }⩽C{\int }_{[T,\infty [\cap \left\{f^2⩾A_{\lambda }\right\}}{\tau }_2^{*}\left(u_0{\left(\frac{f^{\prime }}{f}\right)}^2\right)f^{2}d{\mu }_{\Phi }+\frac{1}{2}{\int }_{[T,\infty [\cap \left\{f^2⩾A_{\lambda }\right\}}{f}^{2}log{f}^{2}d{\mu }_{\Phi }.$$ 

The same method can be used on $]-\infty ,T]$  and then there is $C^{\prime }<\infty$  such that

$${\int }_{-\infty }^T(f-\sqrt{A_{\lambda }}{)}_{+}^{2}log{f}^{2}d{\mu }_{\Phi }⩽C^{\prime }{\int }_{]-\infty ,T]\cap \left\{f^2⩾A_{\lambda }\right\}}{\tau }_2^{*}\left(u_0{\left(\frac{f^{\prime }}{f}\right)}^2\right)f^{2}d{\mu }_{\Phi }+\frac{1}{2}{\int }_{]-\infty ,T]\cap \left\{f^2⩾A_{\lambda }\right\}}{f}^{2}log{f}^{2}d{\mu }_{\Phi }.$$ 

And then we get

$$\int (f-\sqrt{A_{\lambda }}{)}_{+}^{2}log{f}^{2}d{\mu }_{\Phi }⩽(C+C^{\prime }){\int }_{\left\{f^2⩾A_{\lambda }\right\}}{\tau }_2^{*}\left(u_0{\left(\frac{f^{\prime }}{f}\right)}^2\right)f^{2}d{\mu }_{\Phi }+\frac{1}{2}{\int }_{\left\{f^2⩾A_{\lambda }\right\}}{f}^{2}log{f}^{2}d{\mu }_{\Phi }.$$ 

Note that constants $C$  and $C^{\prime }$  don't depend on $T\in [T_1,T_2]$  .

Then by inequality  30 and Lemma  2.5 , Proposition  2.9 is proved. $⊲$ 

Lemma 2.10There exists $u_0>0$  such that, for all $x⩾A_{\lambda }$  we have $${\tau }_2\left(\frac{K^2\left(x\right)}{u_0}\right)⩽\frac{1}{2}log{x}^2.$$ 

Proof $⊳$  Let $\kappa =2\lambda /(1-\epsilon )$  .

For all $x⩾M$  , where $M$  is defined on equation  24 , we have $${\tau }_2\left(h\left(x\right)\right)=\frac{\Phi \left(x\right)}{\kappa },$$  $${\tau }_2\left(\frac{x^2}{\Phi \left(x\right)}\right)=\frac{\Phi \left(x\right)}{\kappa }.$$  ${\tau }_2$  is increasing, then due to the property ( $\mathbf{H}$  ) we have for $x⩾M$  $${\tau }_2\left((1+\epsilon {)}^2\frac{\Phi \left(x\right)}{{\Phi }^{\prime }(x{)}^2}\right)⩽\frac{\Phi \left(x\right)}{\kappa }.$$  Using now inequality  17 one has $$\frac{1}{{\Phi }^{\prime }\left(x\right)}⩾\frac{1}{{\Phi }^{*}\left((1-\epsilon )\Phi \left(x\right)\right)},$$  then for all $x⩾M$  , $${\tau }_2\left((1+\epsilon {)}^2\frac{\Phi \left(x\right)}{{\Phi }^{*}{\left((1-\epsilon )\Phi \left(x\right)\right)}^2}\right)⩽\frac{\Phi \left(x\right)}{\kappa }.$$  Take now $z=(1-\epsilon )\Phi \left(x\right)$  , $${\tau }_2\left(\frac{(1+\epsilon {)}^2}{1-\epsilon }\frac{z}{{\Phi }^{*}{\left(z\right)}^2}\right)⩽\frac{z}{(1-\epsilon )\kappa },$$  to finish take $x=exp\left(\frac{4z}{(1-\epsilon )\kappa }\right)$  to obtain $${\tau }_2\left((1+\epsilon {)}^2\kappa \frac{log{x}^2}{{\Phi }^{*}{\left(\frac{(1-\epsilon )\kappa }{2}log{x}^2\right)}^2}\right)⩽\frac{1}{2}log{x}^2.$$  Recall that $\lambda =(1-\epsilon )\kappa /2$  and let take $u_0=1/\left((1+\epsilon {)}^2\kappa \right),$  to obtain the result for $x⩾C$  , where $C$  is a constant depending on $\Phi$  .

If we have $A_{\lambda }<C$  , one can change the value of $u_0$  to obtain also the results on $[A_{\lambda },C]$  . $⊲$ 

Proof of Theorem  1.2  $⊳$  To give the proof of the theorem we need to give an other result like Proposition  2.4 . By the same argument as in Proposition  2.4 one can also prove that there exists $A,A^{\prime },B,D>0$  such that for any functions $f⩾0$  satisfying $$\int f^{2}d{\mu }_{\alpha }=1\text{and}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\alpha }\left(f^2\right)⩾1$$  we have for some $C^{\prime }\left(A_{\lambda }\right)$  , $C\left(A_{\lambda }\right)$ 

$$\begin{array}{c}\mathbf{E}\mathbf{n}\mathbf{t}{\mu }_{\Phi }\left(f^2\right)⩽C^{\prime }\left(A_{\lambda }\right)\mathbf{V}\mathbf{a}\mathbf{r}{\mu }_{\Phi }\left(f\right)+C\left(A_{\lambda }\right){\int }_{f^2⩾A_{\lambda }}H\left(\frac{f^{\prime }}{f}\right)d{\mu }_{\Phi },\end{array}$$

(35)

where $H_{\Phi }$  is defined on  10 and $A_{\lambda }$  on the Proposition  2.9 . To introduce $A_{\lambda }$  , we just have to change constants in the inequality.

Then the proof of the theorem is a simple consequence of  35 and Proposition  2.9 . $⊲$ 

3 Classical properties and applications

Let us give here properties inherited directly from the methodology known for classical logarithmic Sobolev inequalities.

Proposition 3.1

• 1 This property is known under the name of tensorisation.
  Let ${\mu }_1$  and ${\mu }_2$  two probability measures on ${\mathbb{R}}^{n_1}$  and ${\mathbb{R}}^{n_2}$  . Suppose that ${\mu }_1$  (resp.

  ${\mu }_2$  ) satisfies the a $LSI$  with function $H_{\Phi }$  and constant $A_1$  (resp. with constant $A_2$  ) then the probability ${\mu }_1\otimes {\mu }_2$  on ${\mathbb{R}}^{n_1+n_2}$  , satisfies a $LSI$  with function $H_{\Phi }$  and constant $max\left\{A_1,A_2\right\}$  .
• 2 This property is known under the name of perturbation.
  Let $\mu$  a measure on ${\mathbb{R}}^n$  a $LSI$  with function $H_{\Phi }$  and constant $A$  . Let $h$  a bounded function on ${\mathbb{R}}^n$  and defined $\tilde{\mu }$  as $$d\tilde{\mu }=\frac{e^h}{Z}d\mu ,$$  where $Z=\int e^{h}d\mu$  .

  Then the measure $\tilde{\mu }$  satisfies a $LSI$  with function $H_{\Phi }$  and the constant $D=A{e}^{2\text{osc}\left(h\right)}$  , where $\text{osc}\left(h\right)=sup\left(h\right)-inf\left(h\right)$  .
• 3 Link between $LSI$  of function $H_{\Phi }$  with Poincaré inequality.
  Let $\mu$  a measure on ${\mathbb{R}}^n$  . If $\mu$  satisfies a $LSI$  with function $H_{\Phi }$  and constant $A$  , then $\mu$  satisfies a Poincaré inequality with the constant $A$  . Let us recall that $\mu$  satisfies a Poincaré inequality with constant $A$  if $$\mathbf{V}\mathbf{a}\mathbf{r}\mu \left(f\right)⩽A\int {\left|\nabla f\right|}^{2}d\mu ,$$  for all smooth function $f$  .

Proof $⊳$  One can find the details of the proof of the properties of tensorisation and perturbation and the implication of the Poincaré inequality in chapters 1 and 3 of [ABC + 00] (Section 1.2.6., Theorem 3.2.1 and Theorem 3.4.3). $⊲$ 

Proposition 3.2Assume that the probability measure $\mu$  on $\mathbb{R}$  satisfies a $LSI$  with function $H_{\Phi }$  and constant $A$  . Then there exists constants $B,C,D⩾0$  , independent of $n$  such that: if $F$  is a function on ${\mathbb{R}}^n$  such that $\forall i$  , ${\parallel {\partial }_{i}F\parallel }_{\infty }⩽\zeta$  , then we get for $\lambda ⩾0$  ,

$$\begin{array}{c}{\mu }^{\otimes n}(\left|F-{\mu }^{\otimes n}\left(F\right)\right|⩾\lambda )⩽\{\begin{array}{cc}2exp\left(-nB\Phi \left(C\frac{\lambda }{n\zeta }\right)\right)& \text{if}\lambda >nD\zeta ,\\ 2exp\left(-B\frac{{\lambda }^2}{n{\zeta }^2}\right)& \text{if}0⩽\lambda ⩽nD\zeta .\end{array}\end{array}$$

(36)

Proof $⊳$  Let us first present the proof when $n=1$  . Assume, without loss of generality, that $\int Fd\mu =0$  .

Due to the homogeneous property of  36 on can suppose that $\zeta =1$  .Let us recall briefly Herbst's argument (see Chapter 7 [ABC + 00] for more details). Denote $\psi \left(t\right)=\int e^{tF}d\mu$  , and remark that $LSI$  of function $H_{\Phi }$  applied to $f^2=e^{tF}$  , using basic properties of $H_{\Phi }$  , yields to

$$\begin{array}{c}t{\psi }^{\prime }\left(t\right)-\psi \left(t\right)log\psi \left(t\right)\le A{H}_{\Phi }\left(\frac{t}{2}\right)\psi \left(t\right)\end{array}$$

(37)

which, denoting $K\left(t\right)=(1/t)log\psi \left(t\right)$  , entails $$K^{\prime }\left(t\right)⩽\frac{A}{t^2}{H}_{\Phi }\left(\frac{t}{2}\right).$$  Then, integrating, and using $K\left(0\right)=\int Fd\mu =0$  , we obtain

$$\begin{array}{c}\psi \left(t\right)\le exp\left(At{\int }_0^t\frac{1}{s^2}{H}_{\Phi }\left(\frac{s}{2}\right)ds\right).\end{array}$$

(38)

Then we get using Markov inequality $$\mu (\left|F-\mu \left(F\right)\right|⩾\lambda )⩽2exp\left({min}_{t⩾0}\left\{At{\int }_0^t\frac{1}{s^2}{H}_{\Phi }\left(\frac{s}{2}\right)ds-\lambda t\right\}\right).$$  Let note, for $t⩾0$  , $$G\left(t\right)=At{\int }_0^t\frac{1}{s^2}{H}_{\Phi }\left(\frac{s}{2}\right)ds-\lambda t.$$  An easy study proves that $G$  admits a single minimum on ${\mathbb{R}}^{+}$  (except maybe if $\lambda =0$  ). Then due to the definition of $H_{\Phi }$  we get that $${min}_{t⩾0}\left\{G\left(t\right)\right\}=-\frac{{\lambda }^2}{A},\text{if}\lambda ⩽AD.$$  Assume now that $\lambda ⩾AD$  then we obtain after derivation

$$\begin{array}{c}{min}_{t⩾0}\left\{G\left(t\right)\right\}=-A{\Phi }^{*}\left(t_0\frac{B}{2}\right),\text{with}\lambda t_0=A{t}_0{\int }_0^{t_0}\frac{1}{s^2}{H}_{\Phi }\left(\frac{s}{2}\right)ds+A{H}_{\Phi }\left(\frac{t_0}{2}\right).\end{array}$$

(39)

We first prove that there exists $C⩾0$  such that for all $t_0$  large enough

$$\begin{array}{c}{t}_0{\int }_0^{t_0}\frac{1}{s^2}{H}_{\Phi }\left(\frac{s}{2}\right)ds⩽C{H}_{\Phi }\left(\frac{t_0}{2}\right).\end{array}$$

(40)

For $\kappa ⩾0$  large enough and $t_0⩾\kappa$  we get using then inequality  17 we get $$t_0{\int }_{\kappa }^{t_0}\frac{1}{s^2}{H}_{\Phi }\left(\frac{s}{2}\right)ds⩽C{t}_0{\int }_{\kappa }^{t_0}\frac{1}{s^2}\Phi \left({{\Phi }^{\prime }}^{-1}\left(\frac{s}{2}\right)\right)ds,$$  with $C⩾0$  . Then by a change of variables and integration by parts, for large enough $t_0$  ,

$$\begin{array}{cc}{t}_0{\int }_{\kappa }^{t_0}\frac{1}{s^2}\Phi \left({{\Phi }^{\prime }}^{-1}\left(\frac{s}{2}\right)\right)ds& =\frac{t_0}{2}{\int }_{{{\Phi }^{\prime }}^{-1}\left(\frac{\kappa }{2}\right)}^{{{\Phi }^{\prime }}^{-1}\left(\frac{t_0}{2}\right)}\frac{\Phi \left(u\right)}{{\Phi }^{\prime }(u{)}^2}{\Phi }^{\prime \prime }\left(u\right)du\end{array}$$

$$\begin{array}{cc}& ⩽\frac{t_0}{2}\frac{\Phi \left({\Phi }^{-1}\right(\kappa /2\left)\right)}{{\Phi }^{\prime }\left({\Phi }^{-1}\right(\kappa /2\left)\right)}+\frac{t_0}{2}{{\Phi }^{\prime }}^{-1}(t_0/2)\end{array}$$

$$\begin{array}{cc}& ⩽C{t}_0{{\Phi }^{\prime }}^{-1}(t_0/2),\end{array}$$

for some other $C⩾0$  . Then we get, using inequality  18 , for $t_0$  large enough, $$t_0{\int }_0^{t_0}\frac{1}{s^2}{H}_{\Phi }\left(\frac{s}{2}\right)ds⩽C{t}_0{{\Phi }^{\prime }}^{-1}(t_0/2)⩽C^{\prime }{\Phi }^{*}(t_0/2).$$  for some constant $C^{\prime }⩾0$  and for $t_0$  large enough and inequality  40 is proved. By  40 and  39 one get for $t_0$  large enough, $$\lambda t_0⩽A^{\prime }{\Phi }^{*}\left(\frac{t_0}{2}\right),$$  for some constant $A^{\prime }⩾0$  . But, using inequality  18 we get then $${\Phi }^{\prime }\left(A\lambda \right)⩽C{t}_0,$$  $${min}_{t⩾0}\left\{G\left(t\right)\right\}⩽-A{\Phi }^{*}\left(B{\Phi }^{\prime }\left(C\lambda \right)\right)⩽-A{\Phi }^{*}\left({\Phi }^{\prime }\left(C^{\prime }\lambda \right)\right),$$  if $\lambda$  is large enough and for some other constants $A,B,C,C^{\prime }⩾0$  . Using inequality  17 , we obtain the result in dimension 1.

For the $n$  -dimensional extension, use the tensorisation property of $LSI$  of function $H_{\Phi }$  and $${\sum }_{i=1}^n{H}_{\Phi }\left(\frac{t}{2}{\partial }_{i}F\right)⩽n{H}_{\Phi }\left(\frac{t}{2}\right).$$  Then we can use the case of dimension 1 with the constant $A$  replaced by $An$  . $⊲$ 

Remark 3.3Let us present a simple application of the preceding proposition to deviation inequality of the empirical mean of a function. Consider the real valued function $f$  , with $\left|f^{\prime }\right|\le 1$  . Let apply Proposition  3.2 with the two functions $$F(x_1,...,x_n)=\frac{1}{n}{\sum }_{k=1}^{n}f\left(x_k\right)\text{and}F(x_1,...,x_n)=\frac{1}{\sqrt{n}}{\sum }_{k=1}^{n}f\left(x_i\right).$$  We obtain then $$\mathbb{P}\left(\frac{1}{n}\left|{\sum }_{k=1}^{n}f\left(X_k\right)-\mu \left(f\right)\right|>\lambda \right)⩽\{\begin{array}{cc}2exp\left(-nA\Phi \left(B\lambda \right)\right)& \text{if}\lambda ⩾D,\\ 2exp\left(-nA{\lambda }^2\right)& \text{if}0⩽\lambda ⩽D,\end{array}$$  $$\mathbb{P}\left(\frac{1}{\sqrt{n}}\left|{\sum }_{k=1}^{n}f\left(X_k\right)-\mu \left(f\right)\right|>\lambda \right)⩽\{\begin{array}{cc}2exp\left(-nA\Phi \left(B\frac{\lambda }{\sqrt{n}}\right)\right)& \text{if}\lambda ⩾D\sqrt{n},\\ 2exp\left(-A{\lambda }^2\right)& \text{if}0⩽\lambda ⩽D\sqrt{n}.\end{array}$$ 

References

1. 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.
2. D. Bakry. L'hypercontractivité et son utilisation en théorie des semigroupes. In Lectures on probability theory. École d'été de probabilités de St-Flour 1992, volume 1581 of Lecture Notes in Math., pages 1–114. Springer, Berlin, 1994.
3. F. Barthe, P. Cattiaux, and C. Roberto. Interpolated inequalities between exponential and Gaussian, Orlicz hypercontractivity and isoperimetry. To appear in Rev. Math. Iber., 2005.
4. 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.
5. S. G. Bobkov and F. Götze. Exponential integrability and transportation cost related to logarithmic Sobolev inequalities. J. Funct. Anal., 163(1):1–28, 1999.
6. S. G. Bobkov and M. Ledoux. Poincaré's inequalities and Talagrand's concentration phenomenon for the exponential distribution. Probab. Theory Related Fields, 107(3):383–400, 1997.
7. S. G. Bobkov and M. Ledoux. From Brunn-Minkowski to Brascamp-Lieb and to logarithmic Sobolev inequalities. Geom. Funct. Anal., 10(5):1028–1052, 2000.
8. F. Barthe and C. Roberto. Sobolev inequalities for probability measures on the real line. Studia Math., 159:481–497, 2003.
9. I. Gentil. Logarithmic sobolev inequality for log-concave measure from prekopa-leindler inequality. Preprint, 2005.
10. I. Gentil, A. Guillin, and L. Miclo. Modified logarithmic sobolev inequalities and transportation inequalities. To appear in Probab. Theory Related Fields, 2005.
11. L. Gross. Logarithmic Sobolev inequalities. Amer. J. Math., 97(4):1061–1083, 1975.
12. 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.
13. M. Ledoux. The concentration of measure phenomenon, volume 89 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2001.
14. M. Talagrand. Concentration of measure and isoperimetric inequalities in product spaces. Inst. Hautes Études Sci. Publ. Math., (81):73–205, 1995.

Ivan Gentil, Arnaud Guillin CEREMADE (UMR 7534, Université Paris-Dauphine et CNRS) Place du Maréchal De Lattre de Tassigny, 75775 Paris Cédex 16, France {gentil,guillin}@ceremade.dauphine.fr http://www.ceremade.dauphine.fr/ $\stackrel{~}{}$  {gentil,guillin} Laurent Miclo Laboratoire d'Analyse, Topologie, et Probabilités (UMR 6632, Université de Provence et CNRS) 39, rue F. Joliot Curie, 13453 Marseille Cédex 13, France miclo@cmi.univ-mrs.fr