HAL Id: hal-00609026
https://hal.archives-ouvertes.fr/hal-00609026
Submitted on 17 Jul 2011
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
Dominique Bakry, François Bolley, Ivan Gentil
To cite this version:
Dominique Bakry
∗ †
, François Bolley
‡
and Ivan Gentil
§
September 10, 2010
Introdu tion
In the Eu lidean spa e
R
n
, the lassi al Nashinequality may be statedas
(0.1)
kfk
1+n/2
2
≤ C
n
kfk
1
k∇fk
n/2
2
for all smooth fun tions
f
(with ompa t support for instan e) where the norms areomputed with respe t to the Lebesgue measure. This inequality has been introdu ed
by J. Nash in 1958 (see [9℄) toobtain regularity properties on the solutionsto paraboli
partial dierential equations. The omputation of the optimal onstant
C
n
has beenperformed morere ently in[6℄.
This inequality may be stated in the general framework of symmetri Markov
semi-groups,where itis asimple andpowerful tooltoobtain estimates onthe asso iated heat
kernel. In this ontext, one repla es
k∇fk
2
2
by the Diri hletformE(f, f)
asso iated with thesemigroup, andthe Lebesguemeasurebyitsreversible measure. Moreover, the powerfun tion
x
n
in the inequality is repla ed by a more general onvex fun tion
Φ
, and un-der this form it an be valid(and useful) even in innite dimensional situations su h asthose whi h appear in statisti al me hani s. One an also give weighted forms of these
inequalities : they also lead to pre ise estimates on the semigroup, or on the spe tral
de ompositionof the generator.
TheaimofthisshortnoteistoexplainhowNashinequalitiesleadtosu hestimatesin
ageneral settingand alsoto show simple te hniques used to establishthe required Nash
inequalities. There isno laim for originality,most of the materialin luded here may be
found invariouspapers su h as [1, 2, 5,7, 13℄.
∗
Institut deMathématiquesdeToulouse,UMRCNRS5219,UniversitéPaul-Sabatier
†
Institut UniversitairedeFran e.
‡
Ceremade,UMRCNRS 7534,UniversitéParis-Dauphine.
§
Nashinequalitiesbelongtotheverylargefamilyoffun tionalinequalitiesfor
symmet-ri Markov semigroups whi h have led tomany re ent works. Many of these inequalities
ompare
L
p
norms of fun tions to the
L
2
norms of their gradients, whi h in this
on-text is alled the Diri hlet form; this is the ase of the simplest ones, the spe tral gap
(or Poin aré) inequalities. But one may also onsider
L
1
norms of the gradients, in the
area of isoperimetri inequalities, or
L
p
norms, even
L
∞
norms, when one is on erned
withestimates onLipshitzfun tions,for instan e inthearea of on entration ofmeasure
phenomena.
Here, we shall on entrate only on
L
2
norms of gradients. Even in this setting, there
exists awide variety of inequalities,whi hare adapted tothe kind of measure one wants
to study on one side, and to the properties they des ribe on the other. For example,
measures with polynomial de ay are not overed by the same inequalities as measures
with exponential, orsquare exponentialde ay.
The family of Nash type inequalities we present here belongs to the wide family of
the Sobolev type inequalities. Their main interest is that they easilyprovide good (and
sometimesalmostoptimal) ontrolonheat kernels. Startingfromthe lassi alinequality,
we shall show how to extend them rst by the introdu tion of a rate fun tion
Φ
, and thenbythe extraintrodu tionofaweightfun tionV
(aLyapunovfun tion). As weshall see, the link between Nashinequalities and estimates onthe semigroup spe trumis verysimple and, as usual in the eld, roughly relies on derivation along time and integration
by parts. This is why it is tempting touse itin awide range of situations.
Then, we shall show how to obtain these inequalities in the simplest models on the
realline. Restri tingourselvestothereallinemaybethought aslookingonlyattheeasy
ase. In fa t, by hoosing various measures, one may produ e a lot of dierent model
ases whi h really illustrate what may or may not be expe ted from these inequalities.
Then the extension to higher dimensional situations(like
R
n
ormanifolds) is very often
apure matter of te hni alities, extendingin adire t way the one-dimensional methods.
The paper is organized asfollows. In the rst se tion, we briey present the ontext
of symmetri Markov semigroups, and parti ularlydiusion semigroups. Then, we show
dierent variations of Nash inequalities and how to get estimates on heat kernels from
them. Then, inne, we show howtoprodu e su hNash inequalitiesonthe basi models
onthe real line weare interested in.
1 Symmetri Markov semigroups and difusions
To understand the general ontext of Markov semigroups, we rst onsider a measure
spa e
(E, B, µ)
, whereB
isaσ
-eld andµ
is aσ
-nite measure on it. Although we shall alwaysfo us onexampleswhere(E, B)
isR
n
equipped withthe usualBorelsets (orsome
openset init,oranite dimensionalmanifoldwithorwithoutboundaries), itmaybean
aboutthemeasurablestru ture ofthe spa e. Inany ase, oneshouldalwayssuppose that
(E, B, µ)
is a "reasonable" measure spa e : we shall not say in details what we mean by "reasonable",but resultssu hasthede ompositionofmeasure theoremsshouldbevalid,whi h overs all asesone ouldlook atin pra tise.
Given
(E, B, µ)
asymmetri Markov semigroup is afamily(P
t
)
t≥0
of linear operators mappingthesetofboundedmeasurablefun tionsintoitselfwiththefollowingproperties:(i) Preservation of positivity: if
f ≥ 0
, so isP
t
f
. (ii) Preservation of onstant fun tions :P
t
1 = 1
. (iii) Semigroup property :P
t
◦ P
s
= P
t+s
.(iv) Symmetry :
P
t
mapsL
2
(µ)
intoitself and, forany pair
(f, g) ∈ L
2
(µ)
,one hasZ
E
P
t
f g dµ =
Z
E
f P
t
g dµ.
(v) Continuity at
t = 0
:P
0
= Id
andP
t
f → f
whent → 0
inL
2
(µ)
.
Su hsemigroupsnaturallyappearinprobabilitytheoryas
P
t
f (x) = E(f(X
t
)/X
0
= x)
where(X
t
)
t≥0
is a Markov pro ess. The symmetry property does not always hold and it is equivalent to the reversibility of the pro ess. They also appear in many situationswhenone solves a"heat equation"of the form
∂
t
F (x, t) = LF, F (x, 0) = f (x);
here
L
is ase ond order sub-ellipti (orhypo-ellipti )dierential operator,in whi h aseP
t
f
isthe solutionF (x, t)
at timet
when the initialvalue isF (x, 0) = f (x)
. Letus start with some elementary preliminaryremarks.(a) Sin e
P
t
is symmetri ,andP
t
1 = 1
, one getsR
E
P
t
f dµ =
R
E
f dµ
by takingg = 1
in the property (iv):µ
is invariantunder the semigroup.(b) Sin e
P
t
is linear and positivity preserving,|P
t
f | ≤ P
t
|f|
. This implies thatP
t
is a ontra tioninL
1
(µ)
by invarian eof the measure.
( ) Bythe sameargument,
P
t
is alsoa ontra tioninL
∞
(µ)
and therefore,by
interpola-tion,
P
t
is a ontra tion inL
p
(µ)
for any
p ∈ [1, ∞]
.(d) Sin e
(P
t
)
t
is a semigroup of ontra tions inL
2
(µ)
, by the Hille-Yoshida theory, it
Formally, any property of the semigroup may be translated into a property of the
generator
L
, and vi e versa. For instan e, the preservation of onstants property translates intoL1 = 0
. Also, the symmetry translates into the fa t thatL
is self-adjoint,that is,Z
E
f Lg dµ =
Z
E
gLf dµ.
The positivity preserving property is more subtle. In general, it is translated into
a maximum prin iple of the generator. But this requires a bit more than just a
measurablestru ture on the spa e, and weprefer to translatethis intothe positivity
of the arré du hamp operator, see point(k) below.
(e) The measure
µ
being symmetri (or "reversible") is in general unique up to a nor-malizing onstant (it is however a restri tive ondition that su h a measure exists :see formula(1.3) below). When the measure is nite, we may therefore normalizeit
as tobea probability measure, and we shall always do it. In this ase, the onstant
fun tion
1
isalwaysanormalized eigenve tor,asso iated withthe eigenvalue0
whi h isthe smallestvalueof thespe trumof−L
. Inthe innite ase, thereisno anoni al way of hoosing agoodnormalization.(f) Sin ethemeasurespa e
(E, B, µ)
isa"reasonable"spa e,any su hoperatorP
t
whi h preserves the onstants and positivitymay berepresented asP
t
f (x) =
Z
E
f (y)P
t
(x, dy),
where
P
t
(x, dy)
is akernel of probability measures, that is, aprobability measure onE
dependingonthe parameterx ∈ E
inameasurable way. Thisenables forexample to apply toP
t
any generi property of probability measures, su h as the varian e inequalityP
t
(f
2
) ≥ (P
t
f )
2
.(g) Veryoften(andweshallseethatNashinequalitiesprovideauseful riteriumforthis),
this kernel has a density with respe t tothe reversible measure
µ
,that isP
t
(x, dy) = p
t
(x, y)µ(dy);
here
p
t
(x, y)
is a non negative fun tion whi h is dened almost everywhere (with respe t toµ ⊗ µ
) onE × E
. Then the symmetry property (iv) is equivalent to the symmetryofthiskernelp
t
(x, y) = p
t
(y, x)
. Mu hattentionhas beenbroughtoverthe lastde ades tovariousestimates onthis kernel density (inparti ular in Riemanniangeometry, for heat kernels on Riemannian manifolds, using tools from geometry like
urvature, Riemannian distan e, et ). On e again, Nash inequalities may provide
good su h estimates, as we shall see later on.
(h) When we have su h densities, the semigroup property translates intothe
Chapman-Kolmogorov equation
p
t+s
(x, y) =
Z
E
Hen e, by the Cau hy-S hwarz inequality,
p
2t
(x, y)
2
≤ p
2t
(x, x) p
2t
(y, y).
As the onsequen e, the maximum of
p
t
(x, y)
isalways obtained onthe diagonal. (i) The generatorL
being self-adjoint has a spe tral de omposition with a spe trum in(−∞, 0]
a ording to(1.2).(j) It may be the ase that the spe trum is dis rete, and that we have a omplete
se-quen eoforthonormaleigenve tors
(f
n
)
inL
2
(µ)
,witheigenvalues
−λ
n
forL
. Inthis situation,the kernel densityp
t
(x, y)
is given byp
t
(x, y) =
X
n
e
−λ
n
t
f
n
(x)f
n
(y).
Then wehave the tra e formula
Z
E
p
t
(x, x)dµ(x) =
X
n
e
−λ
n
t
.
On e again, Nash inequalities will provide uniform (or non uniform) bounds on the
densities, and therefore bounds onthe ounting fun tion of the sequen e
(λ
n
)
. (k) Byderivation att = 0
the varian e inequalityP
t
(f
2
) ≥ (P
t
f )
2
givesthe inequalityL(f
2
) ≥ 2fLf.
In parti ular
(1.2)
Z
E
f Lf dµ ≤ 0
by invarian e of
µ
. Of ourse, one has to take are about whi h fun tions these do apply. In general, we assume that there is an algebra of fun tionsA
dense in the domain ofL
, for whi h this is valid. In this ase, one denes the arré du hamp as the bilinearformΓ(f, g) =
1
2
L(f g) − fLg − gLf
.
It satises
Γ(f, f ) ≥ 0
, and in somesense this hara terizes the positivity preserving property ofP
t
.The Diri hlet form asso iated to
P
t
isnally dened byThe last identity is based on the identity
R
E
L(f g)dµ = 0
and is alled the integration by parts formula. The Diri hlet form is in general dened on a larger domain than theoperator
L
itself (formally, it requires only one derivative of the fun tion to be inL
2
(µ)
insteadof
2
for the generator).The knowledge of the measure and of the arré du hamp (or of the Diri hlet form)
entirelydes ribes the operator
L
(and therefore the semigroup), sin eL
may be dened fromΓ
andµ
through the integration by parts formula(see (1.3)).The basi example of su h semigroups is of ourse the standard heat kernel in the
Eu lidean spa e
R
n
; for
t > 0
, its densityp
t
(x, y)
with respe t to the Lebesgue measuredy
isp
t
(x, y) =
1
(4πt)
n/2
exp(−
|x − y|
2
4t
).
Here,
µ(dy) = dy
,L = ∆
andΓ(f, f ) = |∇f|
2
.
This orresponds to the ase studied by Nash in [9℄. It is one of the very few examples
whereone expli itlyknows
P
t
,sin eingeneralweonlyknowL
,andtheissue istodedu e asmu h informationas possible onP
t
fromthe knowledge ofL
.Another model ase is the Ornstein-Uhlenbe k semigroupon
R
n
, for whi hLf (x) = ∆f (x) − x · ∇f(x), Γ(f, f) = |∇f|
2
, µ(dx) =
1
(2π)
n/2
exp(−|x|
2
/2)dx.
Its density with respe t to the Gauss measure
µ(dy)
isp
t
(x, y) = (1 − e
−2t
)
−n/2
exp
−
1
2(1 − e
−2t
)
(|y|
2
e
−2t
− 2 x · ye
−t
+ |x|
2
e
−2t
)
and it behaves in a very dierent way from the previous example as long as fun tional
inequalitiesare on erned.
In the two previous ases, the arré du hamp is the same (and the semigroups only
dierbythemeasure
µ(dx)
). ObservethatΓ(f, g)
isinboth asesarstorderdierential operator in its two arguments. They both belong to the large lass of diusion Markovsemigroups,whi h are semigroups su h that for all smooth fun tions
φ
Γ(φ(f ), g) = φ
′
(f )Γ(f, g)
orequivalently
Lφ(f ) = φ
′
(f )Lf + φ
′′
(f )Γ(f, f ).
This property is alled the hange of variable formula for
L
and is an intrinsi way of saying thatL
is a se ond order dierential operator. The fa t thatL(1) = 0
says that thereis no0
-orderterm inL
. Onemay easily he k that amongall dierentialoperatorson
R
n
Γ(f, f ) ≥ 0
,provided the matrix ofthe se ondorder termsispositive-semideniteatany point.Non diusion ases are of onsiderable interest sin e they are related to Markov
pro- esses with jumps and also naturally appear when one looks at subordinators. However
weshall on entrate here onthe diusion ase, even thoughthe Nashte hniques may be
used inthe same way inthe general ase.
In general,a se ond orderdierentialoperator without
0
-order termsof the formLf =
X
ij
a
ij
(x)∂
ij
2
f +
X
i
b
i
(x)∂
i
f
has a arré du hamp given by
Γ(f, g) =
X
ij
a
ij
(x)∂
i
f ∂
j
g = ∇f · A(x)∇g.
Therefore the positivity of
Γ
is equivalent to the fa t that at any pointx
the matrixA(x) = (a
ij
(x))
is positive-semidenite. Conversely, when the arré du hamp is given
(on aopen set in
R
n
or onasmooth manifold in lo al oordinates) by
Γ(f, g) =
X
ij
a
ij
(x)∂
i
f ∂
j
g,
withpositive-semidenitematri es
(a
ij
(x))
having smooth oe ients,and whenthe
ref-eren e measure
µ(dx)
has a smooth positive densityρ(x)
with respe t to the Lebesgue measure, it orresponds toa unique symmetri diusion operatorL
whi h is(1.3)
Lf =
1
ρ(x)
X
i
∂
i
ρ(x)
X
j
a
ij
(x)∂
j
f
.
Inotherwords,
Γ
odes forthese ondorderpartof theoperatorwhileµ
odesfor the rstorderterms. Observealsothatea h(Γ, µ)
leadstounique symmetriL
,butpossibly several non symmetriL
.A model ase onwhi h weshall fo us is the ase when
E = R
andΓ(f, f ) = f
′2
. We
shall look atthe measures
(1.4)
µ(dx) = C exp(−|x|
a
)dx,
where
a > 0
andC
isa normalizing onstant. In orderto avoidirrelevant di ultiesdue to the non smoothness of|x|
at0
, we shall repla e|x|
by√
1 + x
2
. Depending on the
value of
a
, the orresponding semigroups present diverse behaviours.For
a = 2
, the elebrated Nelson theorem [10℄ asserts that the Ornstein-Uhlenbe k semigroupis "hyper ontra tive", whi h means thatP
t
isbounded fromL
2
(µ)
to
L
for all
t > 0
and someq(t) > 2
. This is equivalent to the also famous Gross logarithmi Sobolev inequality[8℄ (1.5)Z
f
2
log f
2
dµ ≤
Z
f
2
dµ log
Z
f
2
dµ
+ CE(f, f).
When
a > 2
, the semigroup is "ultra ontra tive", whi h means thatP
t
mapsL
1
(µ)
into
L
∞
(µ)
for any
t > 0
, while it is not even hyper ontra tive fora < 2
. Nevertheless, for1 < a < 2
,ithas adis retespe trumanditisHilbert-S hmidt,andweshallsee below howto get estimates onthe spe trumthrough weighted Nashinequalities.For
a = 1
,the spe trumisnolongerdis reteandtheonlypropertyleftistheexisten e of a spe tral gap : the spe trum of−L
lies in{0} ∪ [A, ∞)
for someA > 0
, and this property is equivalentto aspe tralgap (or Poin aré) inequality(1.6)
Z
f
2
dµ ≤
Z
f dµ
2
+
1
A
E(f, f).
When
a < 1
,even the spe tral gap property is lost. Of ourse, one may look at similar models inR
n
, or on Riemannian manifolds with
density measures (with respe t to the Riemann measure) depending on the distan e to
some point. In this latter ase, one would get more ompli ated results, sin e in general
one has to take intoa ount lower bounds onthe Ri i urvature, and even more if one
works with boundaries (with Neumann boundary onditions). We shall not develop this
here.
2 Nash inequalities
Inthe ontextofDiri hletformsasso iatedtosymmetri Markovsemigroupsasdes ribed
above,a Nashinequality isan inequality of the form
(2.7)
kfk
1+n/2
2
≤ kfk
1
k
C
1
kfk
2
2
+ C
2
E(f, f)
n/4
;
herethe norms
L
p
areof ourse omputedwith respe t tothe reversiblemeasure
µ
andn
isany positive parameter(that we allthedimension in the Nashinequality,sin e in thelassi al ase the unique possible value for
n
is really the dimension of the spa e). This inequality should apply for anyf
in the Diri hlet domain, but it is enough to he k it in a dense subspa e of it whi h, in many examples, will be the set of smooth ompa tlysupported fun tions.
It is worth mentioning that when
µ
isa probability measure, thenC
1
≥ 1
(as an be seen by hoosingf = 1
), while for example inR
n
with the Lebesgue measure, one may
When
µ
is a probability measure andC
1
= 1
, we say that the inequality is tight. It thenimpliesaspe tralgap inequality,asonemaysee by applyingthe inequality to1 + ǫf
and lettingǫ
go to0
.Conversely, if (2.7) is valid with
C
1
> 1
and withµ
being a probability measure, together with a spe tral gap inequality, then a tight Nash inequality holds (see [1℄ forexample). In general, we say that a fun tional inequality is tight when one may dedu e
from the inequality that
{E(f, f) = 0}=⇒{f =
onstant}
. Here, whenC
1
= 1
, this is ensured by the equality ase in the inequalitykfk
1
≤ kfk
2
. As we shall see, tightness may beused to ontrol the onvergen e toequilibrium,that is the asymptoti behaviourwhen
t → ∞
, while the general inequality is useful to ontrol the short time behaviour. Mostof fun tionalinequalitiesmay be tightenedin presen e of aspe tral gap inequality,asit isthe ase here.
In the ase of an innite measure, tightness orresponds tothe ase when
C
1
= 0
, as inthe Eu lidean ase.However, there is a strong dieren e between the forms that the Nash inequalities
may take a ording towhether the measure is nite ornot. We know that atight Nash
inequality holds true in the Eu lidean spa e, but it an be proved that the tight Nash
inequality (2.7) may not be validon a nite measure spa e unless the spa e is ompa t.
More pre isely, whenone has atightNashinequality (2.7)onanite measure spa e, one
maygetaboundontheos illationofLip hitzfun tions,when eaboundonthe diameter
of the spa e (this diameter being measured in terms of an intrinsi distan e asso iated
with the arré du hamp [4℄). This explains why belowwe introdu e the extended Nash
inequalities(2.9)and(2.10),whi hmaybevalidonnitemeasurespa eswithunbounded
support, as we shall see.
When
n > 2
, the Nash inequality (2.7) is one of the many forms of the Sobolev inequalitykfk
2
2n/(n−2)
≤ C
1
kfk
2
+ C
2
E(f, f).
Indeed, observe rst that this and Hölder's inequalitieslead tothe Nashinequality (2.7)
with the same "dimension"
n
and the same onstantsC
1
andC
2
. The way ba k is a little more subtle: the argument in [3℄ is based on applying the Nash inequality to thesequen e of fun tions
f
n
= min{(f − 2
n
)
+
, 2
n
}
, addingthe obtained estimates and using the identityP
n
E(f
n
, f
n
) = E(f, f)
. This enables tokeep the same dimensionn
,but not the onstantsC
1
andC
2
.Inthe ontextofDiri hletforms,su hSobolevinequalitiesmayappearunderdierent
forms su h as Energy-Entropy, Gagliardo-Nirenberg, Faber-Krahn et inequalities. We
referto [3℄ for full details. The onne tion between Sobolev (and Nash) inequalities and
various bounds on heat kernels has been explored by many authors, see [1, 7, 11℄ for
example.
Theorem 2.1. Assume that inequality (2.7) holds. Then, (2.8)
kP
t
f k
2
≤ C(t)kfk
1
,
whereC(t) =
max{2C
1
,
2nC
2
t
}
n/4
.
Conversely, if (2.8) holds with
C(t) ≤ a + bt
−n/4
, then a Nash inequality (2.7) holds
with the same dimension
n
and onstantsC
1
andC
2
depending only onn, a
andb
. Proof Letus rewritethe inequality under the formkfk
2
2
kfk
2
1
1+2/n
≤ C
1
kfk
2
2
kfk
2
1
+ C
2
E(f, f)
kfk
2
1
.
Now, hoose apositive fun tion
f
and apply the pre eding bound toP
t
f
. We knowfrom invarian eofµ
thatR
E
P
t
f dµ =
R
E
f dµ
. Letus setH(t) =
kP
t
f k
2
2
kP
t
f k
2
1
=
kP
t
f k
2
2
kf k
2
1
. Wehave∂
t
kP
t
f k
2
2
= 2
Z
P
t
f LP
t
f dµ = −2 E(P
t
f, P
t
f ).
Therefore,
H
isde reasing andH
′
(t) = −2
E(P
t
f, P
t
f )
kP
t
f k
2
1
,
and the Nashinequality (2.7) be omes
H
1+2/n
≤ C
1
H − 2C
2
H
′
.
Now, as long asH ≥ (2C
1
)
n/2
,one hasH
1+2/n
≥ 2C
1
H
,and we getH
1+2/n
≤ −4C
2
H
′
,
and this dierential inequality (with the fa t that
H
isde reasing) givesthe result. To see the reverse way, we may observe that, for a general symmetri Markovsemi-group, the fun tion
t 7→ K(t) = log kP
t
f k
2
2
is onvex. Indeed, the derivative ofK
1
(t) =
kP
t
f k
2
2
is−2
R P
t
f LP
t
f dµ
, whilethe se ondderivativeis4
R (LP
t
f )
2
dµ
,and thereforeby
Cau hy-S hwarz inequality one has
K
1
′
2
≤ K
1
K
1
′′
,
whi hsaysthat
log K
1
= K
is onvex. Thensoisthefun tionh(t) = log
Now, if we have a bound of the form
H(t) ≤ a + bt
−n/2
, we may plug this upper bound
inthe previous inequality, and then optimisein
t
to get the result. In fa t,havinga bound forP
t
asan operatorfromL
1
into
L
2
,we are very lose from
a uniform bound on the kernel
p
t
. Indeed, ifP
t
is bounded fromL
1
into
L
2
with norm
C(t)
, thenby symmetryand duality, itisalsoboundedfromL
2
into
L
∞
with norm
C(t)
,and therefore by omposition and semigroup property
P
2t
is bounded fromL
1
intoL
∞
with normC(t)
2
.Conversely,bytheRiesz-Thorintheorem,if
P
t
isbounded fromL
1
into
L
∞
withnorm
C
1
(t)
, being bounded fromL
1
intoitselfwith norm
1
,it isalsobounded fromL
1
into
L
2
with norm
C
1
(t)
1/2
. Inthe end, we have obtained the following
Theorem2.2. ANashinequality (2.7)holdswithdimension
n
ifandonlyifP
t
isboundedfrom
L
1
into
L
∞
with norm bounded above by
a + bt
−n/2
.
Of ourse, inthe ase when
C
1
= 0
, whi h orresponds tothe lassi alEu lidean Nash inequality, the equivalen e is valid with a bound of the formC(t) = at
−n/2
.
Moreover, a verygeneral fa t(validon"reasonablemeasure spa es"
(E, B, µ)
)asserts thatanoperatorK
isbounded fromL
1
(µ)
into
L
∞
(µ)
ifandonlyifitmayberepresented
by a bounded kernel density
k
:K(f )(x) =
R
E
k(x, y)f (y)µ(dy)
. Moreover, the norm operatorofK
is exa tlytheL
∞
norm of
k
(onE × E
).Sowehaveseen thataNashinequalityisequivalenttoauniformboundonthe kernel
of
P
t
(and also arries the existen e of su h kernel), with very few assumptions on the spa e.Observe that there is noreason why we shouldrestri t ourselves tothe ase ofpower
fun tionsin Nashinequalities. One may onsider extensions of the form
(2.9)
Φ
kfk
2
2
kfk
2
1
≤
E(f, f)
kfk
2
1
,
validsaywhenever
kfk
2
> Mkfk
1
. HereΦ
isasmooth onvexin reasingfun tiondened on an interval(M, ∞)
. (It does not require formally to be onvex in reasing, but it is reallyuseful only inthis ase).Su hinequalities have been introdu edby F.-Y. Wang in[12℄ under the form
kfk
2
2
≤ aE(f, f) + b(a)kfk
2
1
,
alledsuper Poin aré inequalities. Theseinequalitiesmaybeoptimizedunderthe
param-eter
a
togivekfk
2
2
kfk
2
1
≤ Ψ
E(f, f)
kfk
2
1
with some on ave fun tion
Ψ
, whi h isequivalent toinequality (2.9).Then, one an write the argument of Theorem 2.1 with (2.9) instead of (2.7) and we
see that the key assumption is
R
∞
1
Theorem 2.3 (Wang). Assume that an extended Nash inequality (2.9) is valid with a
rate fun tion
Φ
dened on someinterval(M, ∞)
and su hthatR
∞
1
Φ(s)
ds < ∞
. Thenwe havekP
t
f k
2
≤ K(2t) kfk
1
for all
t > 0
and all fun tionsf ∈ L
2
(µ)
; here the fun tion
K
is dened byK(x) =
pU
√
−1
(x)
if 0 < x < U(M),
M
if x ≥ U(M)
where
U
denotes the (de reasing)fun tion dened on(M, +∞)
byU(x) =
Z
∞
x
1
φ(u)
du.
In parti ular, the density
p
t
(x, y)
is bounded from above byK(t)
2
.
Conversely, if there exists a positive fun tion
K
denedon(0, ∞)
su h thatkP
t
f k
2
≤ K(t)kfk
1
for all
t > 0
, then the Nash inequality (2.9) holds withM = 0
and fun tionΦ(x) = sup
t>0
x
2t
log
x
K(t)
2
,
x ≥ 0.
With the te hniques presented in the next se tion, we may see that su h extended
Nashinequalities with fun tions
Φ
of the formx(log x)
α
are adapted tothe study of the
the measures
µ
a
des ribed in (1.4) fora > 2
: as we already mentioned, be ause of non ompa tness, there is no hope inthis ase to have a lassi alNash inequality (2.7) withapowerfun tion
Φ
.In the ase when the measure is nite(and therefore aprobability measure), then we
know that
kfk
2
/kfk
1
≥ 1
. For su h a general inequality, tightness orresponds to the fa t thatΦ(x) → 0
whenx → 1
. (Of ourse, this supposes thatM = 1
in the previous theorem).In this situation,assume that
Φ(x) ∼ λ(x − 1)
whenx → 1
and1/Φ
is integrableat innity. This is the ase in parti ular for the tight form of the lassi al Nashinequal-ity (2.7). Then
K(t) ∼ 1 + Ce
−λt
when
t → ∞
. This shows that the kernelp
t
(x, y)
is bounded from above by a quantity whi h onverges exponentially fast to1
ast
goes to innity. This is what may be expe ted, sin eP
t
f → µ(f)
whent → ∞
. In the ase of a lassi al tight Nashinequality(whi h an onlyo ur when the measure has aboundedsupport), then one may also dedu e a lower uniform bound on the kernel
p
t
whi h also goesexponentially fastto1
, but this requires someother te hniques (see [1℄).The dierent Nash inequalities introdu ed so far may only arry information on the
the general ase when it is not bounded we may still use this method with the tri k of
introdu ingan auxiliaryLyapunov fun tion
V
and weighted Nashinequalities.Forus,aLyapunovfun tion
V
issimplyapositivefun tionV
onE
su hthatLV ≤ cV
forsome onstantc
. Weshallrequire thosefun tionsV
tobeinL
2
(µ)
and inthedomain
toget interesting results,but it is not formallyne essary.
Being a Lyapunov is not a very restri tive requirement for smooth fun tions in the
examples below, as long aswe do not ask
c < 0
(inwhi h ase it annot be true for any fun tionV
in the domain).The weighted Nash inequality takesthen the form
(2.10)
Φ
kfk
2
2
kfV k
2
1
≤
E(f, f)
kfV k
2
1
forallfun tions
f
inthe domainof theDiri hlet formsu h thatkfk
2
2
> M kfV k
2
1
,where the rate fun tionΦ
is dened on(M, ∞)
and su h thatΦ(x)/x
isin reasing.Theorem 2.4 (Wang). Assume that a weighted Nash inequality (2.10) holds with a rate
fun tion
Φ
dened on some interval(M, ∞)
su hthatR
∞
1
Φ(s)
ds < ∞
. ThenkP
t
f k
2
≤ K(2t)e
ct
kfV k
1
for all
t > 0
and all fun tionsf ∈ L
2
(µ)
, where
K
is dened as in Theorem 2.3. Inparti ular, the kernel density
p
t
(x, y)
satisesp
t
(x, y) ≤ K(t)
2
e
ct
V (x)V (y).
Conversely, if there exists a positive fun tion
K
denedon(0, ∞)
su h thatkP
t
f k
2
≤ K(t)kfV k
1
forall
t > 0
, thentheweightedNash inequality (2.10)holdswithM = 0
andratefun tionΦ(x) = sup
t>0
x
2t
log
x
K(t)
2
,
x ≥ 0.
Proof It is given in detail in [2℄. It follows the proof of Theorem 2.1by repla ing the
fun tion
K(t) =
kP
t
f k
2
2
kfk
2
1
byK(t) =
ˆ
kP
t
f k
2
2
kV fk
2
1
. Now, the quantity
R P
t
f V dµ
is no longer invariantin time. Butby properties of the Lyapunov fun tionwe have(2.11)
∂
t
Z
E
V P
t
f dµ = ∂
t
Z
P
t
V f dµ =
Z
P
t
LV f dµ ≤ c
Z
E
V P
t
f dµ,
fromwhi hwe get
Z
E
P
t
f V dµ ≤ e
ct
Z
Using this, we get again a dierential inequality on
K
ˆ
when we apply the Nash in-equality (2.10)toP
t
f
, and theL
1
→ L
2
boundedness result follows.
Togetthe(nonuniform)boundonthekernel,itremainstoobservethatifasymmetri
operator
K
satiseskKfk
2
≤ kfV k
1
, the norms being onsidered with respe t to ameasure
µ
, then the operatorK
1
dened byK
1
(f ) =
1
V
K(f V )
isa ontra tionfromL
1
(ν)
intoL
2
(ν)
, wheredν = V
2
dµ
. Moreover,K
1
is symmetri inL
2
(ν)
, and therefore
K
1
◦ K
1
is a ontra tion fromL
1
(ν)
into
L
∞
(ν)
. It follows that it
hasa density kernelbounded by
1
withrespe t toν
;and this amountstosay thatK
has a density kernel with respe t toµ
bounded above byV (x)V (y)
, sin e the kernel ofK
1
with respe t toν
isk(x, y)
V (x)V (y)
, wherek
isthe kernel ofK
with respe t toµ
.Observe that Theorem 2.4 produ es non uniform bounds on the kernel. Moreover,
when
V ∈ L
2
(µ)
,then theoperator
P
2t
isHilbert-S hmidtsohas adis retespe trumand weget anestimateon the eigenvalues−λ
n
ofL
:X
n
e
−λ
n
t
≤ K
2
(t)e
ct
kV k
2
2
.
3 Weighted Nash inequalities on the real line.
As already mentioned, we shall mainly on entrate on model examples on the real line,
and show elementary te hniques toobtain weighted Nash inequalities for measures with
density
ρ
with respe t tothe Lebesgue measure and the usual arré du hampΓ(f, f ) =
|∇f|
2
= f
′
2
. These te hniques may be easily extended to the
n
-dimensional Eu lideanspa e,and with some extra work toRiemannian manifolds.
Let us rst state a universal weighted Nash inequality in the Eu lidean spa e. We
onsider the ase when
Γ(f, f ) = |∇f|
2
and
µ(dx) = ρ(x)dx
. We are mainly interested inthe ase whenµ
is aprobability measure. Re all that in this situation,there maynot existany lassi alNashinequality ( lassi almeanswithapowerfun tionasratefun tionΦ
) unless the measure is ompa tlysupported.Here, the symmetri operator asso iatedwith the orresponding Diri hlet formis
Lf = ∆f + ∇ log ρ · ∇f.
We may always hoose
V = ρ
−1/2
: it is not hard to he k that
LV ≤ cV
for someonstant
c
toget the universal weighted Nash inequality (with respe t toµ
)Here
C
n
isthe onstant forthe Nashinequalityinthe Eu lideanspa e withthe Lebesgue measure.Tosee this, wejust apply theEu lidean Nashinequality (0.1)to
g = f √ρ
, wheref
is asmooth ompa tly supported fun tion, and observe thatZ
R
n
|∇g|
2
dx =
Z
|∇f|
2
ρdx +
Z
R
n
LV
V
f
2
dµ = E(f, f) +
Z
R
n
LV
V
f
2
dµ,
through integration by parts. Unfortunately, this bound is not very useful sin e
V /
∈
L
2
(µ)
. Nevertheless, with some are to justify the integration by parts in (2.11), (with
extra hypotheses like uniform upper bounds on the Hessian of
log ρ
), it may lead to an upper-bound onthe kernel density.Of ourse, this method has nothing parti ular to do with the Eu lidean ase. It
extends a Nash inequality (without weight) with respe t to a measure
µ
to a weighted Nashinequality with respe t tothe measureρdµ
with weightV = ρ
−1/2
, as soonas the
inequality
LV ≤ cV
issatised.Forexample, one gets with this simpleargument
Corollary 3.1. In
R
n
, withρ(x) = (1 + |x|
2
)
−β
withβ > n
orρ(x) = exp(−(1 + |x|
2
)
a/2
)
with
a > 0
, there exists a onstantC
su h that for allt > 0
andx, y ∈ R
n
the kernel densityp
t
satisesp
t
(x, y) ≤
C
t
n/2
e
Ct
ρ
−1/2
(x) ρ
−1/2
(y).
Butsin eV /
∈ L
2
(µ)
,thismayneverprodu eanyboundonthespe trumforexample.
Soone has to look for more pre iseLyapunov fun tions.
This is what we now perform on our model examples on the real line : we write
T (x) =
√
1 + x
2
and onsider the measure
µ
a
(dx) = C
a
exp(−T (x)
a
)dx,
where
a > 0
andC
a
is a normalizing onstant. We denote byρ
a
the densityexp(−T
a
)
.
Here, the asso iated operatoris
L(f ) = f
′′
− aT
a−1
T
′
f
′
.
In this ontext, it isnot hardto he k that, for any
β ∈ R
,(3.12)
V = T
−β
/
√
ρ
a
isaLyapunov fun tion. If
β > 1/2
,this fun tionisinL
2
(µ
a
)
. The issueisthento hoosethe smallest possible
V ∈ L
2
(µ
a
)
and still have a weighted Nash inequality with rate fun tionΦ
su h that1/Φ
isintegrable atinnity.Theorem 3.2 ([2℄). If
a > 1
, then for anyβ ∈ R
andV
hosen as in (3.12), there exist onstantsC
andλ ∈ (0, 1)
su h that(3.13)
kfk
2
2
≤ C
"
Z
|f|V dµ
a
2
+
Z
|f|V dµ
a
2(1−λ)
E(f, f)
λ
#
for all fun tions
f
. This orresponds to the rate fun tionΦ(x) =
x
C
− 1
1/λ
, x > C.
Although tra table, the expli it value of
λ
in terms of the parametersa
andβ
isnot sosimple. The assumptiona > 1
is ne essary, sin e fora ≤ 1
the spe trum is no longer dis rete(and therefore noweighted Nash inequality ould o ur with anyL
2
(µ
a
)
weightV
). What has tobe underlined here is thatthe introdu tionof a weightallows ustoget polynomialratefun tionsΦ
,althoughwe knowthatsu hpolynomialgrowthisforbidden fornon ompa tly supported nite measures in the absen e of weights. Of ourse,togetthesepolynomialgrowths,onehasto hooseweightswhi harequite losetotheuniversal
weights
1/√ρ
des ribed before. If one hooses mu h smaller weights, the rate fun tion willbe smaller. For example, whena > 2
, one may hooseV = 1
, and in this ase one hasΦ(x) = x(log x)
α
.
The argument of Theorem 3.2 is based on a tail estimate of the measure
µ
a
. Ifq
a
(x) =
R
∞
x
µ
a
(dy)
,then, for some onstantC
,one has(3.14)
q
a
(x) ≤ C
ρ
a
(x)
T (x)
a−1
.
One rst proves a Nash inequality for smooth ompa tly supported fun tions su h
that
f (0) = 0
. We start withLemma3.3. Let
a ≥ 1
,β ∈ R
andV
givenin (3.12). Forallsmooth ompa tlysupported fun tionsf
su hthatf (0) = 0
one hasZ
f
2
dµ
a
≤ CE(f, f)
γ
Z
|f|V dµ
a
2(1−γ)
whereγ = 1 − 2
a − 1
3(a − 1) + 2β
∈
1
3
, 1.
The proof isbased on uttingthe integralon
[0, ∞)
(forinstan e) asZ
∞
0
f
2
dµ
a
=
Z
∞
0
f
2
1l
n
f
kf k 2
≤V Z
−1/2
o
dµ
a
+
Z
∞
0
f
2
1l
n
f
kf k 2
>V Z
−1/2
o
dµ
a
.
for a suitably hosen
Z > 0
. Then both terms are ontrolled by the estimate (3.14), repla ingf
2
by2
R
x
0
f (t)f
′
(t)dt
inthe se ond integral and using Fubini's theorem.
Lemma 3.4. Let
a > 0
,β >
3−a
2
andV
given in (3.12). Then there existθ ∈ (0, 1)
andC
su h thatZ
|f − f(0)|V dµ
a
≤ C
"
Z
|f|V dµ
a
+
Z
|f|V dµ
a
1−θ
E(f, f)
θ/2
#
for all nonnegative smooth ompa tly supported
f
onR
.Although quite similar, this lemma is more restri tive on the values of
β
than the previous one. Passing from fun tions whi h vanish in0
to the general ase is indeed the hardstep. Wereferthe readerto[2℄fordetailsonthe proofs. Itremainstoplug togetherthose inequalitiestoobtain Theorem 3.2.
Corollary 3.5. Let
a > 1
and let(P
t
)
t≥0
be the Markov generator onR
with generatorLf = f
′′
− aT
a−1
T
′
f
′
,
and reversible measure
dµ
a
(x) = ρ
a
(x)dx = C
a
exp(−(1 + |x|
2
)
a/2
)dx
.
Then for all real
β
there existδ > 0
andC
su h that, for allt
,P
t
has a densityp
t
with respe t to the measureµ
a
, whi h satisesp
t
(x, y) ≤
Ce
Ct
t
δ
ρ
−1/2
a
(x)ρ
−1/2
a
(y)
(1 + |x|
2
)
β/2
(1 + |y|
2
)
β/2
for almost every
x, y ∈ R
.Moreover, the spe trum of
−L
is dis rete and its eigenvalues(λ
n
)
n∈N
satisfy the in-equalityX
n
e
−λ
n
t
≤
Ce
Ct
t
δ
for allt > 0
.When
a > 2
,the samete hniques alsolead toaNashinequalityforµ
a
with rate fun -tionΦ(x) = C x (log x)
2(1−1/a)
, and weight
V = 1
. This re overs the ultra ontra tivity resultmentionedearlier. Re allthat whena = 2
thesemigroup isnolonger ultra ontra -tive,but onlyhyper ontra tive,the Nashinequality withrateΦ(x) = x log x
orresponds infa t toanother formof the Logarithmi Sobolevinequality.Referen es
[1℄ D. Bakry. L'hyper ontra tivité et son utilisation en théorie des semigroupes. In
Le tures on probability theory (Saint-Flour, 1992), Le ture Notes in Math. 1581,
[2℄ D.Bakry,F. Bolley,I. Gentil,and P.Maheux. Weighted Nashinequalities. Preprint.
[3℄ D. Bakry, T. Coulhon, M. Ledoux, and L. Salo-Coste. Sobolev inequalities in
disguise. Indiana Univ. Math. J., 44(4):10331074,1995.
[4℄ D.Bakry and M. Ledoux. Sobolevinequalitiesand Myers's diametertheorem foran
abstra t Markov generator. Duke Math. J., 85(1):253270,1996.
[5℄ A. Bendikov, T. Coulhon, and L. Salo-Coste. Ultra ontra tivity and embedding
into
L
∞
. Math. Ann., 337(4):817853, 2007.
[6℄ E.A.CarlenandM. Loss. Sharp onstantinNash'sinequality. Internat. Math.Res.
Noti es, 7:213215, 1993.
[7℄ E. B. Davies. Heat kernels and spe tral theory, volume 92 of Cambridge Tra ts in
Mathemati s. Cambridge University Press, Cambridge, 1990.
[8℄ L.Gross. Logarithmi Sobolev inequalities. Amer. J. Math.,97(4):10611083, 1975.
[9℄ J.Nash. Continuity ofsolutionsof paraboli andellipti equations. Amer. J.Math.,
80:931954,1958.
[10℄ E.Nelson. The free Marko eld. J. Fun tional Analysis,12:211227, 1973.
[11℄ N. Th. Varopoulos. Hardy-Littlewood theory for semigroups. J. Fun t. Anal.,
63(2):240260,1985.
[12℄ F.-Y. Wang. Fun tional inequalitiesfor empty essential spe trum. J. Fun t. Anal.,
170(1):219245,2000.
[13℄ F.-Y. Wang. Fun tional inequalities and spe trum estimates: the innite measure
ase. J.Fun t.Anal.,194(2):288310, 2002.
Institut de Mathématiques de Toulouse,UMR CNRS 5219
Université de Toulouse Routede Narbonne 31062Toulouse - Fran e bakrymath.univ-toulouse.fr Ceremade,UMR CNRS 7534 Université Paris-Dauphine
Pla e du Maré halDe Lattre De Tassigny
75016Paris- Fran e
bolley eremade.dauphine.fr
Institut CamilleJordan, UMR CNRS 5208
43boulevard du 11 novembre 1918
69622Villeurbanne edex - Fran e