ebook img

A dual Moser-Onofri inequality and its extensions to higher dimensional spheres PDF

0.21 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview A dual Moser-Onofri inequality and its extensions to higher dimensional spheres

A dual Moser-Onofri inequality and its extensions to higher dimensional spheres Martial Agueh ∗ Shirin Boroushaki † and Nassif Ghoussoub ‡ 5 DedicatedtoDominiqueBakryontheoccasionofhis60thbirthday 1 0 2 Abstract n a We use optimal mass transport to provide a new proof and a dual formula to the J Moser-OnofriinequalityonS2 inthesamespiritastheapproachofCordero-Erausquin, 6 Nazaret and Villani [4] to the Sobolev inequality and of Agueh-Ghoussoub-Kang[1] to moregeneralsettings. Therearehowevermanyhurdlestoovercomeonceastereographic ] P projection onR2 isperformed: Functionsarenotnecessarily ofcompactsupport,hence A boundary terms need to be evaluated. Moreover, the corresponding dual free energy of the reference probability density µ (x)= 1 is not finite on the whole space, . 2 π(1+|x|2)2 h whichrequirestheintroductionofarenormalizedfreeenergyintothedualformula. We t a also extend this duality to higher dimensions and establish an extension of the Onofri m inequality to spheres Sn with n ≥ 2. What is remarkable is that the corresponding [ free energy is again given by F(ρ) = −nρ1−n1, which means that both the prescribed scalar curvature problemandtheprescribed Gaussian curvature problemleadessentially 1 tothesame dualproblem whose extremals are stationary solutions of thefast diffusion v equations. 7 6 Keywords: Onofri inequality, duality, mass transport, prescribed Gaussian curvature, 2 fast diffusion. 1 0 Mathematics subject classification: 39B62,35J20, 34A34, 34A30. . 1 0 5 1 Introduction 1 : v One of the equivalent forms of Moser’s inequality [10] on the 2-dimensional sphere S2 states i that the functional X r 1 a J(u):= u2dω udω log eudω (1.1) 4ZS2|∇ | −ZS2 − (cid:18)ZS2 (cid:19) is bounded below on H1(S2), where dω is the Lebesgue measure on S2, normalized so that dω = 1. Later, Onofri [11] showed that the infimum is actually zero, and that modulo S2 conformaltransformations,u=0istheoptimalfunction. Notethatthisinequalityisrelated tRo the “prescribed Gaussian curvature” problem on S2, ∆u+K(x)e2u =1 on S2, (1.2) ∗Department of Mathematics and Statistics University of Victoria, Victoria, BC, V8W 3R4, CANADA, [email protected]. †DepartmentofMathematics,1984MathematicsRoad,UniversityofBritishColumbia,BC,V6T1Z2,CANADA [email protected]. ‡DepartmentofMathematics,1984MathematicsRoad,UniversityofBritishColumbia,BC,V6T1Z2,CANADA [email protected] 1 whereK(x)istheGaussiancurvatureassociatedtothemetricg =e2ug onS2,and∆=∆ 0 g0 is the Laplace-Beltrami operator corresponding to the standard metric g . Finding g for a 0 given K leads to solving (1.2). Variationally, this reduces to minimizing the functional dV dV dV (u)= u2 0 +2 u 0 log K(x)e2u 0 onH1(S2), (1.3) F ZS2|∇ | 4π ZS2 4π − (cid:18)ZS2 4π (cid:19) wherethevolumeformissuchthat dV =4π. Onofri’sresultsaysthat,moduloconformal S2 0 transformations, u 0 is the only solution of the “prescribed Gaussian curvature” problem (1.2) for K =1, i.e.≡, 1∆u+eu =1Ron S2, which after rescaling, u 2u, gives 2 7→ ∆u+e2u =1 on S2. (1.4) The proof given by Onofri in [11] makes use of an “exponential” Sobolev inequality due to Aubin [2] combined with the invariance of the functional (1.1) under conformal transforma- tions. Other proofs were given by Osgood-Philips-Sarnak [12] and by Hong [8]. See also Ghoussoub-Moradifam[7]. In this paper we shall give a proof based on mass transport. While this approachhas by nowbecomestandard,therearemanyreasonswhyithasn’tbeensofarspelledout. Thefirst originates in the fact that, unlike the case of Rn, optimal mass transport on the sphere is harder to work with. To avoidthis difficulty, we use an equivalent formulation of the Onofri inequality (1.1), that we call the Euclidean Onofri inequality on R2, namely 1 u2dx+ udµ log eudµ 0 u H1(R2), (1.5) 2 2 16π ZR2|∇ | ZR2 − (cid:18)ZR2 (cid:19)≥ ∀ ∈ which is obtained by projecting (1.1) on R2 via the stereographic projection with respect to the North pole N =(0,0,1),i.e., Π:S2 R2, Π(x):= x1 , x2 where x=(x ,x ,x ). Here µ is the probability density on R2→defined by µ (x(cid:16))1−=x3 1−1x3(cid:17) , and dµ =1µ (2x)d3x. 2 2 π(1+x2)2 2 2 | | One can then try to apply the Cordero-Nazaret-Villani [4] approach as generalized by Agueh-Ghoussoub-Kang [1] and write the Energy-Entropy production duality for functions that are of compact support in Ω, 1 sup (F(ρ)+ x2ρ)dx; ρdx=1 (1.6) − 2| | (cid:26) ZΩ ZΩ (cid:27) =inf α u2 G(ψ u) dx; ψ(u)dx=1 , |∇ | − ◦ (cid:26)ZΩ ZΩ (cid:27) where G(x)=(1 n)F(x)+nxF (x) and where ψ and α are also computable from F. ′ Recall that by−choosing F(x) = nx1 1/n and ψ(t) = t2∗ where 2 = 2n and n > 2, − − | | ∗ n 2 one obtains the following duality formula for the Sobolev inequality − 1 sup n ρ1 1/ndx x2ρdx; ρdx=1 − n ZRn − 2ZRn| | ZRn o 2 =inf 2 n−1 u2dx; u2∗dx=1 , (1.7) n (cid:18)n−2(cid:19) ZRn|∇ | ZRn| | o where u and ρ have compact support in Rn. The extremal u and ρ are then obtained as ∞ ∞ solutions of |x|2 n−1 =0, ρ =u2∗ (Rn). (1.8) ∇ 2 − ρ1/n ! ∞ ∞ ∈P ∞ The bestconstants arethen obtainedby computing ρ from(1.8)andinserting it into (1.7) ∞ in such a way that 2 inf 2 n−1 u2dx: u2∗dx=1 =n ρ1−1/ndx 1 x2ρ dx. n (cid:18)n−2(cid:19) ZRn|∇ | ZRn| | o ZRn ∞ − 2ZRn| | ∞ 2 Note that this duality leads to a correspondencebetween a solutionto the Yamabe equation ∆u= u2∗−2u on Rn (1.9) − | | and stationary solution to the rescaled fast diffusion equation ∂tρ=∆ρ1−n1 +div(xρ) on Rn. (1.10) The abovescheme doesnothoweverapply to inequality(1.5). Forone, the functions euµ = 2 eu(x) do not have compact support, and if one restricts them to bounded domains, we π(1+x2)2 then|n|eed to take into consideration various boundary terms. What is remarkable is that a similar program can be carried out provided the dual formula involving the free energy J (ρ)= (F(ρ)+ x2ρ)dx Ω − | | ZΩ is renormalized by substituting it with J (ρ) J (µ ). Ω Ω 2 − Another remarkable fact is that the corresponding free energy turned out to be F(ρ) = 2ρ21, which is the same as the one associated to the critical case of the Sobolev inequality −F(ρ) = nρ1−n1 when n 3. In other words, the Moser-Onofri inequality and the Sobolev − ≥ inequality“dualize”inthesameway,andboththeYamabeproblem(1.9)andtheprescribed Gaussian curvature problem (1.4) reduce to the study of the fast diffusion equation (1.10), with the caveatthat in dimension n=2, the above equation needs to be considered only on bounded domains, with Neumann boundary conditions. More precisely, we shall show that, when restricted to balls B of radius R in R2, there R is a duality between the “Onofri functional” I (u)= 1 u2dx+ udµ on X := u H1(B ); eudµ = dµ R 16π BR|∇ | BR 2 R { ∈ 0 R BR 2 BR 2} and the free eRnergy R R R JR(ρ)= √2π BR√ρdx− BR|x|2ρdx on YR :={ρ∈P(R2); BRρdx= BR dµ2}. R R R R Once the latter free energy is re-normalized, we then get sup J (ρ) J (µ );ρ Y =0=inf I (u); u X . (1.11) R R 2 R R R { − ∈ } { ∈ } Note that when R + , the right hand side yields the Onofri inequality → ∞ 1 inf u2dx+ udµ ; u H1(R2), eudµ =1 =0, (cid:26)16π ZR2|∇ | ZR2 2 ∈ 0 ZR2 2 (cid:27) while the left-hand side becomes undefined since J (µ ) + as R . R 2 → ∞ →∞ We also show that our approach generalizes to higher dimensions. More precisely, if B R is a ball of radius R in Rn where n 2, and if one considers the probability density µ on n Rn defined by ≥ n µ (x)= n ωn(1+ x n−n1)n | | (ω is the area of the unit sphere in Rn), and the operator H (u,µ ) on W1,n(Rn) by n n n H (u,µ ):= u+ (logµ )n (logµ ))n n (logµ ))n 2 (logµ ) u, n n n n n − n |∇ ∇ | −|∇ | − |∇ | ∇ ·∇ there is then a duality between the functional 1 I (u)= H (u,µ )dx+ u(x)dµ on X := u W1,n(B ); eudµ = dµ R β(n) n n n R { ∈ 0 R BR n BR n} ZBR ZBR R R 3 and the free energy – renormalized by again subtracting J (µ ) – R n JR(ρ)=α(n) ρ(x)n−n1 dx− |x|nn−1ρ(x)dx on YN :={ρ∈P(Rn); BRρdx= BR dµn}. ZBR ZBR R R where α(n)= n ( n )1/n and β(n)=ω ( n )n 1nn. We then deduce the following higher dimensional venr−si1onωonf the Euclidean Onnofrni−in1eq−uality in Rn, n 2: ≥ 1 H (u,µ )dx+ udµ log eudµ 0 for all u W1,n(Rn), (1.12) p n n n β(n)ZRn ZRn − (cid:18)ZRn (cid:19)≥ ∈ and u 0 is the extremal. ≡ We finish this introduction by mentioning that the Euclidean Onofri inequality (1.5) in theradialcasewasrecentlyobtainedin[6]usingmasstransporttechniques. However,toour knowledge, our duality result, the extensions of Onofri’s inequality to higher dimensions, as well as the mass transport proof of the general (non-radial) Onofri inequality are new. The paper is organized as follows. In section 2, we recall the mass transport approach to sharp Sobolev inequalities and some consequences. In section 3, we establish the mass transportdualityprincipleleadingtotheEuclideanOnofriinequality(1.5). Finally,insection 4, we extend the 2-dimensional analysis of section 3 to higher dimensions n 2, and we ≥ producean-dimensionalversionoftheEuclideanOnofriinequality,aswellasotherinteresting inequalities. 2 Preliminaries We start by briefly describing the mass transport approach to sharp Sobolev inequalities as proposed by [4]. We will follow here the framework of [1] as it clearly shows the correspon- dence between the Yamabe equation (1.9) and the rescaled fast diffision equation (1.10). Let ρ ,ρ (Rn). If T is the optimal map pushing ρ forward to ρ (i.e. T ρ = ρ ) 0 1 0 1 # 0 1 ∈ P in the mass transport problem for the quadratic cost c(x y)= |x−y|2 (see [14] for details), then [0,1] t ρ = (T ) ρ is the “geodesic” joinin−g ρ and2ρ in ( (Rn),d ); here t t # 0 0 1 2 ∋ 7→ P T :=(1 t)id+tT and d denotes the quadratic Wasserstein distance (see [14]). Moreover, t 2 given a f−unction F :[0, ) R such that F(0)=0 and x xnF(x n) is convex and non- − ∞ → 7→ increasing, the functional HF(ρ):= F (ρ(x)) dx is displacement convex [9], in the sense Rn that[0,1] t HF(ρ ) R is convex(in the usualsense), for allpairs(ρ ,ρ ) in (Rn). A t 0 1 ∋ 7→ ∈ R P direct consequence is the following convexity inequality, known as “energy inequality”: d HF(ρ1) HF(ρ0) HF(ρt) = ρ0 (F′(ρ0)) (T id) dx, − ≥(cid:20)dt (cid:21)t=0 ZRn ∇ · − which, after integration by parts of the right hand side term, reads as HF(ρ1) HF+nPF(ρ0) ρ0 (F′(ρ0)) T(x)dx, (2.1) − ≤− −ZRn ∇ · where P (x) = xF (x) F(x); here id denotes the identity function on Rn. By the Young F ′ − inequality F (ρ )p T(x)q 1 1 ′ 0 (F′(ρ0)) T(x) |∇ | + | | p,q >1 such that + =1, (2.2) −∇ · ≤ p q ∀ p q (2.1) gives 1 1 HF(ρ1) HF+nPF(ρ0)+ ρ0 F′(ρ0)pdx+ ρ0(x)T(x)qdx, − ≤− pZRn |∇ | q ZRn | | 4 i.e., 1 1 HF(ρ1) y qρ1(y)dy HF+nPF(ρ0)+ ρ0 F′(ρ0)pdx, (2.3) − − q ZRn| | ≤− pZRn |∇ | where we use that T ρ = ρ . Furthermore, if ρ = ρ , then T = id and equality holds in # 0 1 0 1 (2.1). Thenequalityholdsin(2.3)ifitholdsintheYounginequality(2.2). Thisoccurswhen ρ0 = ρ1 satisfies ∇ F′(ρ0(x))+ |xq|q = 0. Therefore, we have established the following duality: (cid:16) (cid:17) 1 sup HF(ρ ) y qρ (y)dy : ρ (Rn) 1 1 1 n− − q ZRn| | ∈P o 1 =inf HF+nPF(ρ0)+ ρ0 F′(ρ0)pdx: ρ0 (Rn) , (2.4) n− pZRn |∇ | ∈P o and an optimal function in both problems is ρ =ρ :=ρ solution of 0 1 ∞ xq F′(ρ (x))+ | | =0. (2.5) ∇ ∞ q (cid:18) (cid:19) In particular, choosing F(x) = nx1 1/n and ρ = u2∗ where 2 = 2n and n > 2, then − − 0 ∗ n 2 HF+nPF =0, and (2.4)-(2.5) gives the duality formula for the Sobolev i−nequality (1.7). Our goal now is to extend this mass transport proof of the Sobolev inequality to the EuclideanOnofri inequality (1.5). As already mentioned in the introduction, a first attempt on this issue was recently made by [6], but the result produced was only restricted to the radial case. Here we show in full generality (without restricting to radial functions u) that the Euclidean Onofri inequality (1.5) can be proved by mass transport techniques. More precisely, we establish an analogue of the duality (1.7) for Euclidean Onofri inequality (see Theorem 3.1), from which we deduce the Onofri inequality (1.5) and its optimal function (see Theorem 3.3). Furthermore, we obtain as for the Sobolev inequality, a correspondence between the prescribed Gaussian curvature problem (1.4) and the rescaled fast diffusion equation (1.10). Finally, we extend our analysis to higher dimensions, and then produce a new version of the Onofri inequality in dimensions n 2 (see Theorem 4.3). ≥ The following general lemma from mass transport theory will be needed in the sequel. Lemma 2.1. Let ρ ,ρ (B ), where (B ) denotes the set of probability densities on 0 1 R R the ball B Rn. Let T∈bePthe optimal mPap pushing ρ forward to ρ (i.e. T ρ = ρ ) in R 0 1 # 0 1 ⊂ the mass transport problem for the quadratic cost. Then 1 ρ1(y)1−n1 dy ρ0(x)1−n1div(T(x))dx. (2.6) ≤ n ZBR ZBR Proof. By [3], the optimal map T : B B is such that T = ϕ where ϕ : B R is R R R → ∇ → convex. Since T ρ =ρ , then it satisfies the Monge-Amp`ere equation # 0 1 ρ (x)=ρ (T(x))det T(x) (2.7) 0 1 ∇ or equivalently ρ (T(x))=ρ (x)[det T(x)] 1. (2.8) 1 0 − ∇ By the arithmetic-geometric-meaninequality 1 [det T(x)]n1 div(T(x)), ∇ ≤ n 5 (2.8) gives 1 ρ1(T(x))−n1 ρ0(x)−n1div(T(x)). (2.9) ≤ n Now using the change of variable y =T(x), we have ρ1(y)1−n1 dy = ρ1(T(x))1−n1 det( T(x))dx, ∇ ZBR ZBR which implies by (2.7) and (2.9), 1 1 ρ1(y)1−n1 dy ρ0(x)−n1div(T(x))ρ0(x)dx= ρ0(x)1−n1div(T(x))dx. ≤ n n ZBR ZBR ZBR 3 Euclidean 2-dimensional Onofri inequality: a duality formula Consider the probability density 1 µ (y)= y R2, 2 π(1+ y 2)2 ∀ ∈ | | and set R2 θ := µ (y)dy = , (3.1) R 2 1+R2 ZBR where B R2. We now establish the following duality. R ⊂ Theorem 3.1. (Duality for the 2-dimensional Euclidean Onofri inequality) Consider the functionals 1 I (u)= u(x)2dx+ u(x)dµ (x) on H1(B ), R 16π |∇ | 2 0 R ZBR ZBR and 2 J (ρ)= ρ(y)dy y 2ρ(y)dy on (R2). R √π − | | P ZBR ZBR p Then, the following holds: sup J (ρ) J (µ ); ρ (R2), ρdy =θ R R 2 R − ∈P (cid:26) ZBR (cid:27) =inf I (u); u H1(B ), eudµ =θ . (3.2) R ∈ 0 R 2 R (cid:26) ZBR (cid:27) Moreover, the maximum on the l.h.s. is attained at ρ = µ , and the minimum on the max 2 r.h.s. is attained at u = 0. Furthermore, the maximum and minimum value of both min functionals in (3.2) is 0. Proof. Applying lemma 2.1 in dimension n=2, and integrating by parts, we get 2 ρ (y) dy ρ (x) div(T(x))dx 1 0 ≤ ZBR ZBR (3.3) p p = (√ρ ) T(x)dx+ ρ (x) T(x) νdS 0 0 − ∇ · · ZBR Z∂BR p 6 where ν = x/x denotes the outward unit normal to the boundary ∂B of B , and dS is R R | | the surface element. By the basic identity 1 (√ρ )= √ρ (logρ ), 0 0 0 ∇ 2 ∇ we can rewrite (3.3) as 4 ρ (y)dy √ρ (logρ ) T(x)dx+2 ρ (x) T(x) νdS. (3.4) 1 0 0 0 ≤− ∇ · · ZBR ZBR Z∂BR p p Now introduce the probability density eu(x)µ (x) 2 ρ (x)= x B , 0 R θ ∀ ∈ R where u H1(B ) satisfies eudµ =θ . Then (3.4) becomes ∈ 0 R BR 2 R R 4 θ ρ dy √euµ (log(euµ )) T(x)dx+2 √euµ T(x) νdS. (3.5) R 1 2 2 2 ≤− ∇ · · ZBR ZBR Z∂BR p We use Young inequality 1 ε √euµ (log(euµ )) T(x) (log(euµ )) 2+ euµ T(x)2 with ε=4√π, 2 2 2 2 − ∇ · ≤ 2ε|∇ | 2 | | and the fact that T ρ =ρ , to rewrite (3.5) as # 0 1 32 πθ ρ dy 16π y 2θ ρ (y)dy 16 √πeuµ T(x) νdS R 1 R 1 2 − | | − · ZBR ZBR Z∂BR p u+ (logµ )2dx. (3.6) 2 ≤ |∇ ∇ | ZBR It is easy to check that u+ (logµ ))2dx= u2dx+ (logµ )2dx 2 u∆(logµ )dx, 2 2 2 |∇ ∇ | |∇ | |∇ | − ZBR ZBR ZBR ZBR and ∆(logµ )= 8πµ . 2 2 − Inserting these identities into (3.6) and using that u =0, we have |∂BR 32 πθ ρ dy 16π y 2θ ρ (y)dy 16 √πµ T(x) νdS R 1 R 1 2 − | | − · ZBR ZBR Z∂BR p u2dx+16π udµ + (logµ )2dx. (3.7) 2 2 ≤ |∇ | |∇ | ZBR ZBR ZBR We now estimate the boundary term. Since T : B B , then T(x) R for all x B , R R R → | |≤ ∈ so that T(x) xdS T(x) x dS R2 dS =2πR3. · ≤ | || | ≤ Z∂BR Z∂BR Z∂BR Then using that ν =x/x, we have | | 1 2πR2 √πµ T(x) νdS = T(x) xdS = 2πθ . (3.8) − 2 · −R(1+R2) · ≥−1+R2 − R Z∂BR Z∂BR 7 We insert (3.8) into (3.7) to have 32 πθ ρ dy 16π y 2θ ρ (y)dy 32πθ R 1 R 1 R − | | − ZBR ZBR p u2dx+16π udµ + (logµ )2dx. (3.9) 2 2 ≤ |∇ | |∇ | ZBR ZBR ZBR By a direct computation, we have (logµ )2dx=16 √πµ dy 16πθ . 2 2 R |∇ | − ZBR ZBR This implies that 32 πθ ρ dy 16π y 2θ ρ (y)dy 16 √πµ dy 16πθ R 1 R 1 2 R − | | − − ZBR ZBR ZBR p u2dx+16π udµ , 2 ≤ |∇ | ZBR ZBR that is, after dividing by 16, 2 πθ ρ dy π y 2θ ρ dy √πµ dy+πθ R 1 R 1 2 R − | | − ZBR ZBR (cid:18)ZBR (cid:19) p 1 u2dx+π udµ . (3.10) 2 ≤ 16 |∇ | ZBR ZBR We easily check that √πµ dy+πθ =2 √πµ dy π y 2µ dy. 2 R 2 2 − | | ZBR ZBR ZBR Inserting this identity into (3.10) yields 2 πθ ρ dy y 2πθ ρ dy 2 √πµ dy y 2πµ dy R 1 R 1 2 2 − | | − − | | ZBR ZBR (cid:18) ZBR ZBR (cid:19) p 1 u2dx+π udµ . (3.11) 2 ≤ 16 |∇ | ZBR ZBR Now, we simplify (3.11) by π and set ρ=θ ρ , to have that R 1 J (ρ) J (µ ) I (u) (3.12) R R 2 R − ≤ forallfunctions uandρ suchthatu H1(B )and eudµ =θ = ρ(y)dy. Observe ∈ 0 R BR 2 R BR that if ρ=µ , then the l.h.s of (3.12) is 0, and so is the r.h.s if u=0. Therefore 2 R R 0 sup J (ρ) J (µ ): ρ(y)dy =θ inf I (u): eudµ =θ 0. R R 2 R R 2 R ≤ ρ (cid:26) − ZBR (cid:27)≤u∈H01(BR)(cid:26) ZBR (cid:27)≤ This completes the proof of the theorem. Remark 3.2. Note that the duality formula (3.2) for Onofri’s inequality is restricted to balls B , while for the Sobolev inequality, it is defined on the whole space Rn, n 3. The reason R ≥ is simply because the free energy 2 J (ρ)= ρ(y)dy y 2ρ(y)dy R √π − | | ZBR ZBR p 8 for(3.2)isnotalwaysfinite onthewholedomainR2,thatiswhenR . Thiscanbe seen →∞ when we choose for example ρ=µ . We then have 2 R2 J (µ )=log(1+R2)+ R 2 1+R2 so that lim J (µ )= . R R 2 →∞ ∞ From Theorem 3.1, we can deduce a certain number of results. One of them is the Euclidean Onofri inequality. Theorem 3.3. (2-dimensional Euclidean Onofri inequality) The following holds: 1 log eudµ udµ u2dx u H1(R2). (3.13) 2 2 (cid:18)ZR2 (cid:19)−ZR2 ≤ 16π ZR2|∇ | ∀ ∈ Moreover u=0 is an optimal function. Proof. By Theorem 3.1, we have for all u H1(B ) such that eudµ =θ = R2 , ∈ 0 R BR 2 R 1+R2 R 1 I (u)= u(x)2dx+ u(x)dµ (x) 0, (3.14) R 2 16π |∇ | ≥ ZBR ZBR with equality if u=0. Letting R , we have θ 1; then (3.14) becomes R →∞ → 1 u(x)2dx+ u(x)dµ (x) 0=log eudµ , 2 2 16π ZR2|∇ | ZR2 ≥ (cid:18)ZR2 (cid:19) for all u H1(R2), with equality if u=0. This completes the proof of the theorem. ∈ Another consequence of Theorem 3.1 is the correspondence between solutions to the prescribed Gaussian curvature problem and stationary solutions to a rescaled fast diffusion equation. Indeed, it is known from the theory of mass transport (see [14] for example) that a maximizer ρ to the supremum problem in (3.2) is a solution to the pde: 1 ∆(√ρ)+div(xρ)=0 in B , R 2√π (with Neumann boundary condition) that is the stationary solution of the rescaled fast dif- fusion equation 1 ∂ ρ= ∆(√ρ)+div(xρ) in B . (3.15) t R 2√π On the other hand, it can be shown that a minimizer u to the infimum problem in (3.2) solves the semilinear pde: 1 ∆u+λµ eu =µ in B , (3.16) 2 2 R 8π where λ is the Lagrange multiplier for the constraint eudµ = θ . The duality formula BR 2 R (3.2) establishes then a correspondence between stationary solutions to the rescaled fast R diffusion equation (3.15) and solutions to the semilinear equation (3.16). 9 4 Extension to higher dimensions Here, we extend the 2-dimensional analysis of section 3 to higher dimensions n 2, thereby ≥ producing a n-dimensional version of the Euclidean Onofri inequality (3.13). We start by proving a n-dimensional version of the duality formula (3.2). Hereafter, we denote by n µ (y)= n ωn(1+ y n−n1)n | | a probability density on Rn, where ω is the area of the unit sphere in Rn. Set n Rn θ := µ (y)dy = . R ZBR n (1+Rnn−1)n−1 Theorem 4.1. (Duality for n-dimensional Euclidean Onofri inequality) Consider the functionals n−1 n JR(ρ)=α(n) ρ(y) n dy y n−1ρ(y)dy − | | ZBR ZBR and 1 I (u)= H (u,µ )dx+ u(x)dµ u W1,n(B ) R β(n) n n n ∀ ∈ 0 R ZBR ZBR where 1/n n 1 α(n)= n n , β(n)= n − nnω n n 1 ω n 1 − (cid:18) n(cid:19) (cid:18) − (cid:19) and Hn(u,µn):= u+ (logµn)n (logµn)n n (logµn)n−2 (logµn) u. |∇ ∇ | −|∇ | − |∇ | ∇ ·∇ Then the following duality holds: sup J (ρ) J (µ );ρ (Rn), ρdy =θ =inf I (u); u W1,n(B ), eudµ =θ . R − R n ∈P R R ∈ 0 R n R (cid:26) ZBR (cid:27) (cid:26) ZBR (cid:27) (4.1) Moreover, the maximum on the l.h.s. is attained at ρ = µ , and the minimum on the max n r.h.s. is attained at u = 0. Furthermore, the maximum and minimum value of both min functionals in (4.1) is 0. Remark 4.2. Beforeprovingthetheorem,wemakeafewremarksonthe operatorH (u,µ ). n n 1. Consider the function c:Rn R, c(z)= z n, n 2. Clearly c is strictly convex, and → | | ≥ c(z)=nz n 2z. So we have the convexity inequality − ∇ | | c(z ) c(z ) c(z ) (z z ) 0 z ,z Rn. (4.2) 1 0 0 1 0 0 1 − −∇ · − ≥ ∀ ∈ Setting z = (logµ ) and z = u+ (logµ ), we see that H (u,µ ) is nothing but 0 n 1 n n n ∇ ∇ ∇ the l.h.s of (4.2); we then deduce that H (u,µ ) 0 u,µ . n n n ≥ ∀ 2. Forallu W1,n(B ),theintegralofH (u,µ )overB involvesawell-knownoperator, ∈ 0 R n n R the n-Laplacian ∆ , defined by n ∆nv :=div( v n−2 v). (4.3) |∇ | ∇ Indeed, this can be seen after performing an integration by parts in the last term of H (u,µ ), n n H (u,µ )dy = u+ logµ ndy (logµ ))ndy+n u∆ (logµ )dy. n n n n n n |∇ ∇ | − |∇ | ZBR ZBR ZBR ZBR 10

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.