# Full text of "Rearrangement inequalities for functionals with monotone integrands"

## See other formats

C3 Rearrangement inequalities for functionals with monotone integrands § ■ Almut Burchard* and Hichem Hajaiej^ (N b : June 2004; final revision March 2006. < Dedicated to Albert Baernstein, II on the occasion of his 65 th birthday. Abstract < C$ , The inequalities of Hardy-Littlewood and Riesz say that certain integrals involving products of two or three functions increase under symmetric decreasing rearrangement. It is known that these inequalities extend to integrands of the form F(ui, ■ ■ ■ , u m ) where F . is supermodular; in particular, they hold when F has nonnegative mixed second deriva- tives didjF for all i ^ j. This paper concerns the regularity assumptions on F and the equality cases. It is shown here that extended Hardy-Littlewood and Riesz inequalities £T) ■ are valid for supermodular integrands that are just Borel measurable. Under some non- degeneracy conditions, all equality cases are equivalent to radially decreasing functions ^ | under transformations that leave the functionals invariant (i.e., measure-preserving maps for the Hardy-Littlewood inequality, translations for the Riesz inequality). The proofs rely on monotone changes of variables in the spirit of Sklar's theorem. 1 Introduction The systematic study of rearrangements begins with the final chapter of "Inequalities" by Hardy, Littlewood, and Poly a 0. Two inequalities are discussed there at length, the Hardy-Littlewood inequality (Theorems 368-370 and 378 of u(x)v(x)dx< / u* (x)v* (x) dx , (1.1) Jr and the Riesz rearrangement inequality (QIIH, Theorem 370 of [ 1 ]) u{x)v{x')w(x — x) dxdx < I I u* (x)v* (x')w* (x — x') dxdx' . (1.2) "University of Toronto, Toronto, Canada M5S 2E4. almut@math . utoronto . ca. t University of Virginia, Charlottesville, VA 22904-4137, USA. hichem . ha jaie j @gmail . com. Here, u, v, and w are nonnegative measurable functions that vanish at infinity, and u* , v* , and w* are their symmetric decreasing rearrangements. The Hardy-Littlewood inequality is a very basic inequality that holds, with suitably defined rearrangements, on arbitrary measure spaces [4]. Its main implication is that rear- rangement decreases L 2 -distances 0. In contrast, the Riesz rearrangement inequality is specific to Z and to M. n , where it is closely related with the Brunn-Minkowski inequality of convex geometry. The generalization of Eq. dl.2t from E to IR n is due to Sobolev [6|, and the inequality is also known as the Riesz-Sobolev inequality. For many applications, the third function in Eq. dl.2t is already radially decreasing, i.e., w(x — x') = K(\x — x'\) with some nonnegative nonincreasing function K, such as the heat kernel or the Coulomb kernel (Theorems 371-373 and 380 of Q). This special case of the inequality also holds on the standard spheres and hyperbolic spaces ElUJ, and it still contains the isoperimetric inequality as a limit. It is a natural question whether these inequalities carry over to more general integral functionals. Under what conditions on F do the extended Hardy -Litdewood inequality j F(u 1 (x),...,u m (x))dx < J F(u*(x), . . . ,u* n (x)) dx and the extended Riesz inequality j ■ ■■ j F{ui{x\), . . .,u m (x m )) Y\Kij(d(x i ,x j ))dx 1 . ..dx m i<j < J-J F(u* 1 (x 1 ),...,u^ n (x m ))Y\Kij(d(x i ,x j ))dxi...dx r (1.3) (1.4) hold for all choices of u\,... ,u m ? In Eq. dl.4t the Kij are given nonnegative nonin- creasing functions on M + , and d(x, y) denotes the distance between x and y. Eq. (I1.3t can be recovered from Eq. dl.4t by choosing K^j as a Dirac sequence and passing to the limit. Note that Eq. dl.4t contains only the case of Eq. dl.2t where the third function is a symmetric decreasing kernel. A larger class of integral kernels K{x\, . . . , x m ) was con- sidered in [9|. The full generalization of Riesz' inequality to products of more than three functions was found by Brascamp-Lieb-Luttinger [ 10 1 ; again, one may ask to what class of integrands the Brascamp-Lieb-Luttinger inequality naturally extends. The main condition on F was identified by Lorentz lITTl as the second-order mono- tonicity property F(y + he, + kej) + F(y) > F(y + he,) + F(y + kej) (i + j, h,k > 0) , (1.5) where y = (yi, . . . , y m ), and e, denotes the i-th standard basis vector in ]R m . Functions satisfying Eq. (I1.5t are called supermodular or 2-increasing in Economics. A smooth func- tion is supermodular, if all its mixed second partial derivatives are nonnegative. Eqs. (I1.3t and il.4l were proved for continuous supermodular integrands depending onm = 2 func- tions by Crowe-Zweibel-Rosenbloom [12| and Almgren-Lieb (Theorem 2.2 of 11731 ). For m > 2, Eq. (I1.3t is due to Brock 1 14 1, and Eq. (I1.4t is a recent result of Draghici [ 15 1. The 2 purpose of this paper is to dispense with the continuity assumptions on F in the theorems of Brock and Draghici, and to characterize the equality cases in some relevant situations. This continues prior work of the second author [16-19]. Acknowledgments. We thank Friedemann Brock, Cristina Draghici, and Loren Pitt for useful discussions, and especially Al Baernstein for drawing our attention to Sklar's the- orem. A.B. was partially funded by grants from the National Science Foundation (NSF), the National Sciences and Engineering Research Council of Canada (NSERC), and a Uni- versity of Toronto Connaught award. H.H. was supported by the Fonds National Suisse de la Recherche Scientifique (FNS). 2 Statement of the results Let X denote either the Euclidean space R n , the sphere S n , or the hyperbolic space EP, equipped with the standard distance function d(-, •) and the uniform volume measure A. Choose a distinguished point x* £ X to serve as the origin or the north pole. Consider a nonnegative measurable function u on X. When X = W n or H n , we require u to vanish at infinity in the sense that all its positive level sets {x G X : u(x) > t} have finite measure; when X = S n this requirement is void. By definition, the symmetric decreasing rearrangement u* of u is the unique upper semicontinuous, nonincreasing function of d(x, x*) that is equimeasurable with u. Explicitly, if p(t) = \({x £ X : u(x) > t}) is the distribution function of u, and B r denotes the open ball of radius r centered at x*, then u*(x) := sup{i > : p{t) > X(B d ^ X}X ^)} . Theorem 1 (Extended Hardy-Littlewood inequality.) Eq. M.3\ holds for all nonnes- ative measurable functions u\ , . . . , u m that vanish at infinity on X = W 1 , § n , or M n , provided that the integrand F is a supermodular Borel measurable function on the closed positive cone with F(0) = 0, and that its negative part satisfies F-(y,i(x) eA dx < oo (2.1) for i = I, ... ,m. Suppose Eq. <\1.3t holds with equality, and the integrals are finite. If F satisfies Eq. <\1.5t with strict inequality for some i ^ j, all y G M.™ and all h, k > 0, then (ui(x) - Ui(x')) (uj(x) - Uj(x')) > for almost all x,x' £ X; in particular, if u; L = u* is strictly radially decreasing, then Uj = u*. 3 / The Borel measurability of F and the integrability assumption in Eq. (I2.lt ensure that the integrals in Eq. (11.31) are well-defined, though they may take the value +00. The left hand side of Eq. J1.3i is invariant under volume-preserving diffeomorphisms of X. More generally, if (f2, p) and (ft , //) are measure spaces and r : Q —>■ ft pushes p forward to p! in the sense that p'(A) = p(r~ 1 (A)) for all ^'-measurable subsets A C ft, then / F(u 1 (u),...,u m (u))d(j,(w)= F(ui ot(uj'),... ,u m oT(u'))df/(u/) . Jq Jq' The right hand side of Eq. J1.3i can also be expressed in an invariant form. Define the nonincreasing rearrangement vfr of u as the unique nonincreasing upper semicontinuous function on M + that is equimeasurable with u, u*(£) : = sup{t > : p(t) > . By construction, {u o r)# = for any map r : Q — > ft that pushes p forward to p! . On X = W 1 , S n and H n , the nonincreasing rearrangement is related with the symmetric decreasing rearrangement by u*(x) = (A {Bm x tX *)))- Theorem [l]implies that F{u x {u>\...,u m {u))dn{u)< F(uf(0,...,u*(0)d£ (2.2) in Jo for all nonnegative measurable functions u\ , . . . , u m on U that vanish at infinity. When fi is a probability measure, Eq. i2.2t says that the expected value of F(Y\ , . . . , Y m ) is maximized among all random variables Y\ , . . . , Y m with given marginal distributions by the perfectly correlated random variables Yf, . . . , Ym - The joint distribution of the max- imizer is uniquely determined, if Yi is continuously distributed for some i and Eq. (I1.5t is strict for all j ^ i. In this formulation, the invariance under measure-preserving trans- formations is evident, since the expected value depends only on the joint distribution of Yi, ... ,Y m . The assumption that F is supermodular signifies that each of the random variables enhances the contribution of the others. Theorem 2 (Extended Riesz inequality.) Eq. M.4\ holds for all nonnegative measurable functions u%, . . . , u m on X = W 1 , § n , or HP that vanish at infinity, provided that F is a su- permodular Borel measurable function on WV with F(0) = 0, each Kij is nonincreasing and nonnegative, and the negative part of F satisfies ■ ■■ F^(ue(xe)ei) Y[ K ij(d(xi,Xj)) dxi ... dx m < 00 (2.3) Jx Jx iKj for £ = 1, . . . , m. Suppose Eq. M.4\ holds with equality. Assume additionally that the integrals are finite, and that Kij(t) > Ofor all i < j and all t < diamX. Let Tq be the graph on the vertex set {1, . . . , m} which has an edge between i and j whenever K{j is a strictly decreasing function, and let i 7^ j be from the same component of Tq. If Eq. M.5t is strict for all y G R!? and all h,k > 0, and if ui and Uj are non-constant, then m = u* o r and Uj = Uj o t for some translation r on X. 4 3 Related work There are several proofs of the extended Hardy-Littlewood inequality in the literature. For continuous integrands, Lorentz showed by discretization and elementary manipulations of the Ui that Eq. ( I2.2t holds for all measurable functions u\, . . . , u m on $7 = (0, 1) if and only if F is supermodular |11|. By the invariance under measure-preserving transforma- tions, this implies Eq. (U.3t . as well as Eq. (I2.2t for arbitrary finite measure spaces Q. However, Lorentz' paper has had little impact on subsequent developments. More than thirty years later, Crowe-Zweibel-Rosenbloom proved Eq. dl.3t for m = 2 on X = W 1 lfT2l . They expressed a given continuous supermodular function F on M. 2 ^ that vanishes on the boundary as the distribution function of a Borel measure fj,p, F(Vi,1to) = Mf([0>l/i) x C '^)) • layer-cake representation j F(u 1 (x),u 2 (x))dx = |y l Ul{x)>yi l U2ix)>y2 dx^ d/j, F (y 1 ,y 2 ) , (3.1) which reduces Eq. dl.3l > to the case where F is a product of characteristic functions (see Theorem 1.13 in [20 1). Another reduction to products was proposed by Tahraoui [21 1. The regularity and boundary conditions on F were relaxed by Hajaiej-Stuart, who assumed it to be supermodular, of Caratheodory type (i.e., Borel measurable in the first, continu- ous in the second variable), and to satisfy some growth and integrability restrictions llT6l . Equality statements for their results were obtained by Hajaiej llT7l[T8l . Using a slightly different layer-cake decomposition, Van Schaftingen-Willem recently established Eq. (I2.2t for m = 2, under additional assumptions on F, for any equimeasurable rearrangement that preserves inclusions [22|. The drawback of the layer-cake representation is that for m > 2 it requires an ru- th order monotonicity condition on the integrand, which amounts for smooth F to the nonnegativity of all (non-repeating) mixed partial derivatives 1 19 1. Brock proved Eq. (I1.3t under the much weaker assumption that F is continuous and supermodular [ 14 1. Carrier viewed maximizing the left hand side of Eq. (I2.2t for a given right hand side as an optimal transportation problem where the distribution functions of m, . . . , u m de- fine mass distributions /ij on R, the joint distribution defines a transportation plan, and the functional represents the cost after multiplying by a minus sign [23 1. He showed that the functional achieves its maximum (i.e., the cost is minimized) when the joint distribu- tion is concentrated on a curve in M m that is nondecreasing in all coordinate directions, and obtained Eq. (I2.2t as a corollary. His proof takes advantage of the dual problem of minimizing m „ X / fi(y) d ^(y) i=i over fi,...,f m , subject to the constraint that £ frfa) > F(y 1 , ... ,y m ) for all y%, . . . , y Tl 5 Theorem ^ can be applied to some integrands that depend explicitly on the radial variable 03] ED E3- If G is a function on R + x R™ such that F(y , ...,y m ) := G(yQ , 2/1, ■ ■ ■ ,y m ) satisfies the assumptions of Theorem ^ then G(\x\,ui(x), . . . ,u m (x)) dx < / G(|x|, u*(x), . . . , u* m (x)) dx . (3.2) Hajaiej-Stuart studied this inequality in connection with the following problem in nonlin- ear optics IT6l IT9l . The profiles of stable electromagnetic waves traveling along a planar waveguide are given by the ground states of the energy functional £( u ) = o / \u'\ 2 dx — I G(\x\,u)dx under the constraint ||u||2 = c. Here, x is the position relative to the optical axis, G is de- termined by the index of refraction, and c > is a parameter related to the wave speed [ 24 1 . If the index of refraction of the optical media decreases with \x\, then F(r, y) = G(r _1 , y) satisfies the assumptions of Theorem ^ Then the first integral shrinks under symmetric decreasing rearrangement by the Polya-Szego inequality, the second integral grows by Eq. d3.2t . and the L 2 -constraint is conserved. Thus, one may rearrange any minimizing sequence to obtain a minimizing sequence of symmetric decreasing functions. This is a crucial step in the construction of ground states — if G violates the monotonicity con- ditions, then a ground state need not exist (25J. Hajaiej-Stuart worried about restrictive regularity assumptions, because G may jump at interfaces between layers of different me- dia. The Riesz inequality in Eq. (I1.4t is non-trivial even when F is just a product of two functions. Ahlfors introduced two-point rearrangements to treat this case on X = S 1 [26|, Baernstein-Taylor proved the corresponding result on S ra [7], and Beckner noted that the proof remains valid on M n and R n [ 8 1. When F is a product of m > 2 functions, Eq. (I1.4t has applications to spectral invariants of heat kernels via the Trotter product formula 11271 . This case was settled by Friedberg-Luttinger [28 1, Burchard-Schmuckenschlager [29], and by Morpurgo, who proved Eq. (II At more generally for integrands of the form m F(y 1 ,...,y m ) = ^(j2y^) (3-3) i=l with <I> convex (Theorem 3.13 of 1301 1. In the above situations, equality cases have been determined l3T1l32ll29ll30l . Almgren-Lieb used the technique of Crowe-Zweibel-Rosen- bloom to prove Eq. (I1.4t for m = 2 |13 |. The special case where F(u, v) = <I>(|n — v |) for some convex function was identified by Baernstein as a 'master inequality' from which many classical geometric inequalities can be derived quickly ll33l . Eq. (II .4t for continuous supermodular integrands with m > 2 is due to Draghici lfT31 . 4 Outline of the arguments In their proofs of Eqs. (U.3t and (U.4t . Brock and Draghici showed that the left hand sides increase under two-point rearrangements if F is any supermodular Borel integrand [14 6 fT5l . Then they approximated the symmetric decreasing rearrangement with sequences of repeated two-point rearrangements. Baernstein-Taylor had established that such sequences can be made to converge to the symmetric decreasing rearrangement in a space of contin- uous functions Q, and Brock-Solynin had proved this convergence in L p -spaces [34). To pass to the desired limits, Brock and Draghici assumed that F is continuous and satisfies some boundary and growth conditions. No new proofs of these inequalities will be given here. Rather, we reduce general su- permodular integrands to the known cases of integrands that are also bounded and contin- uous. This reduction needs more care than the usual density arguments, because pointwise a.e. convergence of a sequence of integrands F^ does not guarantee pointwise a.e. con- vergence of the compositions Fk(u±, . . . , u m ). Approximation within a class of functions with specified positivity or monotonicity properties can be subtle; for instance, nonnega- tive functions of m variables cannot always be approximated by positive linear combina- tions of products of nonnegative functions of the individual variables (contrary to Theorem 2.1 and Lemma 4.1 of ED). In Section |5j we prove a variant of Sklar's theorem [ 35 1 which factorizes a given su- permodular function on W£ as the composition of a Lipschitz continuous supermodular function on M.™ with m monotone functions on R + , and a cutoff lemma that replaces a given supermodular function by a bounded supermodular function. Section|6]is dedicated to the two-point versions of Theorems [2 and |2] Here, we review the proofs of the two- point rearrangement inequalities of Lorentz [ 11 1, Brock [ 14 1, and Draghici [ 15 | and find their equality cases. The main theorems are proved in Section 0by combining the results from Sections |5] and |6] Adapting Beckner's argument from [32], we note that the inequal- ities in Eq. (I1.3t and Eq. (I1.4t are strict unless u\ , . . . , u m produce equality in all of the corresponding two-point inequalities, and then apply the results from Section |6] In the final Section[8l we briefly discuss extensions for the Brascamp-Lieb-Luttinger and related inequalities. 5 Monotone functions In this section, we provide two technical results about functions with higher-order mono- tonicity properties. We begin with an auxiliary lemma for functions of a single variable. Lemma 5.1 (Monotone change of variable.) Let <fi be a nondecreasing real-valued function defined on an interval I. Then, for every function f on I satisfying \f(z)-f(y)\<C(<Kz)-<Ky)) (5.1) for all points y < z £ I with some constant C, there exists a Lipschitz continuous function f : R — * [inf /, sup /] such that f = f o <p. Furthermore, if f is nondecreasing, then f is nondecreasing. PROOF. If t = (f>(y) we set f(t) := f(y). For s < t with s = <f>(y), t = <j>{z), Eq. (BH implies that - f(s)\ = \f(z) - f(y)\ < C(^{z) - 4>{y)) = C(t - s) . (5.2) 7 Since / is uniformly continuous on the image of <p, it has a unique continuous extension to the closure of the image. The complement consists of a countable number of open disjoint bounded intervals, each representing a jump of (p, and possibly one or two unbounded intervals. On each of the bounded intervals, we interpolate / linearly between the values that have already been assigned at the endpoints. If (f> is bounded either above or below, we extrapolate / to t > sup <fi and t < inf 4> by constants. By construction, f = f ° 4> and f(M) = [inf /, sup/]. The continuous extension and the linear interpolation preserve the modulus of continuity of /, and hence, by Eq. (I5.2t . \Rt)-f(s)\<C\t-s\ (5.3) for all s, t G M.. If / is nondecreasing, then / is nondecreasing on the image of <p by defini- tion, and on the complement by continuous extension and linear interpolation. ■ Lemma l5~Tl is related to the elementary fact that a continuous random variable can be made uniform by a monotone change of variables. More generally, if <fi is nondecreasing and right continuous, and its generalized inverse is defined by rp(t) = inf{y : <j>(y) > t}, then the cumulative distribution functions of two random variables that are related by Y = ip(Y) satisfy F(y) = P(Y <y) = P(Y < 0(y)) = F^y)) , i.e., F = F o (p. Choosing 4> = F results in a uniform distribution for Y. The corresponding result for m > 2 random variables is known as Sklar's theo- rem [35 1 . The theorem asserts that a collection of random variables Y\, . . . , Y m with a given joint distribution function F can be replaced by random variables Y\ , . . . , Y m whose marginals % are uniformly distributed on [0, 1], and whose joint distribution function jP is continuous. The next lemma contains Sklar's theorem for supermodular functions. Since the lemma follows from the arguments outlined in [36| rather than from the statement of the theorem, we include its proof for the convenience of the reader. We first introduce some notation. Let F be a real- valued function on the closed positive cone M™. For i = 1, . . . , m and h > 0, consider the finite difference operators A l F(y;/ l ) := F(y + he t ) - F(y) . The operators commute, and higher order difference operators are defined recursively by A h _ ie F(y;hi,. . . ,h e ) := A il ...^_ 1 A k F((y; hi); h u ■ ■ ■ M-i) ■ If F is £ times continuously differentiable, then &i 1 ...i t F(y;hi,...,hi) = •••/ d h . . . d ie F[ y + V] ) dt x . . . dte . Jo Jo v i=1 J A function F is nondecreasing in each variable if Aji 7 > for i = 1, . . . , m; it is super- modular, if AijF > for all i / j. The joint distribution function of m random variables satisfies A^ . . . Ai e F > for any choice of distinct indices i\, . . . , 8 Lemma 5.2 (Sklar's theorem.) Assume that F is bounded, nondecreasing in each variable, and supermodular on R™. Then there exist bounded nondecreasing functions 01, ... , (p m on with 4>i(0) = and a Lipschitz continuous function F on such that F(yi, ■■■;y m ) = F{fa(y{), . . .,<t> m {Vm)) ■ Furthermore, F is bounded, nondecreasing in each variable, and supermodular. If in addition, A^^^F > on MJ? xl^. for some distinct indices i±, . . . , %i, then A^.^F > 0. Proof. Set (y) = lim iF(yi,...,y m ) - F(y 1} . . . ,y m ) yi=y Vi=0 These functions are nonnegative and bounded by sup F — inf F. Since F is nondecreasing in each variable, they are nonnegative, and since F is supermodular, they are nondecreas- ing and satisfy F(y + he i )-F(y)<Myi + h)-Myi) (5-4) for all y = (y h . . . , y m ) G and all h > 0. We construct F by changing one variable at a time. For the first variable, we write y = (y, y) where y G M + and y G By Eq. (Bil l, for each y G IR+ -1 , the function f(y) = F(y,y) satisfies Eq. (I5.U with C = 1 and (p = fa- By Lemma |5~T1 there exists a function F\ satisfying F( y ,y) = F 1 (My),y) for all (y, y) G M.™. Furthermore, Fi is Lipschitz continuous in the first variable, iFx^-Fx^y)! < \t-s\ . We claim that F\ satisfies Eq. (15 .4t for all j : > 1 with the same function fa as F. To see this, note that for each h > and every y, f(y) = A j F(y,y;h) satisfies the assumptions of Lemma l5~Tl with C = 2 and <p = fa. A moment's considera- tion shows that f(t) = A j F 1 (t,y;h) and the claim follows since sup / = sup / < fa (yj + h) — <pj (yj ) by Lemma |5~T1 We next verify that Fx has the same monotonicity properties as F. Suppose that &-ii...i t F > for some set of £ > 1 distinct indices £]_,... , t£. If 1 . . . we apply Lemma l5~Tl to /(y) = Ai 1 ,„i e F(y, y; /ii, . . . , which satisfies Eq. (I5.lt with C = 2* and </> = fa for all y G R m ~ l and all h u ...,h t > 0. It follows that f{t) = Aj 1 ...i f i ; i(t, y; foi, . . . , hi) > 0. On the other hand, if z'i = 1, we apply Lemma l5~T1 to /(y) = Ai 2r „ ; i t F(y,y; fi2, ■ ■ ■ , fog). Since /(y) is nondecreasing by as- sumption, fit) = Ai 2i ... ) ^i ? i(t, y; /12, . . . , /i^) is again nondecreasing, and we conclude that A^ . ^Fi > also in this case. 9 Iterating the change of variables for i = 2, . . . , m gives functions Fj satisfying Fj-i^i, . . .,U-i,yi, . . .,y m ) = Fi(ti, . . . (/)i(yi), yi+i, . . . y m ) , as well as 0<A j F i (t 1 ,...,t i ,y i+1 ,...y m ;h)<S [ ^ + fc) _ ^ J .f l . (5.5) Finally, we set F = F m . It follows from Eq. (I5.5t that F satisfies the Lipschitz condition |F(z) — F(y)| < \ z i ~ Vi\ ^ \An |z — y| for all y, z G ■ The distribution function of a Borel measure on can be conveniently approximated from below by restricting the measure to a large cube [0, L) m . The next lemma constructs the corresponding approximation for functions with weaker monotonicity properties. Lemma 5.3 (Cutoff.) Given a real-valued function F in WV, set F L (yi, ...,y m ):= F(mm{yi,L}, . . . ,min{y m ,L}) . If F is nondecreasing in each variable, then F L < F. If Ai 1 . ^ e F > on W± X M.+ for some distinct indices i\, . . . , then Aj lj ... ) j £ i ?lZ ' > 0. In particular, if F is supermodular, so is F L . If F has the property that A^.^.F > on R'P x W+for every set of distinct indices ix,...,ig, then F — F L also has this property. PROOF. As in the proof of Lemma l5~2l we modify the variables one at a time. The func- tion F 1,L (y, y) := F(min{y, L}, y) has the same monotonicity properties as F because min{y, L} is nondecreasing in y. If Aj 1 ...j { i ?1 > for all collections of distinct indices i\, . . . , i#, we write F(y, y) - F l ' L {y, y) = & x F{y, y; [y - L]+) , and it follows that A^ . j^F — F l,L ) > whenever 1 ^ . . . , ig}. For i\ = 1, we write A X {F(y, y; h) - F 1,L (y, y; h) = A 1 F(max{y, L},y; [h - [L - y]+} + ) , and conclude that Aj 1 ...j ( ,(F — F 1, ) > also in this case. Repeating the construction for the variables y%, . . . , y m gives the claims. ■ 6 Two-point rearrangements Let X be R n , S n , or H n . A reflection on X is an isometry characterized by the properties that (i) a 2 x = x for all x G X; (hJ the fixed point set Hq of a separates M into two half-spaces H + and that are interchanged by a; and (Hi) d(x, x') < d(x, ax') for all x, x' G We call and the positive and negative half-spaces associated with a. By convention, we always choose H + to contain the distinguished point x* of X in 10 its closure. The two-point rearrangement, or polarization of a real- valued function u with respect to a reflection a is defined by or \ _ \ max{«(x),u((Tx)} , x G H + U [ mm{u(x), u(ax)} , x G ff_ . This definition makes sense, and the two-point versions of Eqs. (I1.3t and (ll.4t hold for any space with a reflection symmetry. On X = R n , § n , and H n , any pair of points is connected by a unique reflection. The space of reflections forms an n-dimensional submanifold of the n{n + l)/2-dimensional space of isometries, and thus has a natural uniform metric. If u is measurable, both the composition u o a and the rearrangement u a depend continuously on a in the sense that <7fc — > a implies that u o o~k — > u o a and u°" fc — > u a in measure. Two-point rearrangements are particularly well-suited for identifying symmetric de- creasing functions, because u = u* u = u a for all a . (6.1) Functions that are radially decreasing about some point are characterized by u = u* o t for some translation r <^=^ for all a, either u = u a or u = u a o a (6.2) (see Lemma 2.8 of 11291). Integral inequalities for two-point rearrangements typically reduce to elementary com- binatorial inequalities for the integrands. The following lemma supplies the elementary inequality for the Hardy-Littlewood and Riesz functionals. Lemma 6.1 (Lorentz two-point inequality.) A real-valued function F on M.™ is super- modular, if and only if for every pair of points z, w G M.™. F(z 1 ,...,z m )+F(w 1 ,...,w m ) < F(Taax{z 1 ,Wx},...,max{z m ,w m }) +F(mm{zi,w 1 }, . . . , min{z m , w m }) . If AijF > Ofor some i ^ j then Eq. \6.3\ is strict unless [z% — Wi)(zj — Wj) > 0. PROOF. Given z, w G M™, define y, h G by yi = min{zj, Wi} and hi = \zi — Wi\ for i = l,...m. If / C {1, . . . , m}, we use the notation h/ = X^ig/ ^i e «- Subtracting the left hand side of Eq. (I6.3t from the right hand side results in the equivalent statement A/jF(y; h/, hj) := F(y + hjuj) - F(y + h/) - F(y + hj) + F(y) > , (6.4) where I = {i : z\ < Wi}, and J = {i : z-i > Wi}. If either / or J is empty, Eq. (I6.4t is trivially satisfied. If / and J each have exactly one element, Eq. (I6.4t is equivalent to Eq. dl.5t . If one of the sets, say /, has several elements, then decomposing it into disjoint subsets as I = I' U I" gives A/jF(y; h/, hj) = A JV F(y + h 7 », h 7 ,, hj) + A / » J F(y, hj,, hj) , 11 and Eq. (I6.4t follows by recursion. The same recursion implies that if \jF > and Wi and Zj — Wj have opposite signs, then the inequality in Eq. ( I6.4t is strict whenever I contains i, J contains j, and hi, hj > 0. ■ Brock proved that the left hand side of Eq. (I1.3t increases under two-point rearrange- ment FPU : Lemma 6.2 (Hardy-Littlewood two-point inequality.) Let F be a supermodular Borel measurable function on W?, and let u±, . . . , u m be nonnegative measurable functions on X satisfying the integrability condition in Eq. ( I2.il) . Then, for any reflection a on X, F(ui(x), . . . ,u m (x)) dx < / F(u1(x), . . . ,v% n (x)) dx . (6.5) Jx Assume furthermore that A-ijF > on M.™ x (0, oo) 2 for some i ^ j. If Eq. ( 16. 51 ) holds with equality and the integrals are finite, then [ui{x] — Ui(ax)) (uj(x) — Uj(ax)) > a.e. . In particular, ifm = u* is strictly radially decreasing and cr(x*) ^ x*, then Uj = u°. PROOF. The inequality XTH : The left hand side of Eq. d6-5l > can be written as an integral over the positive half-space, T(ux, . . . ,u m ) := / F(u 1 (x),...u m (x)) + F(ux(o-x),...u m (ax)) dx . By Lemma |6~T1 with Zi = Ui(x) and Wi = Ui(ax), the integrand satisfies F(m(x), . . . u m (x)) + F(ui(ax), . . . u m (ax)) < F(uf(x), . . . u° m {x)) + F(ul(ax), . . . <H) for all x € H + . Integrating over H + yields Eq. (16.5b . Equality statement: Assume that I(u\, u m ) = Eq. ( 16.6b must hold with equality almost everywhere on H + . If A^F > on W? x (0,oo) 2 , then Lemma l6~Tl implies that Ui(x) — Ui(ox) and Uj(x) — Uj{ax) cannot have opposite signs except on a set of zero measure. If moreover U{ = u* is strictly radially decreasing and ax* ^ x* , then u,i(x) > Ui(ax) for a.e. x £ H + , and Lemma l6~Tl implies that Uj(x) > Uj(ax) for a.e. x £ H + . ■ Brock completed the proof of Eq. dl.3t by approximating the symmetric decreasing rear- rangement with a sequence of two-point rearrangements a la Baernstein-Taylor Q. We sketch his argument in the simplest case where F is a continuous supermodular function that vanishes on the boundary of the positive cone W?, and u%,..., u m are bounded and compactly supported. 12 By Theorem 6.1 of [34] there exists a sequence of reflections {crfc}fc>i such that u <n,—>°k _^ u * i n measure (k — ► oo) (6.7) for i = 1, . . . ,m. By Lemma l6~2l the functional increases monotonically along such a sequence. If B is a ball centered at x* that contains the supports of m, . . . , u m , then the rearranged functions u^ 1 ''"' <7/s are also supported on B, and dominated convergence yields i^.^g^iK 1 (fc-oo). (6.8) The corresponding results for Eq. dl.4t are due to Draghici lfT5l . The two-point in- equality is not an immediate consequence of Lemma l6~Tl but requires an additional com- binatorial argument. This argument was used previously by Morpurgo 1 30 1 , and a simpler version appears in [29|. Lemma 6.3 (Riesz two-point inequality.) Assume that F is a supermodular Borel mea- surable function on M™. For each pair of indices 1 < i < j < m, let Kij be a nonin- creasing function on R + , and let u\, . . . ,u m be nonnegative measurable functions on X satisfying the integrability condition in Eq. A2.3i . Then, for any reflection a, ■■ F(m(xi), . . . ,u m (x m ))T\ Kij(d(xi,Xj)) dxi . . .dx m Jx Jx T<4 I f TT (6 - 9) Jx Jx f Kj Assume additionally that that Kij(t) > Ofor all i < j and all t < diamX. Let Tq be the graph on { 1 , . . . , m} with an edge between i and j whenever K^ is strictly decreasing. If Ay F > Ofor some i ^ j lying in the same connected component of To, and that Ui and Uj are not symmetric under a. If the integrals in Eq. \6.9\ have the same finite value, then either Uj = u" and Uj = u", or m = u° o a and Uj = uj o a. PROOF. The inequality KT5V : The left hand side of Eq. < I6.9I > can be written as an m-fold integral over the positive half-space T(ui,...,u m ) :-- / .../ Yl <F(u 1 (a £l x 1 ),...,u m (a £m x m )) x Jh+ Jh+ e . 6{0)1 } )i=1) ... jm I x Y\ K ij (d(a £l Xi,a ej Xj)) > dx\ . . . dx m . (6.10) Fix xi, . . . , x m £ H + . For each pair of indices i < j, set = K^ (d(xi,axj)) and bij = (d(xi,Xj)) — (d(xi,crxj)) , so that Kij{d(o- £t Xi,cj e ixj)) = + bijl £i=e . . 13 The product term in Eq. (I6.10t expands to 11 K ij (d(a £ ^x l , a £ i Xj )) = J2( II ( II h: ' X ) =: Crl^^ieE , i<j r ijeE where T runs over all proper graphs on the vertex set V = {1, . . . , m}, and 2? is the set of edges of T. Inserting the expansion into Eq. (I6.10t and exchanging the order of summation shows that each graph contributes a nonnegative term C r F{M^ El xx),---,u m (a Em x m ))l £i=£dj i jeE (6.11) e t e{0,l},i&V to the integral in Eq. (I6.10t . If T is connected, then F(ni(<T £l 2;i), . . . ,u m (a £m x m ))l £i=£ ^ ij( zE e;6{0,l},«eV = F(u 1 (x 1 ), . . .,u m (x m )) + F(ux((TXi), . . .,u m (ax m )) < F«(si)),...,u^(x m )) + F«(<7Xi)),...,< l (<7s m )) (6.12) £,£{0,11,16^ where the second step follows from Lemma l6~Tl with Zi = Ui{xi) and Wi = Ui(axi). If T is not connected, choose a connected component V and let r" be its complement. Let E' , E" , V, and V" be the corresponding edge and vertex sets. The sum in Eq. (16.1 It can be decomposed as ^2 \ ^2 F(ui{a £l xi), ... ,u m (a £m x m )) l £l=£] , ijeE' > l £% =e 3 ,ijaE" ■ The key observation is that Eq. (I6.12t applies to the term in braces for fixed £j,i G V"; in other words, the contribution of T can only increase if u, is replaced by v!J L for all i 6 V . An induction over the connected components of F shows that ^ F(ui(a £l xi), . . . 1 U m (a £m X rn ))l £l=£ ^ ij( zE e l G{0,l},«6V e,e{0,l},i£V for any graph T = (E?, F). Adding the contributions of all graphs shows that the integrand in Eq. (I6.10t increases pointwise under two-point rearrangement, and Eq. (I6.9E follows. Equality statement: Let To be the graph defined in the statement of the lemma, and let Eq be its edge set. By assumption, Cvo = ( II Kij(d(xi,Xj)) - K ij {d(axi,x j ))^(^ Ky (d(a Xi , Xj))j >0 ij0nEg ij&Eo 14 for a.e. x\,..., x m E H + . If AijF > 0, then Lemma l6Jl implies that Eq. (I6.12t is strict unless (v,i(xi) — Ui{axij) (uj{xj) — Uj(axj)) > , a.e. Xi,Xj G H + . If Ui and uj are not symmetric under a, the product is not identically zero. Since Xi and Xj can vary independently, this means that Ui(x) — m{ax) and Uj(x) — Uj(ax) cannot change sign on H + . We conclude that equality in Eq. (I6.9t implies that either m = u? and Uj = Uj, or m = u° o a and Uj = u" o a. ■ Draghici also used Baernstein-Taylor approximation to obtain Eq. dl.4t from Eq. <6.9b . If F is bounded and continuous and ify is bounded for 1 < j < j < m, then for bounded functions u\ , . . . , u m that are supported in a common ball the inequality follows from Lemma lOl bv approximating the symmetric decreasing rearrangement with a sequence of two-point rearrangements, see Eq. <6.7b . Dominated convergence applies as in Eq. J6.8t . since the integrations extend only over the bounded set B m . 7 Proof of the main results PROOF OF Theorem Q The inequality for Borel integrands: Let F be a supermod- ular Borel function with F(0) = 0, and let and u\, . . . ,u m be nonnegative measurable functions that vanish at infinity, as in the statement of the theorem. Denote by l(ui, . . . ,u m ) := / F(ui(x),...,Um(x))dx Jx the left hand side of Eq. (11.3b - Replacing F(y) by F(y) — YliLi F{Vi e i) an d using that F{ui{-) Bi) and F(u*{-) e«) contribute equally to the two sides of Eq. dl.3t . we may as- sume F to be nondecreasing in each variable. Fix L > 0, and replace itj by the bounded function uf(x) :=min{it i (a;) ) L}l { | a .| <£ ,} for i = 1, . . . ,m. Then F(u^,...,ui) = F L (ui,...,ui) , (7.1) where F L is the function defined in Lemma 1531 By construction, F L is bounded, and by Lemma l531 it is nondecreasing and supermodular. By Lemma l5^2l there exist nondecreas- ing functions 4>i with 4>i(0) = and a continuous supermodular function F L on R!^ such that F L ( yi , . . . ,y m ) = F L (Myi), ■ ■ ■ ,4m(y m )) ■ (7.2) Since (pi is nondecreasing and vanishes at zero, u\ is compactly supported, and (uf)* < (u*) L pointwise by construction, we have o uf)* = fa o {ufY < (Pi o (u*) L (7.3) 15 for i = 1, . . . , m. By Theorem 1 of [ 14 1), F L (<t> 1 o U L(x),...,(f> m ou% l (x))dx < [ F L ((0 1 ou^)*(x),...,(0 m ou^)*(x))dx . With Eqs. (fTlVlfOl. this becomes J(4,...,n^)<T(K) i ,...,(n^) i ). Since nf(x) = Ui{x) for L > max{-Uj(x), |x|}, we see that F(uf(x), . . . ,u^(x)) con- verges pointwise to F(u\(x), . . . ,u m (x)), and Eq. (I1.3t follows by monotone conver- gence. Equality statement: Combining Eq. (I6.5t with Eq. dl.3t and using that is equimea- surable with Uj, we see that U m ) < I(Ui, U a m ) < I(u*, U* m ) . Hence equality in Eq. dl.3t implies equality in Eq. (I6.5t for every choice of the reflection a. Given two points x, x' in X, choose a such that a{x) = x' . If A^F > for some i 7^ j, then Ui(x) — ui(x') and Uj(x) — Uj(x') cannot have opposite signs by Lemma l6!2l If Ui = u* is strictly radially decreasing, then it follows that uj = Uj for every reflection a that does not fix x*. By Eq. (I6.U . Uj = u* as claimed. ■ PROOF OF THEOREM EJ The inequality for Borel integrands: The proof of Eq. ( 11.41) proceeds along the same lines as the proof of Eq. dl.3t . Let l(ui, . . . ,u m ) := / ••• / F(u 1 (xi),...,u m (x m ))T\K ij (d(x i ,x j ))dxi...dx m be the left hand side of Eq. fll.4t . As before, we may assume that F is nondecreasing in each variable. We replace F with F L , Ui with fa o uf, Kij with = min{Kij, L}, and set 2 L (ui, . . . ,u m ) := / ••• / F L (ui(x 1 ), . . . ,u m (x m )) T\ K ij{d(xi,Xj)) dxi . . .dx m . Jx Jx Applying Theorem 2.2 of lfT31 . we obtain with the help of Eqs. <I7. H - d7.3b l L (u^...,ui)<l L ((ul) L ,...,(u* m ) L ) . Eq. (I1.4t follows by taking L —> oo and using monotone convergence. Equality statement: Consider the set Si of all reflections a of X that fix Uj. If is non-constant, then Si is a closed proper subset of the space of all reflections on X. This subset is nowhere dense, since any open set of reflections generates the entire isometry group of X. If Eq. (I1.4t holds with equality, then the two-point rearrangement inequality in Eq. (I6.9t holds with equality for every reflection a. For a Si, Lemma lOl implies that either Uj = Uj or Uj = u" o a. Since Si is nowhere dense, it follows from the continuous dependence of u a on a that uj agrees with either u a - or nj o a also for a G Si. By Eq. J6.2t . there exists a translation r such that uj = u* o r. Lemma IQI implies furthermore that Ui agrees with uf when uj = u a -, and with u" o a when Uj = u° o a. We conclude that u,' = u* or. ■ 16 8 Concluding remarks In the proof of Eq. il.4l and its two-point version in Eq. (I6.9I) . the kernels Ky played a very different role from the functions u\, . . . , u rn that enter into the integrand. However, the Riesz functional on the left hand side of Eq. dl.2t depends equally on u, v, and w. We will use the connection of Riesz' inequality with the Brunn-Minkowski inequality to construct examples where the two-point rearrangement fails for Eq. <1.2b . The Brunn-Minkowski inequality says that the measures of two subsets A, B C W 1 are related to the measure of their Minkowski sum A + B = {a + b : a G A,b G B} by Recognizing the two sides of the inequality as proportional to the radii of the balls A* + B* and (A + B)*, we rewrite it as the rearrangement inequality Eq. (18.11) follows rather directly from Riesz' inequality in Eq. J1.2i . because the sup- port of the convolution of two nonnegative functions is essentially the Minkowski sum of their supports. Conversely, the Brunn-Minkowski inequality enters into the proof of the Brascamp-Lieb-Luttinger inequality [ 10 1, of which Eqs. (11.21) and HAt are special cases. Equality in the Brunn-Minkowski inequality implies that A and B differ only by sets of measure zero from two independently scaled and translated copies of a convex body 1371 . Let A = B be an ellipsoid in M. n with n > 1 that is centered at a point c ^ 0, so that Eq. (I8.lt holds with equality. If a is the reflection at a hyperplane through c that is not a hyperplane of symmetry for A and B, then A a and B a are non-convex, and therefore Choosing u, v, and w as the characteristic functions of A, A + B, and B provides an example where the Riesz functional strictly decreases under two-point rearrangement. For an example of this phenomenon in one dimension, consider the symmetric decreasing functions \{Af/ n + \{B) l / n < X(A + B) \(A* +B*) < \{A + B) . (8.1) X(A a + B a ) > X(A* + B*) = X(A + B) . u(x) = 1 x-2\<s i V{x) = U)(x) = l\ x -!\ <e , and let a be the reflection at x = 1. Then u a (x) = 1 X \<E ) V a {x) = W a {x) = l|a._i| <e and if < e < ^, Riesz' inequality fails for a, 17 While the two-point rearrangement is not useful for Eq. (I1.2I) . the layer-cake represen- tation of Crowe-Zweibel-Rosenbloom shows that F(u(x),v(x'),w(x — x')) dxdx' < / / F(u* (x),v* (x'),w* (x — x')) dxdy for any integrand that can be written as the joint distribution function of a Borel measure F{y 1 ,y 2 ,y 3 ) = Pf([0,Vl) x [0,y 2 ) x [0,y 3 )) . Such integrands are left continuous, vanish at the origin, and satisfy Aj lv .. ^ e F > for every choice of £ < 3 distinct indices. Lemma I5T21 allows to accommodate integrands in Eq. d8.2t that are only Borel measurable. The main condition is that A123-F > 0; the second-order monotonicity conditions can be replaced by integrability assumptions on the negative part F_ similar to Eq. (12.31 . To ensure that the functional is finite at least when u,v,w are bounded and compactly supported, F should vanish on the coordinate axes. For example, Eq. (I8.2t holds for UV1V F(u, v, w) = — — r - (uv + uw + vw) (1 + u)(l + v)(l + w) since A 123 F > 0, even though \jF < for all i / j. For Borel integrands satisfying A123-F > 0, equality in Eq. d8.2t implies that every triple of level sets of u, v, w produces equality in Eq. dl.2b . These equality cases were described in [38 1. In particular, if two of the three functions u,v,w are known to have continuous distribution functions and the value of the functional is finite, then equality implies that u,v,w are equivalent to u*,v*,w* under the symmetries of the functional (see Theorem 2 of [38 1). By the same line of reasoning, the Brascamp-Lieb-Luttinger inequality ifTOl implies that l(ui, . . . ,u m ) := / ... / F[ui (V] aijXj) (V] a m jXj) ) dxi . . . dx k jR n JR n V j =1 . =1 7 increases under symmetric decreasing rearrangement, if Aj x ... > for all choices of distinct indices ii, . . . , with £ < m. Interesting examples are integrands of the form in Eq. (13.3b . where $ is completely monotone in the sense that all its distributional derivatives are nonnegative. If A^ . ^F > for all choices of i\, . . . then the last statement of Lemma l531 can be used to show that the extended Brascamp-Lieb-Luttinger inequality has the same equality cases as the original inequality. The characterization of these equality cases remains an open problem. References [1] G. E. Hardy, J. E. Littlewood, and G. Polya. Inequalities. First / second edition, Cambridge University Press, London and New York, 1934 / 1952. 18 [2] F. Riesz. Sur une inegalite integrale. /. London Math. Soc, 5:162-168, 1930. [3] A. Zygmund. On an integral inequality. /. London Math. Soc. 8:175-178, 1933. [4] M. G. Crandall and L. Tartar. Some relations between nonexpansive and order preserving mappings. Proc. Amer. Math. Soc, 78:385-390, 1980. [5] B. Kawohl. Rearrangements and convexity of level sets in PDE. Springer Lecture Notes in Mathematics 1 150, 1980. [6] S. L. Sobolev On a theorem of functional analysis. Mat. Sb. (N.S.) 4:471-497, 1938 lAMSTransl. (2), 34:39-68, 1963. [7] A. Baernstein, II and B. A. Taylor. Spherical rearrangements, subharmonic functions, and *-functions in n-space. Duke Math. J., 43:245-268, 1976. [8] W. Beckner. Sobolev inequalities, the Poisson semigroup and analysis on the sphere S n . Proc. N.A.S., 89:4816^1819, 1992. [9] C. Draghici. A general rearrangement inequality. Proc. Amer. Math. Soc, 133: 735- 743, 2005. [10] H. J. Brascamp, E. H. Lieb, and J. M. Luttinger. A general rearrangement inequality for multiple integrals. /. Funct. Anal., 17:227-237 ', 1974. [1 1] G. G. Lorentz. An inequality for rearrangements. Amer. Math. Monthly, 60:176-179, 1953. [12] J. A. Crowe, J. A. Zweibel, and P. C. Rosenbloom. Rearrangements of functions. J. Funct. Anal, 66:432^138, 1986. [13] F. Almgren and E. H. Lieb. Symmetric decreasing rearrangement is sometimes con- tinuous. Journal of the AMS, 2:683-773, 1989. [14] F. Brock. A general rearrangement inequality a la Hardy-Littlewood. /. Ineq. Appi, 5:309-320, 2000. [15] C. Draghici. Rearrangement inequalities with applications to ratios of heat kernels. Potential Analysis, 22:351-374 (2005). [16] H. Hajaiej and C. A. Stuart. Symmetrization inequalities for composition operators of Carafheodory type. Proc. London Math. Soc, 87:396^118, 2003. [17] H. Hajaiej. Cases of equality and strict inequality in the Extended Hardy Littlewood inequalities. Proc. Royal Soc. Edinburgh A 135:643-661, 2005. [18] H. Hajaiej. Extended Hardy-Littlewood inequalities and applications. Trans. Amer. Math. Soc, 357: 4885-4896 (2005). [19] H. Hajaiej and C. A. Stuart. Extensions of the Hardy-Littlewood inequalities for Schwarz symmetrization. Int. J. Math. Math. Sci., 59:3129-3150, 2004. [20] E. H. Lieb and M. Loss. Analysis. First / second edition, AMS Graduate Studies in Mathematics, 1996/2001. [21] R. Tahraoui. (a) Symmetrization inequalities, (b) Corrigendum. Nonlinear Anal, 27:933-955, 1996 and 39:535, 2000. 19 [22] J. Van Schaftingen and M. Willem, Set transformations, symmetrizations and isoperi- metric inequalities. In: Nonlinear analysis and applications to physical sciences, 135-152, Springer Italia, Milan, 2004. [23] G. Carlier. On a class of multidimensional optimal transportation problems. /. Con- vex Anal, 10:517-529, 2003. [24] C. A. Stuart. Guidance properties of nonlinear planar wave guides. Arch. Rat. Mech. Anal, 125:145-200, 1993. [25] H. Hajaiej and C. A. Stuart. Existence and non-existence of Schwarz symmetric ground states for eigenvalue problems. Matematica Pura ed Applicata, 186, 2005. [26] L. V. Ahlfors. Conformal Invariants: Topics in Geometric Function Theory. McGraw-Hill Series in Higher Mathematics, 1973. [27] J. M. Luttinger. Generalized isoperimetric inequalities, I, II, III. /. Math. Phys., 14:586-593, 1444-1447 and 1448-1450, 1973. [28] R. Friedberg and J. M. Luttinger. (a) Rearrangement inequalities for periodic func- tions, (b) A new rearrangement inequality for multiple integrals. Arch. Rat. Mech. Anal, 61:35^14 and 45-64, 1976. [29] A. Burchard and M. Schmuckenschlager. Comparison theorems for exit times. Geom. Fund. Anal, 11:651-692, 2001. [30] C. Morpurgo. Sharp inequalities for functional integrals and traces of conformally invariant operators. Duke Math. J., 114:477-553, 2002. [31] E. H. Lieb. Existence and uniqueness of the minimizing solution of Choquard's nonlinear equation. Studies in Applied Mathematics, 57:93-105, 1977. [32] W. Beckner. Sharp Sobolev inequalities on the sphere and the Moser-Trudinger in- equality. Ann. Math. (2), 138:213-242, 1993. [33] A. Baernstein, II. A unified approach to symmetrization. In: Partial differential equa- tions of elliptic type (Cortona, 1992), 47-91, Sympos. Math., XXXV, Cambridge Univ. Press, Cambridge, 1994. [34] F. Brock and A. Yu. Solynin. An approach to symmetrization via polarization. Trans. Amer. Math. Soc, 352:1759-1796, 2000. [35] A. Sklar. Fonctions de repartition a n dimensions et leurs marges. Inst. Statist. Univ. Paris, 8:229-231, 1959. [36] A. Sklar. Random variables, joint distribution functions, and copulas. Kybernetika (Prague), 9:449^160, 1973. [37] H. Hadwiger and D. Ohmann. Brunn-Minkowskischer Satz und Isoperimetrie. Math. Z, 66:1-8, 1956. [38] A. Burchard. Cases of equality in the Riesz rearrangement inequality. Ann. Math. (2), 143:499-527, 1996. 20