A discrete version and stability of Brunn Minkowski inequality

In the first part of the paper, we define an approximated Brunn-Minkowski inequality which generalizes the classical one for length spaces. Our new definition based only on distance properties allows us also to deal with discrete spaces. Then we show…

Authors: ** Michel Bonnefont **

A discrete version and stability of Brunn Minkowski inequality
A disrete v ersion and stabilit y of Brunn Mink o wski inequalit y Mi hel Bonnefon t Institut de mathématiques, Lab oratoire de Statistique et Probabilités, Univ ersité P aul Sabatier, 118 route de Narb onne, 31062 T oulouse, FRANCE No v em b er 16, 2018 Abstrat In the rst part of the pap er, w e dene an appro ximated Brunn- Mink o wski inequalit y whi h generalizes the lassial one for length spaes. Our new denition based only on distane prop erties allo ws us also to deal with disrete spaes. Then w e sho w the stabilit y of our new inequalit y under a on v ergene of metri measure spaes. This result giv es as a orol- lary the stabilit y of the lassial Brunn-Mink o wski inequalit y for geo desi spaes. The pro of of this stabilit y w as done for dieren t inequalities (ur- v ature dimension inequalit y , metri on tration prop ert y) but as far as w e kno w not for the Brunn-Mink o wski one. In the seond part of the pap er, w e sho w that ev ery metri measure spae satisfying lassial Brunn-Mink o wski inequalit y an b e appro ximated b y disrete spaes with some appro ximated Brunn-Mink o wski inequalities. 1 In tro dution Let us reall some fats ab out the Brunn-Mink o wski inequalit y . First the in- equalit y w as set in R n for on v ex b o dies b y Brunn and Mink o wski in 1887 (for more details ab out the inequalit y and its birth, one an refer to the great sur- v eys [1 , 5℄ and the referene therein). It an b e read as if K and L are on v ex b o dies (ompat on v ex sets with non empt y in terior) of R n and 0 < t < 1 then V n ((1 − t ) K + tL ) 1 /n ≥ (1 − t ) V n ( K ) 1 /n + tV n ( L ) 1 /n (1) where V n is the Leb esgue measure on R n and + the Mink o wski sum whi h is giv en b y A + B = { a + b , a ∈ A, b ∈ B } 1 for A and B t w o sets of R n . Equalit y holds if and only if K and L are equals up to translation and dilatation. Brunn-Mink o wski inequalit y is a v ery p o w erful inequalit y with a lot of ap- pliations. F or example it implies v ery qui kly the isop erimetri inequalit y for on v ex b o dies in R n whi h reads  V n ( K ) V n ( B )  1 /n ≤  s ( K ) s ( B )  1 / ( n − 1) (2) where K is a on v ex b o dy of R n and s the surfai measure, with equalit y if and only if K is a ball. The Brunn-Mink o wski inequalit y is not only true for on v ex b o dies but also for all ompat sets and ev en for all measurable sets of R n (with the little diult y that the Mink o wski sum of t w o mesurable sets is not neessary mea- surable). One w a y to pro v e it is to pro v e a funtional inequalit y kno wn as Prek opa-Leindler inequalit y whi h applied to  harateristi funtions of sets giv es the m ultipliativ e Brunn-Mink o wski inequalit y V n ((1 − t ) K + tL ) ≥ V n ( K ) 1 − t V n ( L ) t (3) where V n is the Leb esgue measure on R n , K and L t w o measurable sets of R n . By homogenit y of the v olume V n , it an b e sho wn that this a priori w eak inequalit y is in fat equiv alen t to the n -dimensional one (1). All this w as to sho w that Brunn-Mink o wski inequalit y has a v ery geometri meaning and it is natural to ask on whi h more general spaes than R n the inequalit y an b e extended. One rst answ er is w e an  hange the measure, for example a measure log- ona v e on R n satisfy m ultipliativ e Brunn Mink o wski. But to b e able to quit R n , w e ha v e to generalize the Mink o wski sum. This an b e done on length spaes b y using ideas of optimal transp ortation (refer to [3 ℄ for length spae, [7 ℄ for optimal transp oration, and for exemple [4℄ for this generalisation). F ollo wing an idea of this pap er, for t w o sets K and L of a metri spae X w e dene what w e are going to all the s -in termediate set b et w een K and L b y Z s ( K, L ) =  z ∈ X ; ∃ ( k , l ) ∈ K × L , d ( k , z ) = sd ( k , l ) d ( z , l ) = (1 − s ) d ( k , l )  (4) This set will pla y the role set of baryen ters of the Mink o wski sum. In fat the authors in [4 ℄ use it only for a Riemannian manifold but it mak es sense for all metri spaes ev en if it is in teresting only for length spae. In this on text w e will sa y a metri measure spae ( X, d, m ) satises the N -dimensionnal Brunn- Mink o wski inequalit y if m 1 / N ( Z s ( K, L )) ≥ (1 − s ) m 1 / N ( K ) + s m 1 / N ( L ) (5) for all 0 < s < 1 and K , L ompats of X . W e will refer in the sequel at (5) as the "lassial" N -dimensionnal Brunn-Mink o wski inequalit y . It is pro v en in 2 [4 ℄ that a Riemannian manifold M of dimension n whose Rii's urv ature is alw a ys non negativ e satises (5) with dimension N = n and with the anonial v olume of the Riemannian manifold as measure, i.e. v ol ( Z s ( K, L )) 1 /n ≥ (1 − s ) v ol ( K ) 1 /n + s vol ( L ) 1 /n (6) for all ompats K , L of M where v ol denotes the anonial v olume of the Riemannian manifold. In fat they obtain more preise results on funtionnal inequalities lik e Prek opa-Leindler and Borell-Brasamp-Lieb inequalities. Reen tly , there ha v e b een a lot of w orks on geometry of metri measure spaes. Lott-Villani and Sturm ha v e giv en indep enden tly a syn theti treatmen t of metri spaes ha ving Rii urv ature b ounded b elo w b y k (see [7, 9 , 10 ℄). All these w orks b egan b y the result of preompatness of Gromo v: the lass of Riemannian manifolds of dimension n and Rii urv ature b ounded b elo w b y some onstan t k is preompat for a Gromo v-Hausdor metri. So the notion they dev elop for metri spaes has to generalize the one for Riemannian mani- folds and has to b e stable b y Gromo v-Hausdor on v ergene. Their denition is ab out on v exit y prop erties of relativ e en trop y on the W asserstein spae of probabilit y and is link ed with optimal transp ortation. Sturm in this on text denes a Brunn Mink o wski inequalit y with urv ature k (see [10 ℄). The meaning of this inequalit y ma y b e not totally satisfatory . Indeed the inequalit y is dep ending on parameter Θ whi h equals inf k ∈ K,l ∈ L d ( k , l ) or sup k ∈ K,l ∈ L d ( k , l ) whether the urv ature is p ositiv e (or n ull) or negativ e. It orresp onds to the minimal or maximal length of geo desis b et w een the t w o ompats K and L . Ho w ev er this is a diret impliation from its dimension- urv ature ondition C D ( k , N ) and this is this inequalit y that giv es all the geo- metri onsequenes of their theory lik e for example a Bishop-Gromo v theorem on the gro wth of balls. There is another w eak onept of urv ature whi h is kno wn as metri on- tration prop ert y (see [8 , 10 , 6℄) and whi h is implied b y this Brunn-Mink o wski inequalit y at least in the ase of urv ature 0 and the m ⊗ m a.s. uniqueness of geo desis b et w een t w o p oin ts of X . As far as I kno w stabilit y of Brunn-Mink o wski inequalit y w as not pro v en y et. This is the most in teresting result w e ha v e in the pap er (orollary 2.4 ). F or simpliit y w e will w ork only with the lassial Brunn-Mink o wski (i.e. with urv ature 0) and explains ho w to extend our results in the general ase, with urv ature k , in a remark. F or doing this w e in tro due an appro ximated Brunn mink o wski inequalit y sine w e need it during the pro of. This fat is in teresting in itself sine it allo ws us to deal with disrete spaes. In the seond part of the pap er w e sho w that ev ery metri measure spae satisfying lassial Brunn-Mink o wski inequalit y an b e appro ximated b y disrete spaes with some appro ximated Brunn-Mink o wski inequalities. T o a v oid some problems b et w een sets with zero measure w e will w ork only with metri spaes ( X, d, m ) where ( X, d ) is P olish and m a Borel measure on ( X, d ) with full supp ort, i.e. that  harges ev ery ball of X . 3 2 Stabilit y of Brunn-Mink o wski inequalit y Denition 2.1. Given h ≥ 0 and N ∈ N , N ≥ 1 , we say that a metri me a- sur e sp a e ( X, d, µ ) satises the h Brunn-Minkowski ine quality of dimension N denote d by B M ( N , h ) if ∀ C 0 , C 1 ⊂ X  omp ats, ∀ s ∈ [0 , 1] , we have: µ 1 / N ( C s h ) ≥ (1 − s ) µ 1 / N ( C 0 ) + s µ 1 / N ( C 1 ) (7) wher e C h s =  x ∈ X/ ∃ ( x 0 , x 1 ) ∈ C 0 × C 1 / | d ( x 0 , x ) − sd ( x 0 , x 1 ) | ≤ h | d ( x, x 1 ) − (1 − s ) d ( x 0 , x 1 ) | ≤ h  (8) W e all the set C h s the set of h (-appro ximated) s -in termediate p oin ts b et w een C 0 and C 1 . One an note that if X is a geo desi spae and h = 0 , it giv es ba k the lassial Brunn-Mink o wski inequalit y for geo desi spaes. W e shall often note B M ( N ) instead of B M ( N , 0 ) . Another remark to b e done is that this denition an b e used for disrete spaes. One an also note that if X satisfy B M ( N , h ) it will also satisfy B M ( N , h ′ ) for all h ′ ≥ h . In these notes w e use the follo wing distane D b et w een abstrat metri measure spaes. W e refer to [9 ℄ for its prop erties. Denition 2.2. L et ( M , d, m ) and ( M ′ , d ′ , m ′ ) b e two metri me asur e sp a es, their distan e D is given by D (( M , d, m ) , ( M ′ , d ′ , m ′ )) = inf ˆ d,q  Z M × M ′ ˆ d 2 ( x, x ′ ) dq ( x, y )  1 / 2 wher e ˆ d is a pseudo metri on M ⊔ M ′ whih  oinides with d on M and with d ′ on M ′ and q a  oupling of the me asur es m and m ′ . Theorem 2.3. L et ( X n , d n , m n ) b e a se quen e of  omp at metri me asur e sp a es whih  onver ges with r esp e t to the distan e D to another  omp at met- ri me asur e sp a e ( X, d, m ) . If ( X n , d n , m n ) satises B M ( N , h n ) with h n → h when n go es to innity, then ( X, d, m ) satises B M ( N , h ) . In partiular for ompat geo desi spaes it implies diretly the stabilit y of the lassial Brunn-Mink o wski inequalit y with resp et to the D -on v ergene: Corollary 2.4. L et ( X n , d n , m n ) b e a se quen e of  omp at ge o desi sp a es whih  onver ges with r esp e t to the distan e D to another  omp at metri me a- sur e sp a e ( X, d, m ) , then X is also a ge o desi sp a e. If ( X n , d n , m n ) satises B M ( N ) then ( X, d, m ) satises also B M ( N ) . W e will mak e the pro of of theorem 2.3 only for ompat sets of stritly p ositiv e measure. The remarks after the pro of will giv e the inequalit y for all 4 mesurable sets. The idea of the pro of is quite simple. W e  ho ose t w o ompats of the limit set X . Then w e  ho ose a go o d oupling of X n and X and w e onstrut t w o om- pats of X n b y dilating these ompats with resp et to the pseudo-distane of the oupling and taking the restrition of this t w o sets with X n . The fat whi h mak es things w ork is that the op eration w e did do esn't lose t w o m u h measure. So, w e an dene a s -in termediate set in X n and apply Brunn-Mink o wski in- equalit y in X n . By the same onstrution as b efore, w e onstrut a set in the limit set X from the s -in termediate set in X n without lo osing a lot of measure. T o onlude w e ha v e to study the link b et w een this set and set of appro ximate s -in termediate p oin ts b et w een initial ompats. Pro of of Theorem 2.3 Let C 0 , C 1 t w o ompats of X of stritly p ositiv e measure. Let s ∈ [0 , 1] . Cho ose n so that D ( X n , X ) ≤ 1 2 n . By denition of D , there exists ˆ d a pseudo-metri on X n ⊔ X and q a oupling of m n and m so that  Z X n × X ˆ d 2 ( x, y ) dq ( x, y )  1 / 2 ≤ δ n = 1 n F or ε n > 0 dene C ε n n,i = { x ∈ X n / ˆ d ( x, C i ) ≤ ε n } for i = 1 , 2 , these are ompats of X n . They are indeed not empt y for n large enough and ε n w ell  hosen, sine b eing of stritly p ositiv e measure as w e will see it. W e ha v e m ( C 0 ) = q ( X n × C 0 ) = q ( C ε n n, 0 × C 0 ) + q ( { X n \ C ε n n, 0 } × C 0 ) But if ( x, y ) ∈ { X n \ C ε n n, 0 } × C 0 , then ˆ d ( x, y ) ≥ ε n , so q ( { X n \ C ε n n, 0 } × C 0 ) ≤ Z { X n \ C ε n n, 0 }× C 0 ˆ d 2 ( x, y ) ε n 2 dq ( x, y ) ≤ δ 2 n ε n 2 whi h equals 1 n for δ n = 1 n and ε n = 1 √ n . On the other hand, w e ha v e: m n ( C ε n n, 0 ) = q ( C ε n n, 0 × X ) ≥ q ( C ε n n, 0 × C 0 ) Consequen tly , m n ( C 1 √ n n, 0 ) ≥ m ( C 0 ) − 1 n (9) and iden tially m n ( C 1 √ n n, 1 ) ≥ m ( C 1 ) − 1 n . (10) 5 No w onsider the set C ε n ,h n n,s ⊂ X n dened as in the denition (2.1 ) b y C ε n ,h n n,s =  x ∈ X n / ∃ ( x n, 0 , x n, 1 ) ∈ C ε n n, 0 × C ε n n, 1 / | d ( x n, 0 , x ) − sd ( x n, 0 , x n, 1 ) | ≤ h n | d ( x, x n, 1 ) − (1 − s ) d ( x n, 0 , x n, 1 ) | ≤ h n  This is the set of all the h n s -in termediate p oin ts b et w een C ε n n, 0 and C ε n n, 1 . Sine X n satises B M ( N , h n ) , m 1 n n ( C ε n ,h n n,s ) ≥ (1 − s ) m 1 / N n ( C ε n n, 0 ) + s m 1 / N n ( C ε n n, 1 ) (11) W e an no w dene C ε n ,h n s ⊂ X b y C ε n ,h n s = { y ∈ X , ∃ x ∈ C ε n ,h n n,s ˆ d ( x, y ) ≤ ε n } Similary to (9) w e ha v e m ( C 1 √ n ,h n s ) ≥ m n ( C ε n n,s ) − 1 n (12) No w sine ( x − 1 n ) 1 / N + ≥ x 1 / N − ( 1 n ) 1 / N for all x ≥ 0 , om bining the inequalities (9), (10 ), (12 ) and (11) giv e us, for ε n = 1 √ n , m 1 / N ( C ε n ,h n s ) ≥ m 1 / N n ( C ε n ,h n n,s ) − ( 1 n ) 1 / N ≥ (1 − s ) m 1 / N n ( C ε n n, 0 ) + s m 1 / N n ( C ε n n, 1 ) − ( 1 n ) 1 / N ≥ (1 − s ) m 1 / N ( C 0 ) + s m 1 / N ( C 1 ) − 2( 1 n ) 1 / N C ε n ,h n s is inluded in the set K h n +4 ε n s of all the h n + 4 ε n s -in termediate p oin ts b et w een C 0 and C 1 . Indeed, let y ∈ C ε n ,h n s , b y denition of this set, there exists x ∈ C ε n ,h n n,s so that ˆ d ( x, y ) ≤ ε n . By denition of C ε n ,h n n,s , it follo ws that there exists ( x n, 0 , x n, 1 ) ∈ C ε n n, 0 × C ε n n, 1 satisfying | d n ( x, x n, 0 ) − s d n ( x n, 0 , x n, 1 ) | ≤ h n | d n ( x, x n, 1 ) − (1 − s ) d n ( x n, 0 , x n, 1 ) | ≤ h n . There exists, b y denition of C ε n n,i for i = 1 , 2 , ( y 0 , y 1 ) ∈ C 0 × C 1 with ˆ d ( x n, 0 , y 0 ) ≤ ε n and ˆ d ( x n, 1 , y 1 ) ≤ ε n . It follo ws: | ˆ d ( y , y 0 ) − s ˆ d ( y 0 , y 1 ) | ≤ | ˆ d ( y , y 0 ) − ˆ d ( x, x n, 0 ) | + | ˆ d ( x, x n, 0 ) − s ˆ d ( x n, 0 , x n, 1 ) | + s | ˆ d ( y 0 , y 1 ) − ˆ d ( x n, 0 , x n, 1 ) | ≤ h n + 4 ε n . and | ˆ d ( y , y 1 ) − (1 − s ) ˆ d ( y 0 , y 1 ) | ≤ h n + 4 ε n . 6 The sequene ( h n + ε n ) n is on v erging to h . W e an extrat a monotone sequene from it whi h will still b e denoted b y h n + ε n . There are t w o ases. The rst one is when the extrating subsequene is non-dereasing. Then w e ha v e K h n +4 ε n s ⊂ K h s . So, for all n , m 1 / N ( K h s ) ≥ m 1 / N ( K h n +4 ε n s ) ≥ (1 − s ) m 1 / N ( C 0 ) + s m 1 / N ( C 1 ) − 2( 1 n ) 1 / N . Letting n go es to innit y giv es the onlusion. The seond one, more in teresting, is when the extrated subsequene is non- inreasing. Then w e ha v e K h s = \ n K h n +4 ε n s . Indeed if y ∈ T n K h n +4 ε n s , for all n ∈ N , ∃ ( y n, 0 , y n, 1 ) ∈ C 0 × C 1 so that | d ( y , y n, 0 ) − s d ( y n, 0 , y n, 1 ) | ≤ h n + 4 ε n | d ( y , y n, 1 ) − (1 − s ) d ( y n, 0 , y n, 1 ) | ≤ h n + 4 ε n . By ompatness of C 0 and C 1 w e an extrat another subsequene so that y n, 0 → y 0 ∈ C 0 and y n, 1 → y 1 ∈ C 1 and w e ha v e | d ( y , y 0 ) − s d ( y 0 , y 1 ) | ≤ h | d ( y , y 1 ) − (1 − s ) d ( y 0 , y 1 ) | ≤ h . The other inlusion is immediate. This in tersetion is non-inreasing so m 1 / N ( K h s ) = lim n →∞ m 1 / N ( K h n +4 ε n s ) whi h giv es the onlusion m 1 / N ( K h s ) ≥ (1 − s ) m 1 / N ( C 0 ) + s m 1 / N ( C 1 ) . Remark 1. B M ( N ) is diretly implied b y the ondition C D ( O , N ) of Sturm or Lott and Villani for the ompat sets with a stritly p ositiv e measure (in fat for mesurable sets with stritly p ositiv e measure) (see [10 ℄). But if the mea- sure m is  harging all the balls of the spae and (if the spae is geo desi), then the fat of ha ving B M ( N ) for all the ompats subspae with stritly p ositiv e measure implies B M ( N ) for all ompat subspaes. Indeed if ( X, d, m ) saties B M ( N ) for all the ompat sets with a stritly p ositiv e measure and if the measure m is  harging all the balls, if C 0 , C 1 are om- pats with m ( C 0 ) = 0 and m ( C 1 ) > 0 (the ase m ( C 0 ) = m ( C 1 ) = 0 is trivial) and s ∈ [0 , 1] . Dene C ε 0 = { y ∈ X, ∃ x ∈ C 0 /d ( x, y ) ≤ ε } , m ( C ε 0 ) > 0 . Dene H ε s the set of all the s -in termediate p oin ts b et w een C ε 0 and C 1 , By Brunn-Mink o wski inequalit y w e ha v e: m 1 / N ( H ε s ) ≥ (1 − s ) m 1 / N ( C ε 0 ) + s m 1 / N ( C 1 ) ≥ s m 1 / N ( C 1 ) 7 H ε s is inluded in K 2 ε s the set of all 2 ε s -in termediate p oin ts b et w een C 0 and C 1 . As b efore T ε> 0 K 2 ε s is an non-inreasing in tersetion equal to K 0 s the set of all the exat s -in termediate p oin ts b et w een C 0 and C 1 . So m ( K 0 s ) = lim ε → 0 K 2 ε s whi h giv es the annoned result. Consequen tly , on a metri measure spae where the measure  harges all the balls, C D (0 , N ) implies B M ( N ) for all ompats whi h in turns implies M C P (0 , N ) 2. In P olish spaes, Borel measures are regular whi h p ermits to pass from ompat sets to measurable ones. More preisely , if a P olish spae satisfy B M ( N , h ) for all his ompat subsets, it also satises it for all his measur- able subsets. Therefore, if the spaes X n and X are only P olish (no more ompats), the sets C ε n n,i for i = 1 , 2 dened as ab o v e ma y b e no more ompats. Ho w ev er they will still b e measurable sine losed, so (11 ) will still sta y true in this more general on text. W e an, onsequen tly , drop the assumption of ompatness of X n and X in the theorem ( 2.3 ) and its orollarry (2.4 ). 3. W e an do the same for the Brunn-Mink o wski inequalit y with urv ature k b y using the denition giv en in [10 ℄. The only additional thing to do is to on trol the parameter Θ . But, with preeeding notations, w e ha v e | Θ( C 0 , C 1 ) − Θ( C ε n n, 0 , C ε n n, 1 ) | ≤ 2 ε n . 4. W e an pro v e also the same theorem for the m ultipliativ e Brunn-Mink o wski inequalit y (3). 3 Disretizations of metri spaes Let ( M , d, m ) b e a giv en P olish measure spae. F or h > 0 , let M h = { x i , i ≥ 1 } b e a oun table subspae of M with M = S i ≥ 1 B h ( x i ) . Cho ose A i ⊂ B h ( x i ) , x i ∈ A i m utually disjoin t and mesurable so that S i ≥ 1 A i = M . Con- sider the measure m h on M h giv en b y m h ( { x i } ) = m ( A i ) for i ≥ 1 . W e all ( M h , d, m h ) a disretization of ( M , d, m ) . It is pro v ed in [2 ℄ that if m ( M ) < ∞ then ( M h , d, m h ) D − → ( M , d, m ) . Theorem 3.1. If ( M , d, m ) satises B M ( N ) then ( M h , d, m h ) satises B M ( N , 4 h ) . The pro of is based on the t w o follo wing fats. Lemma 3.2. 1. If H ⊂ M h then m ( H h ) ≥ m h ( H ) (13) wher e H h = { x ∈ M , d ( x, H ) ≤ h } . 8 2. If A ⊂ M mesur able and A h = { x i ∈ M h , d ( x i , A ) ≤ h } then m h ( A h ) ≥ m ( A ) . (14) Pro of of lemma 3.2 First, let H ⊂ M h , w e ha v e m h ( H ) = X i/x i ∈ H m ( A i ) = m ( ⊔ i/x i ∈ H A i ) ≤ m ( H h ) sine ⊔ i/x i ∈ H A i ⊂ H h = { x ∈ M , d ( x, H ) ≤ h } . F or the seond p oin t, let A ⊂ M mesurable, dene A h as ab o v e, then m h ( A h ) = X i/x i ∈ A h m ( A i ) = m ( ⊔ i/x i ∈ A h A i ) ≥ m ( A ) sine ⊔ i/x i ∈ A h A i ⊃ A . Indeed if for some j , A j ∩ A 6 = ∅ then there exists a ∈ A with d ( x j , a ) ≤ h so x j ∈ A h . Pro of ot theorem 3.1 Let H 0 , H 1 b e t w o ompats of M h and s ∈ [0 , 1] . H 0 and H 1 onsist of a nite or oun table n um b er of p oin ts x j . Dene H h 0 , H h 1 ⊂ M b y H h i = { x ∈ M , ∃ x j ∈ H i /d ( x j , x ) ≤ h } for i = 1 , 2 . By the rst p oin t of the lemma, for i = 1 , 2 m ( H h i ) ≥ m h ( H i ) . (15) Let ( H h ) s ⊂ M b e the set of all the s -in termediate p oin ts b et w een H h 0 and H h 1 in the en tire spae M , i.e. ( H h ) s =  x ∈ M , ∃ ( x 0 , x 1 ) ∈ H h 0 × H h 1 /     d ( x, x 0 ) = s d ( x 0 , x 1 ) d ( x, x 1 ) = (1 − s ) d ( x 0 , x 1 )  B M ( N ) inequalit y on M giv es us m 1 / N (( H h ) s ) ≥ (1 − s ) m 1 / N ( H h 0 ) + s m 1 / N ( H h 1 ) . (16) As b efore b y triangular inequalit y , w e an see ( H h ) s is inlude in the set ˜ C 3 h s of 3 h s -in termediaire p oin ts in the whole spae M b et w een H 0 and H 1 . So the set ˜ H 4 h s ⊂ M h of 4 h s -in termediate p oin ts b et w een H 0 and H 1 in the disrete spae M h on tains the restrition at M h of the h dilated of ( H h ) s . By the seond p oin t of the lemma w e ha v e m h ( ˜ H 4 h s ) ≥ m (( H h ) s ) . (17) 9 Com bining inequalities (15), (16 ) and (17) ends the pro of of the theorem. Remark If ( M , d, m ) satises B M ( N , k ) then ( M h , d, m h ) satises B M ( N , k + 4 h ) . Referenes [1℄ F. Bar the , A utour de l'iné galité de Brunn-Minkowski . Ann. F a. Si. T oulouse Math. (6), ( 2003 ) v ol 12, 27178 [2℄ A.I. Bonioa t and K.T. Sturm , Mass tr ansp ortation and r ough urvatur e b ounds for disr ete sp a es . Preprin t [3℄ D. Bura go , Y. Bura go and S. Iv ano v , A  ourse in metri ge ometry . Graduate Studies in Mathematis 33. Amerian Mathematial So iet y , Pro videne, RI.( 2001 ) [4℄ D. Corder o-Era usquin , R. MCann and M. Shmukenshläger , A Riemannian interp olation ine quality à la Bor el l, Br as amp and Lieb . In v en t. Math. ( 2001 ) v ol 146, 219257, [5℄ R.J. Gardner , The Brunn-Minkowski ine quality . Bulletin of the Amerian Mathematial So iet y ( 2001 ) v ol 39, n ◦ 3, 355-405 [6℄ N. Juillet Ge ometri Ine qualities and Gener alise d R i i Bounds in Heisenb er g Gr oup , preprin t [7℄ J. Lott and C. Villani , R i i urvatur e for metri-me asur e sp a es via optimal tr ansp ort . Ann. of Math. (to app ear). [8℄ S.I. Oht a On the me asur e  ontr ation pr op erty of metri me asur e sp a es Commen t. Math. Helv. ( 2007 ) Commen t. Math. Helv. v ol 82, 805828 [9℄ K.T. Sturm , On the ge ometry of metri me asur e sp a es. I . A ta Math., in press. [10℄ K.T. Sturm , On the ge ometry of metri me asur e sp a es. II . A ta Math., in press. [11℄ C. Villani T opis in optimal tr ansp ortation . Graduate Studies in Mathematis 58. Amer- ian Mathematial So iet y ( 2003 ) 10

Original Paper

Loading high-quality paper...

Comments & Academic Discussion

Loading comments...

Leave a Comment