aa r X i v : . [ m a t h . DG ] M a r Lecture notes on differential calculus on
RCD spaces
Nicola Gigli ∗ March 21, 2017
Contents L -normed modules, cotangent module and differential . . . . . . . . . . . . . 61.2.1 L -normed modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.2.2 Cotangent module and differential . . . . . . . . . . . . . . . . . . . . 81.3 Duality and the tangent module . . . . . . . . . . . . . . . . . . . . . . . . . 111.3.1 The module dual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.3.2 The tangent module . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141.4 Link with the metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161.4.1 Pullback of a module . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161.4.2 Speed of a test plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181.5 Maps of bounded deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . 201.6 Infinitesimally Hilbertian spaces and Laplacian . . . . . . . . . . . . . . . . . 24 RCD spaces 27
RCD spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.2 Measure-valued Laplacian and test functions . . . . . . . . . . . . . . . . . . 282.3 The space W , ( X ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302.3.1 Tensor product of Hilbert modules . . . . . . . . . . . . . . . . . . . . 302.3.2 Definition of W , ( X ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 312.3.3 Existence of W , functions . . . . . . . . . . . . . . . . . . . . . . . . 322.3.4 Calculus rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.4 Covariant derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.4.1 Sobolev vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.4.2 Calculus rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372.4.3 Flow of vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382.5 Exterior derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.5.1 Exterior power of a Hilbert module . . . . . . . . . . . . . . . . . . . . 392.5.2 Sobolev differential forms and basic calculus rules . . . . . . . . . . . 402.5.3 de Rham cohomology and Hodge theorem . . . . . . . . . . . . . . . . 422.6 Ricci curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442.7 Some properties in the finite dimensional case . . . . . . . . . . . . . . . . . . 47 ∗ SISSA. email: [email protected] ntroduction These are extended notes of the course given by the author at RIMS, Kyoto, in October2016. The aim is to give a self-contained overview on the recently developed approach todifferential calculus on metric measure spaces, with most, but not all, the material comingfrom [25]. The effort is directed into giving as many ideas as possible, without losing too muchtime in technical details and utmost generality: for this reason many statements are givenunder some simplifying assumptions and proofs are sometimes only sketched.The notes are divided in two parts: in the first one we study the first-order structureof general metric measure spaces, then, building on top of this, in the second we study thesecond-order differential structure of spaces with (Riemannian) Ricci curvature bounded frombelow.For what concerns the first part, a crucial role is played by the concept of L -normed L ∞ -module, which provides a convenient abstraction of the notion of ‘space of L sections ofa vector bundle’. This is a variant of the similar notion of L ∞ -module introduced by Weaverin [47] who was also interested in developing a calculus on non-smooth spaces. In fact, someof the statements which we shall present in Sections 1.2 and 1.3 can be seen as technicalvariants of analogous statements given in [47]. Still, our axiomatization and the study ofSobolev functions carried out in [7] allow to produce new and interesting links between theabstract differential calculus and the structure of the space: for instance, in Theorem 1.32we shall see that we can associate to ‘almost every absolutely continuous curve’ a derivativewhose modulus coincides with the metric speed of the curve itself. This kind of statement,whose precise formulation requires the notions of ‘test plan’ and of ‘pullback of a module’, iscrucial in applications to geometry, see for instance [19].We also remark that the definition of cotangent module that we give here can be canon-ically identified with the cotangent bundle as built by Cheeger in [16]. We won’t insist onthis point (referring to [25] for more details) because the two approaches are very different inspirit: in [16], working on doubling spaces supporting a Poincar´e inequality, Cheeger gave ametric version of Rademacher’s theorem, which results in much more than a mere definitionof cotangent bundle. Here, instead, we are only interested in giving an abstract and weaknotion of differential of a Sobolev function and we shall do so without imposing any dou-bling or Poincar´e inequality. In any case, our first-order theory should mostly be regarded asfoundational material for the second-order one on RCD spaces.In the second part of the notes we shall work in
RCD spaces, mostly without imposingany dimension bound (we confine to the final Section 2.7 some recent results about calculuson finite dimensional spaces). The definition of
RCD ( K, ∞ ) spaces that we shall adopt is theone, coming from [9], based on the appropriate weak formulation of the Bochner inequality∆ |∇ f | ≥ h∇ f, ∇ ∆ f i + K |∇ f | . (0.1)There is a certain amount of ‘cheating’ in choosing this approach, because it is the closestto differential calculus and the furthest from the fact, crucial for the theory, that the classof RCD ( K, ∞ ) spaces is closed w.r.t. measured-Gromov-Hasdorff convergence. Nevertheless,the validity of Bochner inequality on RCD spaces is now well-established within the theory,so that possibly there is no much harm in taking it as starting point for our discussion. The2eader interested in the stability issue might want to start from the lecture notes [6] for anaccount of the path which starts from the original approach of Lott-Sturm-Villani ([37], [44])and uses the heat flow ([23], [27], [7]) to isolate ‘Riemannian’ spaces ([8]) by also providing astable version of the Bochner inequality ([9]).From the technical point of view, the main result of this second part of the notes (Lemmas2.8 and 2.33) is the improvement of the Bochner inequality from (0.1) to:∆ | X | ≥ |∇ X | − h X, (∆ H X ♭ ) ♯ i + K | X | (0.2)in the appropriate weak sense. Notice that for X = ∇ f , (0.2) reduces to (0.1) with theadditional non-negative contribution | Hess f | on the right hand side. Here the language of L -normed modules provides natural spaces where objects like the Hessian or the covariantderivative belong, and one of the effects of the improved formula (0.2) is the bound Z | Hess f | d m ≤ Z (∆ f ) − K |∇ f | d m (0.3)obtained integrating (0.2) for X = ∇ f (Corollary 2.10). Since functions with gradient andLaplacian in L are easy to build using the heat flow, (0.3) grants that there are ‘many’functions with Hessian in L . Starting from this, it will not be hard to build a second ordercalculus and an indication of the novelty of the theory is in the fact that we can prove thatthe exterior differential is a closed operators on the space of k -forms for any k ∈ N (Theorem2.24), whereas previously known results only covered the case k = 0 ([16], [47], [17]). Inparticular, quite natural versions of the De Rham cohomology and of the Hodge theorem canbe provided (Section 2.5.3)Another consequence of the fact that we have well-defined differential operators is thatwe can define the Ricci curvature as the quantity for which the Bochner identity holds: Ric ( X, X ) := ∆ | X | − |∇ X | + h X, (∆ H X ♭ ) ♯ i . It turns out that
Ric ( X, X ) is a measure-valued tensor and the role of (0.2) is to grant thatthe Ricci curvature is bounded from below by K , as expected.Finally, a feature of the language proposed here is that the differential operators arestable w.r.t. measured-Gromov-Hausdorff convergence of the base spaces in a quite naturalsense. To keep the presentation short we won’t discuss this - important and under continuousdevelopment - topic, referring to [33], [12], [10] for recent results. Acknowledgment
I wish to thank RIMS for the invitation in giving a course there andthe very warm hospitality. This project has also been partly financed by the MIUR SIR-grant‘Nonsmooth Differential Geometry’ (RBSI147UG4).
For the purpose of this note a metric measure space ( X , d , m ) is a complete separable metricspace ( X , d ) endowed with a non-negative (and not zero) Borel measure m giving finite massto bounded sets. 3 ( X ) is the space of Borel probability measures on X and C ([0 , , X ) the space of con-tinuous curves with value in X endowed with the sup norm. For t ∈ [0 ,
1] the evaluation mape t : C ([0 , , X ) → X is defined bye t ( γ ) := γ t , ∀ γ ∈ C ([0 , , X ) . Recall that γ : [0 , → X is absolutely continuous provided there is f ∈ L (0 ,
1) such that d ( γ t , γ s ) ≤ Z st f ( r ) d r, ∀ t, s ∈ [0 , , t < s. (1.1)In this case, for a.e. t ∈ [0 ,
1] there exists | ˙ γ t | := lim h → d ( γ t + h ,γ t ) | h | and | ˙ γ t | is the least, inthe a.e. sense, function f ∈ L (0 ,
1) for which (1.1) holds (see e.g. Theorem 1.1.2 of [5] for aproof).By LIP( X ) (resp. LIP b ( X )) we mean the space of Lipschitz (resp. Lipschitz and bounded)functions on X .There are several equivalent definitions of Sobolev functions on a metric measure space([16], [43], [7]), here we shall adopt one of those proposed in the latter reference, where thenotion of Sobolev function is given in duality with that of test plan: Definition 1.1 (Test Plans) . Let π ∈ P ( C ([0 , , X )) . We say that π is a test plan providedfor some C > we have (e t ) ∗ π ≤ C m , ∀ t ∈ [0 , , Z Z | ˙ γ t | d t d π ( γ ) < ∞ . The least such C is called compression constant of π and denoted as Comp ( π ) . Recall that L ( X ) is the space of (equivalence classes w.r.t. m -a.e. equality of) Borel realvalued functions on X . Definition 1.2 (The Sobolev class S ( X , d , m )) . The Sobolev class S ( X , d , m ) , or simply S ( X ) is the space of all functions f ∈ L ( X ) such that there exists a non-negative G ∈ L ( m ) ,called weak upper gradient of f , for which it holds Z | f ( γ ) − f ( γ ) | d π ( γ ) ≤ Z Z G ( γ t ) | ˙ γ t | d t d π ( γ ) , ∀ π test plan . (1.2)Notice that the assumptions on π grant that the integrals are well defined and that theone in the right hand side is finite. With an argument based on the stability of the class oftest plans by ‘restriction’ and ‘rescaling’ it is not hard to check that f ∈ S ( X ) with G beinga weak upper gradient if and only if for any test plan π and any t, s ∈ [0 , t < s it holds | f ( γ s ) − f ( γ t ) | ≤ Z st G ( γ r ) | ˙ γ r | d r π -a.e. γ. (1.3)Then an application of Fubini’s theorem (see [26] for the details) shows that this is in turnequivalent to: for any test plan π and π -a.e. γ , the function t f ( γ t ) is in W , (0 ,
1) and (cid:12)(cid:12) dd t f ( γ t ) (cid:12)(cid:12) ≤ G ( γ t ) | ˙ γ t | , a . e . t. (1.4)4t is then easy to check that there exists a minimal G in the m -a.e. sense for which (1.2) holds:such G will be called minimal weak upper gradient and denoted by | D f | .From the definitions it is clear that S ( X ) is a vector space and that | D( αf + βg ) | ≤ | α || D f | + | β || D g | ∀ f, g ∈ S ( X ) , α, β ∈ R . (1.5)Beside this, the two crucial properties of minimal weak upper gradients that we shall use are:Lower semicontinuity of minimal weak upper gradients. Let ( f n ) ⊂ S ( X ) and f ∈ L ( X ) besuch that f n → f as n → ∞ in L ( X ) (i.e. m -a.e.). Assume that ( | D f n | ) converges to some G ∈ L ( X ) weakly in L ( X ).Then f ∈ S ( X ) and | D f | ≤ G, m -a.e. . (1.6)Locality. The minimal weak upper gradient is local in the following sense: | D f | = 0 , m -a.e. on { f = 0 } , ∀ f ∈ S ( X ) . (1.7)(1.6) follows quite easily from the very definition of S ( X ), while (1.7) comes from the char-acterization (1.4) and the analogous property of functions in W , (0 , W , ( X ) := L ∩ S ( X ) endowed with the norm k f k W , ( X ) := k f k L ( X ) + k| D f |k L ( X ) . is a Banach space. It is trivial to check that Lipschitz functions with bounded support are in W , ( X ) with | D f | ≤ lip( f ) m -a.e. , where lip( f )( x ) := lim y → x | f ( y ) − f ( x ) | d ( x, y ) if x is not isolated, 0 otherwise.In particular, W , ( X ) is dense in L ( X ). On the other hand it is non-trivial that for every f ∈ W , ( X ) there exists a sequence ( f n ) of Lipschitz functions with bounded support convergingto f in L such that Z | D f | d m = lim n Z lip ( f n ) d m . We shall not use this fact (see [7] for the proof).We conclude recalling that, as shown in [2],if W , ( X ) is reflexive, then it is separable. (1.8)This can be proved considering a countable L -dense set D of the unit ball B of W , ( X ).Then for f ∈ B find ( f n ) ⊂ D converging to f in L ( X ): being ( f n ) bounded in W , ( X ), upto subsequences it must have a weak limit in W , ( X ) and this weak limit must be f . Hencethe weak closure of D is precisely B and by Mazur’s lemma this is sufficient to conclude.5 .2 L -normed modules, cotangent module and differential L -normed modulesDefinition 1.3 ( L ( X )-normed L ∞ ( X )-modules) . A L ( X ) -normed L ∞ ( X ) -module, or simplya L ( X ) -normed module, is a structure ( M , k · k , · , | · | ) wherei) ( M , k · k ) is a Banach spaceii) · is a bilinear map from L ∞ ( X ) × M to M , called multiplication by L ∞ ( X ) functions,such that f · ( g · v ) = ( f g ) · v, (1.9a) · v = v, (1.9b) for every v ∈ M and f, g ∈ L ∞ ( X ) , where is the function identically equal to 1.iii) | · | is a map from M to L ( X ) , called pointwise norm, such that | v | ≥ m -a.e. (1.10a) | f v | = | f | | v | m -a.e. (1.10b) k v k = sZ | v | d m , (1.10c) An isomorphism between two L ( X ) -normed modules is a linear bijection which preserves thenorm, the product with L ∞ ( X ) functions and the pointwise norm. We shall typically write f v in place of f · v for the product with L ∞ ( X ) function.Notice that thanks to (1.9b), for λ ∈ R and v ∈ M the values of λv intended as comingfrom the vector space structure and as the product with the function constantly equal to λ agree, so that the expression is unambiguous. Also, from (1.10b) and (1.10c) we obtain that k f v k ≤ k f k L ∞ k v k . We also remark that the pointwise norm satisfies | λv | = | λ | | v || v + w | ≤ | v | + | w | , m -a.e. for every v, w ∈ M and λ ∈ R . Indeed, the first comes from (1.10b), while for thesecond we argue by contradiction. If it were false, for some v, w ∈ M , Borel set E ⊂ X with m ( E ) ∈ (0 , ∞ ) and positive real numbers a, b, c with a + b < c we would have m -a.e. on E | v + w | ≥ c | v | ≤ a | w | ≤ b However, this creates a contradiction with (1.10c) and the fact that k · k is a norm because k χ E v k + k χ E w k = k χ E | v |k L + k χ E | w |k L ≤ p m ( E ) ( a + b ) < p m ( E ) c ≤ k χ E | v + w |k L = k χ E ( v + w ) k . In the following for given v, w ∈ M and Borel set E ⊂ X we shall say that v = w m -a.e. on E , provided χ E ( v − w ) = 0 or equivalently if | v − w | = 0 m -a.e. on E. xample 1.4. Consider a manifold X equipped with a reference measure m and with anormed vector bundle. Then the space of L ( X , m )-sections of the bundle naturally carriesthe structure of L ( X )-normed module. This is the example which motivates the abstractdefinition of L ( X )-normed module. (cid:4) We say that f ∈ L ∞ ( X ) is simple provided it attains only a finite number of values. Definition 1.5 (Generators) . We say that V ⊂ M generates M provided finite sums of theform P i χ E i v i with ( E i ) Borel partition of X and ( v i ) ⊂ V are dense in M . By approximating L ∞ functions with simple ones, it is easy to see that V generates M ifand only if L ∞ -linear combinations of elements of V are dense in M .A particularly important class of modules is that of Hilbert modules , i.e. modules H which are, when seen as Banach spaces, Hilbert spaces. It is not hard to check that in thiscase the pointwise norm satisfies the pointwise parallelogram identity | v + w | + | v − w | = 2( | v | + | w | ) m -a.e. ∀ v, w ∈ H and thus that by polarization it induces a pointwise scalar product h· , ·i : H → L ( X ) whichis L ∞ ( X )-bilinear and satisfies |h v, w i| ≤ | v | | w | h v, v i = | v | , m -a.e. for every v, w ∈ H .It is at times convenient to deal with objects with less integrability; in this direction, thefollowing concept is useful: Definition 1.6 ( L -normed module) . A L -normed module is a structure ( M , τ, · , | · | ) where:i) · is a bilinear map, called multiplication with L functions, from L ( X ) × M to M forwhich (1.9a) , (1.9b) hold for any f ∈ L ( X ) , v ∈ M ,ii) |·| : M → L ( X ) , called pointwise norm, satisfies (1.10a) and (1.10b) for any f ∈ L ( X ) , v ∈ M ,iii) for some Borel partition ( E i ) of X into sets of finite m -measure, M is complete w.r.t.the distance d ( v, w ) := X i i m ( E i ) Z E i min { , | v − w |} d m (1.11) and τ is the topology induced by the distance.An isomorphims of L -normed modules is a linear homeomorphism preserving the pointwisenorm and the multiplication with L -functions. It is readily checked that the choice of the partition ( E i ) in ( iii ) does not affect thecompleteness of M nor the topology τ . 7 heorem/Definition 1.7 ( L completion of a module) . Let M be a L -normed module.Then there exists a unique couple ( M , ι ) , where M is a L -normed module and ι : M → M is linear, preserving the pointwise norm and with dense image.Uniqueness is intended up to unique isomorphism, i.e.: if ( ˜ M , ˜ ι ) has the same properties,then there exists a unique isomorphism Φ : M → ˜ M such that ˜ ι = Φ ◦ ι .proof Uniqueness is trivial. For existence define M to be the metric completion of M w.r.t. thedistance defined in (1.11) and ι as the natural embedding, then observe that the L -normedmodule structure of M can be extended by continuity and induce an L -normed modulestructure on M . (cid:3) The cotangent module L ( T ∗ X ) and the differential d : S ( X ) → L ( T ∗ X ) are defined, up tounique isomorphism, by the following theorem. The elements of the cotangent module will becalled 1-forms. Theorem/Definition 1.8.
There exists a unique couple ( L ( T ∗ X ) , d) with L ( T ∗ X ) being a L -normed module and d : S ( X ) → L ( T ∗ X ) a linear map such that:i) for any f ∈ S ( X ) it holds | d f | = | D f | m -a.e.,ii) L ( T ∗ X ) is generated by { d f : f ∈ S ( X ) } .Uniqueness is intended up to unique isomorphism, i.e.: if ( M , d ′ ) is another such couple, thenthere is a unique isomorphism Φ : L ( T ∗ X ) → M such Φ(d f ) = d ′ f for every f ∈ S ( X ) . Note: we shall call a form ω ∈ L ( T ∗ X ) simple if it can be written as P i χ A i d f i for afinite Borel partition ( A i ) of X and ( f i ) ⊂ S ( X ). proof Uniqueness
Consider a simple form ω ∈ L ( T ∗ X ) and notice that the requirements that Φis L ∞ -linear and such that Φ(d f ) = d ′ f force the definitionΦ( ω ) := X i χ A i d ′ f i for ω = X i χ A i d f i . (1.12)The identity | Φ( ω ) | = X i χ A i | d ′ f i | ( i ) for M = X i χ A i | D f i | ( i ) for L ( T ∗ X ) = X i χ A i | d f i | = | ω | shows in particular that the definition of Φ( ω ) is well-posed, i.e. Φ( ω ) depends only on ω andnot on the way we represent it as finite sum. It also shows that Φ preserves the pointwisenorm of simple forms and thus, since Φ is clearly linear, grants that Φ is continuous. Beingsimple forms dense in L ( T ∗ X ) (by property ( ii ) for L ( T ∗ X )), Φ can be uniquely extendedby continuity to a map from L ( T ∗ X ) to M and this map is clearly linear, continuous andpreserves the pointwise norm. Also, from the very definition (1.12) we see that Φ( f ω ) = f Φ( ω ) for simple f and ω , so that by approximation we see that the same holds for general f ∈ L ∞ ( X ), ω ∈ L ( T ∗ X ). Property (1.10c) grants that Φ also preserves the norm, so that to8onclude it is sufficient to show that its image is the whole M . This follows from the densityof simple forms in M (property ( ii ) for M ). Existence
We define the ‘Pre-cotangent module’
Pcm to be the set of finite sequences( A i , f i ) with ( A i ) being a Borel partition of X and ( f i ) ⊂ S ( X ). Then we define an equivalencerelation on Pcm by declaring ( A i , f i ) ∼ ( B j , g j ) iff for every i, j we have | D( f i − g j ) | = 0 , m -a.e. on A i ∩ B j . Denoting by [ A i , f i ] the equivalence class of ( A i , f i ), we endow Pcm / ∼ with a vector spacestructure by putting [ A i , f i ] + [ B j , g j ] := [ A i ∩ B j , f i + g j ] ,λ [ A i , f i ] := [ A i , λf i ] . Notice that thanks to the locality property (1.7) of the minimal weak upper gradient, thesedefinitions are well posed. For the same reason, the quantity k [ A i , f i ] k := sX i Z A i | D f i | d m is well defined, and from (1.5) we see that it is a norm. Let ( L ( T ∗ X ) , k · k ) be the completionof ( Pcm / ∼ , k · k ) and d : S ( X ) → L ( T ∗ X ) be the map sending f to [ X , f ]. By construction, L ( T ∗ X ) is a Banach space and d is linear. We want to endow L ( T ∗ X ) with the structure of L ( X )-normed module and to this aim we define | · | : Pcm / ∼→ L ( X ) by | [ A i , f i ] | := X i χ A i | D f i | and a bilinear map { simple functions } × Pcm / ∼ → Pcm / ∼ by (cid:16) X j α j χ E j (cid:17) · [ A i , f i ] := [ A i ∩ E j , α j f i ] , where ( E j ) is a finite partition of X . It is readily verified that these definitions are well posedand that properties (1.9) and (1.10) hold for simple functions and elements of Pcm / ∼ . It isalso clear that || ω | − | ω || ≤ | ω − ω | m -a.e. for every ω , ω ∈ Pcm / ∼ and therefore wehave k| ω | − | ω |k L ≤ k ω − ω k , showing that the pointwise norm can, and will, be extended by continuity to the whole L ( T ∗ X ). Similarly, for h : X → R simple and ω ∈ Pcm / ∼ from the identity | hω | = | h || ω | weobtain k hω k = Z | hω | d m ≤ k h k L ∞ Z | ω | d m = k h k L ∞ k ω k , showing that the multiplication by simple functions on Pcm / ∼ can, and will, be extended bycontinuity to a multiplication by L ∞ ( X ) functions on L ( T ∗ X ).The fact that properties (1.9) and (1.10) hold for these extensions follows trivially byapproximation. Hence L ( T ∗ X ) is a L ( X )-normed module.To conclude, notice that property ( i ) is a direct consequence of the definition of d andof the pointwise norm. The fact that L ( T ∗ X ) is generated by { d f : f ∈ S ( X ) } also followsby the construction once we observe that the typical element [ A i , f i ] of Pcm / ∼ is equal to P i χ A i d f i by the very definitions given. (cid:3) emark 1.9. By a simple cut-off and truncation argument we see that { d f : f ∈ W , ( X ) } also generates L ( T ∗ X ). Hence, slightly more generally, we also have that if D is a densesubset of W , ( X ), then { d f : f ∈ D } generates L ( T ∗ X ).This also shows that if W , ( X ) is separable, then so is L ( T ∗ X ). (cid:4) Remark 1.10.
It is not hard to check that if X is a smooth Finsler manifold, then W , ( X )as we defined it coincides with the Sobolev space defined via charts and that | D f | coincidesa.e. with the norm of the distributional differential.From this fact and Theorem 1.8 it follows that the cotangent module can be identifiedwith the space of L sections of the cotangent bundle via the map which sends d f to thedistributional differential of f . (cid:4) Proposition 1.11 (Closure of the differential) . Let ( f n ) ⊂ S ( X ) be a sequence m -a.e. con-verging to some function f ∈ L ( X ) . Assume that (d f n ) converges to some ω ∈ L ( T ∗ X ) inthe weak topology of L ( T ∗ X ) seen as Banach space.Then f ∈ S ( X ) and d f = ω .proof By applying Mazur’s lemma we can assume that the convergence of (d f n ) to ω isstrong in L ( T ∗ X ). In particular ( | d f n | ) converges to | ω | in L ( X ) and by (1.6) this grantsthat f ∈ S ( X ). For any m ∈ N we have that f n − f m → f − f m m -a.e., thus using again (1.6)we have k d f − d f n k L ( T ∗ X ) = k| D( f − f n ) |k L ( X ) ≤ lim m k| D( f m − f n ) |k L ( X ) = lim m k d f m − d f n k L ( T ∗ X ) and the conclusion follows letting n → ∞ using the fact that, being (d f n ) strongly convergingin L ( T ∗ X ), it is a Cauchy sequence. (cid:3) Proposition 1.12 (Calculus rules) . The following holds.-
Locality
For every f, g ∈ S ( X ) we have d f = d g m -a.e. on { f = g } . (1.13) - Chain rule
For every f ∈ S ( X ) and ϕ ∈ LIP ∩ C ( R ) we have ϕ ◦ f ∈ S ( X ) and d( ϕ ◦ f ) = ϕ ′ ◦ f d f. (1.14) - Leibniz rule
For every f, g ∈ L ∞ ∩ S ( X ) we have f g ∈ S ( X )d( f g ) = f d g + g d f. (1.15) proof Locality
By the linearity of the differential the claim is equivalent tod f = 0 m -a.e. on { f = 0 } which follows directly from | d f | = | D f | m -a.e. and the locality property (1.7) of | D f | . Chain rule
The fact that Lip( ϕ ) | D f | ∈ L ( X ) is a weak upper gradient for ϕ ◦ f is obvious,hence in particular ϕ ◦ f ∈ S ( X ). 10o prove (1.14), start noticing that taking into account the linearity of the differentialand the fact that constant functions have 0 differential (because trivially their minimal weakupper gradient is 0), the chain rule (1.14) is trivial if ϕ is affine. Hence by the locality property(1.13) the chain rule (1.14) holds if ϕ is piecewise affine. Notice that this also forces d f to be0 m -a.e. on f − ( z ) for any z ∈ R , and thus also m -a.e. on f − ( N ) for N ⊂ R countable.Let now ϕ ∈ LIP ∩ C ( R ) and find a sequence ( ϕ n ) of equi-Lipschitz and piecewise affinefunctions such that ( ϕ n ) , ( ϕ ′ n ) uniformly converge to ϕ, ϕ ′ respectively. From these, whatpreviously said and the closure of the differential we can pass to the limit ind( ϕ n ◦ f ) = ϕ ′ n ◦ f d f and conclude. Leibniz rule
From the characterization (1.4) it easily follows that | g || D f | + | f || D g | ∈ L ( X )is a weak upper gradient for f g , so that f g ∈ S ( X ). Now assume that f, g ≥ m -a.e.. Thenalso f g ≥ m -a.e. and we can apply the chain rule with ϕ = log, which is Lipschitz on theimage of f, g and f g , to getd( f g ) f g = d(log( f g )) = d(log f + log g ) = d log f + d log g = d ff + d gg , which is the thesis. The general case now follows easily replacing f, g by f + C, g + C for C ∈ R large enough. (cid:3) (Dual of a module) . Let M be a L ( X ) -normed module. Its dual M ∗ is thespace of linear continuous maps L : M → L ( X ) such that L ( f v ) = f L ( v ) , ∀ f ∈ L ∞ ( X ) , v ∈ M . We equip M ∗ with the operator norm, i.e. k L k ∗ := sup v : k v k≤ k L ( v ) k L . The multiplication of f ∈ L ∞ ( X ) and L ∈ M ∗ is defined as ( f L )( v ) := L ( f v ) , ∀ v ∈ M . Finally, the pointwise norm | L | ∗ of L ∈ M ∗ is defined as | L | ∗ := ess-sup v : | v |≤ m -a.e. | L ( v ) | . The only non-trivial thing to check in order to show that the structure just defined is a L -normed module is property (1.10c) (which also grants that | L | ∗ belongs to L ( X )). Fromthe definition it is not hard to check that | L ( v ) | ≤ | L | ∗ | v | m -a.e. ∀ v ∈ M , L ∈ M ∗ , and thus by integration we get k L ( v ) k L ≤ k v kk| L | ∗ k L showing that k L k ∗ ≤ k| L | ∗ k L .For the opposite inequality notice that from the basic properties of the essential supremumthere is a sequence ( v n ) ⊂ M such that | v n | ≤ m -a.e. for every n ∈ N satisfying | L | ∗ =11up n | L ( v n ) | . Put ˜ v := v and for n > A n := {| L ( v n ) | > | L (˜ v n − ) |} and˜ v n := χ A n v n + χ A cn ˜ v n − . Then | ˜ v n | ≤ m -a.e. and the sequence ( | L (˜ v n ) | ) is increasing andconverges m -a.e. to | L | ∗ . Pick f ∈ L ∩ L ∞ ( X ) arbitrary, notice that k f ˜ v n k = k| f ˜ v n |k L ≤k f k L and thus Z | f || L (˜ v n ) | d m = Z | L ( f ˜ v n ) | d m ≤ k f ˜ v n k k L k ∗ = k f k L k L k ∗ ∀ n ∈ N . By the monotone convergence theorem the integral on the left goes to R | f || L | ∗ d m as n → ∞ ,hence passing to the limit we obtain Z | f || L | ∗ d m ≤ k f k L k L k ∗ and being this true for every f ∈ L ∩ L ∞ ( X ) we conclude that k| L | ∗ k L ≤ k L k ∗ , as desired.We shall frequently use the fact that for L : M → L ( X ) linear and continuous we have L ∈ M ∗ ⇔ L ( χ E v ) = χ E L ( v ) for every E ⊂ X Borel and v ∈ M , (1.16)which can be proved by first checking that L ( f v ) = f L ( v ) holds for simple f and then arguingby approximation.Denote by M ′ the dual of M seen as Banach space, so that M ′ is the Banach space of linearcontinuous maps from M to R equipped with its canonical norm k · k ′ . Integration provides anatural map Int : M ∗ → M ′ sending L ∈ M ∗ to the operator Int ( L ) ∈ M ′ defined as Int ( L )( v ) := Z L ( v ) d m , ∀ v ∈ M . Proposition 1.14.
The map
Int is a bijective isometry, i.e. k L k ∗ = k Int ( L ) k ′ for every L ∈ M ∗ .proof The trivial bound | Int ( L )( v ) | = (cid:12)(cid:12)(cid:12) Z L ( v ) d m (cid:12)(cid:12)(cid:12) ≤ k L ( v ) k L ≤ k v kk L k ∗ shows that k Int ( L ) k ′ ≤ k L k ∗ . To prove the converse, fix L ∈ M ∗ , ε > v ∈ M suchthat k L ( v ) k L ≥ k v k ( k L k ∗ − ε ). Put ˜ v := χ { L ( v ) ≥ } v − χ { L ( v ) < } v , notice that | ˜ v | = | v | and L (˜ v ) = | L ( v ) | m -a.e. and conclude by k Int ( L ) k ′ k ˜ v k ≥ | Int ( L )(˜ v ) | = (cid:12)(cid:12)(cid:12) Z L (˜ v ) d m (cid:12)(cid:12)(cid:12) = k L ( v ) k L ≥ k v k ( k L k ∗ − ε ) = k ˜ v k ( k L k ∗ − ε )and the arbitrariness of ε >
0. Thus it remains to prove that
Int is surjective.Pick ℓ ∈ M ′ , fix v ∈ M and consider the map sending a Borel set E to µ v ( E ) := ℓ ( χ E v ) ∈ R . It is additive and given a disjoint sequence ( E i ) of Borel sets we have | µ v ( ∪ n E n ) − µ v ( ∪ Nn =1 E n ) | = | µ v ( ∪ n>N E n ) | = | ℓ ( χ ∪ n>N E n v ) | ≤ k ℓ k ′ k χ ∪ n>N E n v k and since k χ ∪ n>N E n v k = R ∪ n>N E n | v | d m → µ v is a Borel measure. By construction, it is also absolutely continuous w.r.t. m and thusit has a Radon-Nikodym derivative, which we shall denote by L ( v ) ∈ L ( X ).12he construction trivially ensures that v L ( v ) is linear and since for every E, F ⊂ X Borel the identities µχ E v ( F ) = ℓ ( χ F χ E v ) = ℓ ( χ E ∩ F v ) = µ v ( E ∩ F ) grant that R F L ( χ E v ) = R E ∩ F L ( v ), we see that L ( χ E v ) = χ E L ( v ) ∀ v ∈ M , E ⊂ X Borel . (1.17)Now given v ∈ M we put ˜ v := χ { L ( v ) ≥ } v − χ { L ( v ) < } v so that | ˜ v | = | v | and, by (1.17) and thelinearity of L we have | L ( v ) | = L (˜ v ) m -a.e.. Then k L ( v ) k L = Z L (˜ v ) d m = µ ˜ v,ℓ ( X ) = ℓ (˜ v ) ≤ k ℓ k ′ k ˜ v k = k ℓ k ′ k v k , i.e. v L ( v ) is continuous. The conclusion follows from (1.17) and (1.16). (cid:3) The Hanh-Banach theorem grants that for every v ∈ M there exists ℓ ∈ M ′ with k ℓ k ′ = k v k and | ℓ ( v ) | = k v k . Putting L := Int − ( v ), from the fact that the inequalities k v k = ℓ ( v ) = Z L ( v ) d m ≤ Z | L | ∗ | v | d m ≤ k| v |k L k| L | ∗ k L = k v kk L k ∗ = k v kk ℓ k ′ = k v k are in fact equalities we deduce that m -a.e. it holds | L | ∗ = | v | L ( v ) = | v | . (1.18)It follows that the natural embedding I : M → M ∗∗ sending v to the map L L ( v ),which is trivially L ∞ -linear, preserves the pointwise norm. Indeed, since for any v, L wehave | I ( v )( L ) | = | L ( v ) | ≤ | v || L | ∗ we have | I ( v ) | ∗∗ ≤ | v | , while the opposite inequality comesconsidering L such that (1.18) holds.Modules M for which I is surjective will be called reflexive . Proposition 1.15 (Riesz theorem for Hilbert modules and reflexivity) . Let H be an Hilbertmodule and consider the map sending v ∈ H to L v ∈ H ∗ given by L v ( w ) := h v, w i .Then this map is an isomorphism of modules. In particular, Hilbert modules are reflexive.proof The only non-trivial claim about the map v L v is surjectivity. To check it, let L ∈ H ∗ , consider Int ( L ) ∈ H ′ and apply the standard Riesz theorem to find v ∈ H suchthat Z L ( w ) d m = Int ( L )( w ) = h v, w i H = Z h v, w i d m ∀ w ∈ H , where h· , ·i H is the scalar product in the Hilbert space H and the last identity follows from(1.10c) by polarization. Writing χ E w in place of w in the above for E ⊂ X Borel arbitrarywe see that L ( w ) = h v, w i m -a.e., i.e. L = L v . The claim about reflexivity is now obvious. (cid:3) Proposition 1.16.
Let M be a L ( X ) -normed module V ⊂ M a vector subspace which gen-erates M and L : V → L ( X ) a linear map. Assume that for some g ∈ L ( X ) we have | L ( v ) | ≤ g | v | m -a.e. ∀ v ∈ V. (1.19) Then there is a unique ˜ L ∈ M ∗ such that ˜ L ( v ) = L ( v ) for every v ∈ V and for such ˜ L wehave | ˜ L | ∗ ≤ g . roof Any extension ˜ L of L which is L ∞ ( X )-linear must be such that˜ L ( v ) = X i χ E i L ( v i ) , for v = X i χ E i v i (1.20)where ( E i ) is a finite partition of X and ( v i ) ⊂ V . For ˜ L defined in this way, the bound (1.19)gives that | ˜ L ( v ) | = X i χ E i | L ( v i ) | ≤ X i χ E i g | v i | = g (cid:12)(cid:12)(cid:12) X i χ E i v i (cid:12)(cid:12)(cid:12) = g | v | and in particular k ˜ L ( v ) k L ( X ) ≤ k g k L ( X ) k v k . This shows that the definition (1.20) is well-posed - in the sense that ˜ L ( v ) depends only on v and not on the way to represent it as P i χ E i v i - and that it is continuous. Since by assumption the set of v ’s of the form P i χ E i v i is dense in M , we can uniquely extend ˜ L to a continuous operator ˜ L : M → L ( X ). The factthat such ˜ L is linear is obvious and the definition (1.20) easily gives that ˜ L ( f v ) = f ˜ L ( v ) holdsfor simple functions f . Then L ∞ -linearity follows by approximation. (cid:3) We conclude with the following proposition, which in some sense says that the operationsof taking the dual and of taking the L -completion (recall Theorem 1.7) commute: Proposition 1.17.
Let M be a L -normed module. Then the duality pairing M × M ∗ → L ( X ) uniquely extends to a continuous duality pairing M × ( M ∗ ) → L ( X ) . Moreover, if L : M → L ( X ) is such that for some g ∈ L ( X ) it holds | L ( v ) | ≤ g | v | m -a.e. ∀ v ∈ M , (1.21) then L ∈ ( M ∗ ) (in the sense of the previously defined pairing).proof The claim about the unique continuous extension is a trivial consequence of the defi-nitions. For the second part of the claim just notice that we can always find a sequence ( E n )of Borel sets such that χ E n g ∈ L ( X ) for every n ∈ N and ( χ E n g ) → g in L ( X ). Then from(1.21) and Proposition 1.16 above with V = M we see that the map v L n ( v ) := χ E n L ( v )belongs to M ∗ . Since clearly | L n − L m | ∗ ≤ | χ E n − χ E m | g , the sequence ( L n ) is Cauchy in( M ∗ ) and its limit is easily seen to be equal to L . (cid:3) (Tangent module) . The tangent module L ( T X ) is defined as the dual ofthe cotangent module L ( T ∗ X ) . Its elements are called vector fields. To keep consistency with the notation used in the smooth setting, we shall denote thepointwise norm in L ( T X ) as |·| , rather than |·| ∗ , and the duality pairing between ω ∈ L ( T ∗ X )and X ∈ L ( T X ) as ω ( X ). Definition 1.19 ( L derivations) . A L -derivation is a linear map L : S ( X ) → L ( X ) forwhich there is g ∈ L ( X ) such that | L ( f ) | ≤ g | D f | ∀ f ∈ S ( X ) . (1.22)14otice that the concept of derivation has a priori nothing to do with the notion of L -normed module. It is therefore interesting to see that such notion emerges naturally from theconcept of derivation, because as the following theorem shows, derivations and vector fieldsare two different points of view on the same kind of object. The same result, in conjunctionwith the Leibniz rule (1.15), also shows that, although not explicitly encoded in the definition,derivations satisfy the Leibniz rule L ( f g ) = f L ( g ) + gL ( f ) for any f, g ∈ L ∞ ∩ S ( X ). Theorem 1.20 (Derivations and vector fields) . For any vector field X ∈ L ( T X ) the map X ◦ d : S ( X ) → L ( X ) is a derivation.Conversely, given a derivation L there exists a unique vector field X ∈ L ( T X ) such thatthe diagram S ( X ) L ( T ∗ X ) L ( X )d XL commutes.proof The first claim follows from the linearity of X ◦ d, the fact that | X | ∈ L ( X ) and theinequality | d f ( X ) | ≤ | X | | d f | = | X | | D f | valid m -a.e. for any f ∈ S ( X ).For the second, let L be a derivation, put V := { d f : f ∈ S ( X ) } and define ˜ L : V → L ( X )by ˜ L (d f ) := L ( f ). Inequality (1.22) grants that this is a good definition, i.e. ˜ L (d f ) dependsonly on d f and not on f , and that | ˜ L (d f ) | ≤ g | d f | . The conclusion then follows from Proposition 1.16 recalling that V generates L ( T ∗ X ). (cid:3) Taking the adjoint of the differential leads to the notion of divergence:
Definition 1.21 (Divergence) . We say that X ∈ L ( T X ) has divergence in L , and write X ∈ D (div) provided there is h ∈ L ( X ) such that Z f h d m = − Z d f ( X ) d m ∀ f ∈ W , ( X ) . (1.23) In this case we shall call h the divergence of X and denote it by div( X ) . Notice that by the density of W , ( X ) in L ( X ) there is at most one h satisfying (1.23),hence the divergence is unique.It is also easily verified that for X ∈ D (div) and g ∈ LIP b ( X ) we have gX ∈ D (div) withdiv( gX ) = d g ( X ) + g div( X ) , (1.24)indeed, start observing that replacing f with min { max { f, − n } , n } in (1.23) and then sending n → ∞ , we can reduce to check (1.23) for f ∈ L ∞ ∩ W , ( X ). For such f we can apply theLeibniz rule (1.15) to get Z f (d g ( X ) + g div( X )) d m = Z f d g ( X ) − d( f g )( X ) d m = − Z g d f ( X ) d m , D (div) contains an non-zero vector field;in this direction, see (1.43). The concept of pullback of a module mimics the one of pullback of a bundle.
Definition 1.22 (Maps of bounded compression) . Let ( X , m X ) and (Y , m Y ) be measuredspaces. We say that ϕ : Y → X has bounded compression provided ϕ ∗ m Y ≤ C m X for some C > . The least such constant C is called compression constant and denoted by Comp( ϕ ) . Theorem/Definition 1.23 (Pullback module and pullback map) . Let M be a L ( X ) -normedmodule and ϕ : Y → X a map of bounded compression.Then there exists a unique couple ( ϕ ∗ M , ϕ ∗ ) with ϕ ∗ M being a L (Y) -normed module and ϕ ∗ : M → ϕ ∗ M linear and continuous such thati) for every v ∈ M it holds | ϕ ∗ v | = | v | ◦ ϕ m Y -a.e.ii) ϕ ∗ M is generated by { ϕ ∗ v : v ∈ M } .Uniqueness is intended up to unique isomorphism, i.e.: if ( ] ϕ ∗ M , e ϕ ∗ ) is another such couple,then there is a unique isomorphism Φ : ϕ ∗ M → ] ϕ ∗ M such that Φ( ϕ ∗ v ) = e ϕ ∗ v for any v ∈ M , Note: we call an element of ϕ ∗ M simple if it can be written as P i χ A i ϕ ∗ v i for some finiteBorel partition ( A i ) of Y and elements v i ∈ M . Sketch of the proof
Uniqueness
As in the proof of Theorem 1.8, any such Φ must send the simple element P i χ A i ϕ ∗ v i to P i χ A i e ϕ ∗ v i and properties ( i ) , ( ii ) grant that this is a good definition and thatΦ can uniquely be extended by continuity to a map which is the desired isomorphism. Existence
Consider the set ‘Pre-Pullback Module’
Ppb defined as
Ppb := { ( A i , v i ) i =1 ,...,n : n ∈ N , ( A i ) is a Borel partition of Y and v i ∈ M ∀ i = 1 , . . . , n } , define an equivalence relation on it by declaring ( A i , v i ) ∼ ( B j , w j ) provided | v i − w j | ◦ ϕ = 0 m Y -a.e. on A i ∩ B j , ∀ i, j and the map ϕ ∗ : M → Ppb / ∼ which sends v to the equivalence class of (Y , v ). Theconstruction now proceeds as for the cotangent module given in Theorem 1.8: one defineson Ppb / ∼ a vector space structure, a multiplication by simple functions on Y, a pointwisenorm and a norm, then passes to the completion to conlude. We omit the details. (cid:3) Example 1.24. If M = L ( X ), then ϕ ∗ M is (=can be identified with) L (Y), the pullbackmap being given by ϕ ∗ f = f ◦ ϕ . (cid:4) xample 1.25. If (Y , m Y ) is the product of ( X , m X ) and another measured space (Z , m Z )and ϕ : Y → X is the natural projection, then the pullback of M via ϕ is (=can be identifiedwith) L (Z , M ) with the pullback map being the one assigning to a given v ∈ M the functionidentically equal to v .Notice indeed that L (Z , M ) admits a canonical multiplication with functions in L ∞ (Y) = L ∞ ( X × Z): the product of z v ( z ) ∈ M and f ( x, z ) ∈ L ∞ ( X × Z) is z f ( · , z ) v ( z ) ∈ M .Also, on L (Z , M ) there is a natural pointwise norm: the one assigning to z v ( z ) ∈ M themap ( x, z )
7→ | v ( z ) | ( x ).The claim is now easily verified. (cid:4) Proposition 1.26 (Universal property of the pullback) . Let M be a L ( X ) -normed module, ϕ : Y → X a map of bounded compression, N a L (Y) -normed module and T : M → N linearand such that for some C > it holds | T ( v ) | ≤ C | v | ◦ ϕ m Y -a.e. . Then there exists a unique L ∞ (Y) -linear and continuous map ˆ T : ϕ ∗ M → N such that ˆ T ( ϕ ∗ v ) = T ( v ) ∀ v ∈ M . Sketch of the proof
Consider the space V := { ϕ ∗ v : v ∈ M } , which generates ϕ ∗ M , and themap L : V → N given by L ( ϕ ∗ v ) := T ( v ), then argue as for Proposition 1.16. (cid:3) Remark 1.27 (Functoriality of the pullback) . A direct consequence of this last proposition isthat if ϕ : Y → X and ψ : Z → Y are both of bounded compression and M is a L ( X )-normedmodule, then ψ ∗ ϕ ∗ M can be canonically identified to ( ψ ◦ ϕ ) ∗ M via the only isomorphismwhich sends ψ ∗ ϕ ∗ v to ( ψ ◦ ϕ ) ∗ v for every v ∈ M . (cid:4) Remark 1.28 (The case of invertible ϕ ) . If ϕ is invertible with inverse of bounded deforma-tion, then the previous remark grants that ϕ ∗ is bijective. Moreover, the right compositionwith ϕ provides an isomorphism of L ∞ ( X ) and L ∞ (Y) and under this isomorphism the mod-ules M and ϕ ∗ M can be identified, the isomorphism being ϕ ∗ . (cid:4) Consider now also the dual M ∗ of the module M and its pullback ϕ ∗ M ∗ . There is a naturalduality relation between ϕ ∗ M and ϕ ∗ M ∗ : Proposition 1.29.
There exists a unique L ∞ (Y) -bilinear and continuous map from ϕ ∗ M × ϕ ∗ M ∗ to L (Y) such that ϕ ∗ ω ( ϕ ∗ v ) = ω ( v ) ◦ ϕ ∀ v ∈ M , ω ∈ M ∗ (1.25) and for such map it holds | W ( V ) | ≤ | W | ∗ | V | ∀ V ∈ ϕ ∗ M , W ∈ ϕ ∗ M ∗ . (1.26) proof Considering simple elements W ∈ ϕ ∗ M ∗ and V ∈ ϕ ∗ M we see that the requirement(1.25) and L ∞ ( Y )-bilinearity force the definition W ( V ) := X i,j χ A i ∩ B j ω i ( v j ) ◦ ϕ for W = X i χ A i ϕ ∗ ω i V := X j χ B j ϕ ∗ v j . (1.27)17he bound (cid:12)(cid:12)(cid:12) X i,j χ A i ∩ B j ω i ( v j ) ◦ ϕ (cid:12)(cid:12)(cid:12) ≤ X i,j χ A i ∩ B j | ω i | ◦ ϕ | v j | ◦ ϕ = X i χ A i | ω i | ◦ ϕ X j χ B j | v j | ◦ ϕ = | W | | V | shows that the above definition is well posed, in the sense that the definition of W ( V ) dependsonly on V, W and not on the way they are written as finite sums. The same bound also showsthat (1.26) holds for simple elements and that k W ( V ) k L ( Y ) ≤ k W k ϕ ∗ M ∗ k V k ϕ ∗ M .Since the definition (1.27) also trivially grants that ( f W )( gV ) = f gW ( V ) for f, g simple,all the conclusions follow by the density of simple elements in the respective modules. (cid:3) The last proposition can be read by saying that there is a natural embedding I of ϕ ∗ M ∗ into ( ϕ ∗ M ) ∗ which sends W ∈ ϕ ∗ M ∗ into the map ϕ ∗ M ∋ V W ( V ) ∈ L (Y) . Routine computations shows that I is a module morphism which preserves the pointwisenorm. It is natural to wonder whether it is surjective, i.e. whether ϕ ∗ M ∗ can be identified ornot with the dual of ϕ ∗ M . Example 1.25 and Proposition 1.14 show that in general the answeris negative, because in such case our question can be reformulated as: is the dual of L (Z , M )given by L (Z , M ∗ )? It is known (see e.g. [20]) that the answer to this latter question is yes ifand only if M ∗ has the Radon-Nikodym property and that this is ensured if M ∗ is separable.In our case we have the following result: Theorem 1.30 (Identification of ϕ ∗ M ∗ and ( ϕ ∗ M ) ∗ ) . Let ( X , d X , m X ) , (Y , d Y , m Y ) be twocomplete and separable metric spaces equipped with non-negative Borel measures finite onbounded sets and ϕ : Y → X of bounded compression. Let M be a L ( X ) normed module suchthat its dual M ∗ is separable.Then I : ϕ ∗ M ∗ → ( ϕ ∗ M ) ∗ is surjective. The proof of this result is rather technical: we shall omit it, referring to [25] for the details.Here we instead prove the following much simpler statement:
Proposition 1.31.
Let ( X , m X ) and (Y , m Y ) be two measured spaces, ϕ : Y → X of boundedcompression and H an Hilbert module on X .Then I : ϕ ∗ H ∗ → ( ϕ ∗ H ) ∗ is surjective.proof The pointwise norm of H satisfies the pointwise parallelogram identity, hence thesame holds for the pointwise norm of ϕ ∗ H (check first the case of simple elements, thenargue by approximation). Thus ϕ ∗ H is a Hilbert module. Now let R : H → H ∗ andˆ R : ϕ ∗ H → ( ϕ ∗ H ) be the respective Riesz isomorphisms (recall Proposition 1.15), consider ϕ ∗ ◦ R : H → ϕ ∗ ( H ∗ ) and the induced map \ ϕ ∗ ◦ R : ϕ ∗ H → ϕ ∗ ( H ∗ ) as given by Proposition1.26.It is then readily verified that \ ϕ ∗ ◦ R ◦ ˆ R − : ( ϕ ∗ H ) ∗ → ϕ ∗ H ∗ is the inverse of I : ϕ ∗ H ∗ → ( ϕ ∗ H ) ∗ , thus giving the result. (cid:3) With the aid of the concept of pullback of a module we can now assign to any test planits ‘derivative’ π ′ t for a.e. t . The maps of bounded compression that we shall consider are18he evaluation maps e t from C ([0 , , X ) endowed with a test plan π as reference measure to( X , d , m ). In this case, we shall denote the pullback of the tangent bundle L ( T X ) via e t by L ( T X , e t , π ). Theorem/Definition 1.32.
Let ( X , d , m ) be a metric measure space such that L ( T X ) isseparable and π a test plan.Then for a.e. t ∈ [0 , there exists a unique vector field π ′ t ∈ L ( T X , e t , π ) such that forevery f ∈ W , ( X ) the identity lim h → f ( γ t + h ) − f ( γ t ) h = (e ∗ t d f )( π ′ t )( γ ) , (1.28) holds, the limit being intended in the strong topology of L ( π ) . For these π ′ t ’s we also have | π ′ t | ( γ ) = | ˙ γ t | , π × L | [0 , -a.e. ( γ, t ) . (1.29) Sketch of the proof
Start observing that since L ( T X ) is separable and isometric to theBanach dual of L ( T ∗ X ) (Proposition 1.14), L ( T ∗ X ) is also separable. Then observe thatsince f ( f, d f ) is an isometry of W , ( X ) into L ( X ) × L ( T ∗ X ) with the norm k ( f, ω ) k := k f k L ( X ) + k ω k L ( T ∗ X ) , the space W , ( X ) is separable as well.Now pick f ∈ W , ( X ), define [0 , ∋ t F t , G t ∈ L ( π ) as F t ( γ ) := f ( γ t ) G t ( γ ) := | D f | ( γ t ) | ˙ γ t | , and notice that (1.3) can be written as | F s − F t | ≤ Z st G r d r π -a.e. . (1.30)Integrating this bound w.r.t. π we see in particular that the map t F t ∈ L ( π ) is abso-lutely continuous. Although this is not sufficient to deduce that such curve is differentiableat a.e. t (because the Banach space L ( π ) does not have the Radon-Nikodym property), thepointwise bound (1.30) grants uniform integrability of the incremental ratios F t + h − F t h and inturn this grants that for some h n ↓ F · + hn − F · h n converges in the weak topology of L ( L | [0 , × π ) to a limit function Der · ( f ) which by (1.30) and the definition of G t satisfies | Der t ( f ) | ( γ ) ≤ | D f | ( γ t ) | ˙ γ t | = | e ∗ t d f | ( γ ) | ˙ γ t | L | [0 , × π -a.e. ( t, γ ) . (1.31)From the definition of Der t ( f ) it also follows that F s − F t = Z st Der r ( f ) d r ∀ t, s ∈ [0 , , t < s, and this in turn implies that F t + h − F t h converge to Der t ( f ) strongly in L ( π ) as h → t ∈ [0 , W , ( X ) is separable, we can thensee that the exceptional set of t ’s is independent on f , so that for a.e. t we have: ∀ f ∈ W , ( X ) f ◦ e t + h − f ◦ e t h converge in L ( π ) to some Der t ( f ) for which (1.31) holds.19ix t for which this holds and let L t : { e ∗ t d f : f ∈ W , ( X ) } → L ( π ) be defined as L t (e ∗ t d f ) := Der t ( f ). The bound (1.31) grants that this is a good definition, then using Proposition 1.16and Theorem 1.30 (recall that we assumed L ( T X ) to be separable) we deduce that thereexists a unique π ′ t ∈ L ( T X , e t , π ) such thate ∗ t d f ( π ′ t ) = Der t ( f ) ∀ f ∈ W , ( X )and that inequality ≤ in (1.29) holds. To prove ≥ notice that for f ∈ W , ∩ LIP( X ) and γ absolutely continuous the map t f ( γ t ) is absolutely continuous. Therefore the derivative dd t f ( γ t ) is well defined for π × L | [0 , -a.e. ( γ, t ) and it is easy to check that it π × L | [0 , -a.e.coincides with Der t ( f )( γ ). Thus π × L | [0 , -a.e. ( γ, t ) we havedd t f ( γ t ) = e ∗ t d f ( π ′ t )( γ ) ≤ | e ∗ t d f | ( γ ) | π ′ t | ( γ ) = | d f | ( γ t ) | π ′ t | ( γ ) ≤ Lip( f ) | π ′ t | ( γ ) . Hence to conclude it is sufficient to show that there exists a countable family D of 1-Lipschitzfunctions in W , ( X ) such that for any absolutely continuous curve γ we havesup f ∈ D dd t f ( γ t ) ≥ | ˙ γ t | , a . e . t. (1.32)Let ( x n ) ⊂ X be countable and dense and define f n,m ( x ) := max { , m − d ( x, x n ) } . It is clearthat f n,m ∈ W , ∩ LIP( X ) and that d ( x, y ) = sup n,m f n,m ( x ) − f n,m ( y ), thus for γ absolutelycontinuous we have d ( γ s , γ t ) = sup n,m f n,m ( γ s ) − f n,m ( γ t ) = sup n,m Z st dd r f n,m ( γ r ) d r ≤ Z st sup n,m dd r f n,m ( γ r ) d r and the claim (1.32) follows. (cid:3) In applications one can often find explicit expressions for the vector fields π ′ t in terms ofthe data of the problem, so that this last theorem can be used to effectively calculate thederivative of f ◦ e t , see for instance Remark 1.45. Here we introduce maps between metric measure spaces which are ‘first-order smooth’ and seethat they naturally induce a pull-back of 1-forms and, by duality, that they have a differential.
Definition 1.33 (Maps of bounded deformation) . Let ( X , d X , m X ) and (Y , d Y , m Y ) be metricmeasure spaces. A map ϕ : Y → X is said of bounded deformation provided it is Lipschitz andof bounded compression (Definition 1.22). A map of bounded deformation induces by left composition a map ˆ ϕ : C ([0 , , Y) → C ([0 , , X ). It is clear that if γ is absolutely continuous then so is ˆ ϕ ( γ ) and, denoting byms t ( ˆ ϕ ( γ )) its metric speed at time t , thatms t ( ˆ ϕ ( γ )) ≤ Lip( ϕ ) | ˙ γ t | a . e . t. (1.33)Also, for µ ∈ P (Y) such that µ ≤ C m Y we have ϕ ∗ µ ≤ C Comp( ϕ ) m X . It follows that if π isa test plan on Y, then ˆ ϕ ∗ π is a test plan on X .20y duality, we now check that for f ∈ S ( X ) we have f ◦ ϕ ∈ S (Y) with | d( f ◦ ϕ ) | ≤ Lip( ϕ ) | d f | ◦ ϕ m Y -a.e. . (1.34)Indeed, let π be a test plan on Y and notice that Z | f ( ϕ ( γ )) − f ( ϕ ( γ )) | d π ( γ ) = Z | f (˜ γ ) − f (˜ γ ) | d ˆ ϕ ∗ π (˜ γ )because ˆ ϕ ∗ π is a test plan on X ≤ Z Z | d f | (˜ γ t )ms t (˜ γ ) d ˆ ϕ ∗ π (˜ γ )= Z Z | d f | ( ϕ ( γ t ))ms t ( ˆ ϕ ( γ )) d π ( γ )by (1.33) ≤ Lip( ϕ ) Z Z | d f | ( ϕ ( γ t )) | ˙ γ t | d π ( γ ) , which, by the arbitrariness of π and the very definition of S (Y) and minimal weak uppergradient, gives the claim.A direct consequence of this simple observation is: Theorem/Definition 1.34 (Pull-back of 1-forms) . Let ϕ : Y → X be of bounded deforma-tion. Then there exists a unique linear and continuous map ϕ ∗ : L ( T ∗ X ) → L ( T ∗ Y) , calledpull-back of 1-forms, such that ϕ ∗ (d f ) = d( f ◦ ϕ ) ∀ f ∈ S ( X ) (1.35) ϕ ∗ ( gω ) = g ◦ ϕ ϕ ∗ ω ∀ g ∈ L ∞ ( X ) , ω ∈ L ( T ∗ X ) , (1.36) and for such map it holds | ϕ ∗ ω | ≤ Lip( ϕ ) | ω | ◦ ϕ m Y -a.e. ∀ ω ∈ L ( T ∗ X ) . (1.37) proof For a simple form W = P i χ A i d f i ∈ L ( T ∗ X ) the requirements (1.35),(1.36) force thedefinition ϕ ∗ W := P i χ A i ◦ ϕ d( f i ◦ ϕ ). The inequality (cid:12)(cid:12) X i χ A i ◦ ϕ d( f i ◦ ϕ ) (cid:12)(cid:12) = X i χ A i ◦ ϕ | d( f i ◦ ϕ ) | (1.34) ≤ Lip( ϕ ) X i ( χ A i | d f i | ) ◦ ϕ = Lip( ϕ ) | W | ◦ ϕ shows that the definition of ϕ ∗ W is well-posed - i.e. it depends only on W and not on theway we write it as P i χ A i d f i - and that (1.37) holds for simple forms. In particular we have k ϕ ∗ W k L ( T ∗ Y) ≤ Lip( ϕ ) sZ | W | ◦ ϕ d m Y ≤ Lip( ϕ ) p Comp( ϕ ) k W k L ( T ∗ X ) , ∀ W simpleshowing that the map ϕ ∗ so defined is continuous from the space of simple 1-forms on X to L ( T ∗ Y). Hence it can be uniquely extended to a linear continuous map from L ( T ∗ X ) to L ( T ∗ Y), which clearly satisfies (1.37). Thus by construction we have (1.35) and (1.36) forsimple functions; the validity (1.36) for any g ∈ L ∞ ( X ) then follows by approximation. (cid:3) ϕ ◦ ψ ) ∗ = ψ ∗ ◦ ϕ ∗ We remark that given a map of bounded deformation ϕ : Y → X we have two (very) differentways of considering the pull-back of 1-forms: the one defined in the previous theorem, whichtakes values in L ( T ∗ Y), and the one in the sense of pull-back modules, which takes valuesin the pullback ϕ ∗ L ( T ∗ X ) of L ( T ∗ X ) via ϕ . To avoid confusion, we shall denote the lattermap by [ ϕ ∗ ] keeping the notation ϕ ∗ for the former.With this said, by duality we can now define the differential of a map of bounded defor-mation: Theorem/Definition 1.35 (Differential of a map of bounded deformation) . Let ϕ : Y → X be of bounded deformation and assume that L ( T X ) is separable. Then there exists a unique L ∞ (Y) -linear and continuous map d ϕ : L ( T Y) → ϕ ∗ L ( T X ) , called differential of ϕ , suchthat [ ϕ ∗ ω ] (cid:0) d ϕ ( v ) (cid:1) = ϕ ∗ ω ( v ) ∀ ω ∈ L ( T ∗ X ) , v ∈ L ( T Y) (1.38) and it satisfies | d ϕ ( v ) | ≤ Lip( ϕ ) | v | m Y -a.e. ∀ v ∈ L ( T Y) . (1.39) proof Let v ∈ L ( T Y) and consider the map L v : { ϕ ∗ ω : ω ∈ L ( T X ) } → L (Y) sending ϕ ∗ ω to ϕ ∗ ω ( v ). The bound (1.37) and the identity | ω | ◦ ϕ = | [ ϕ ∗ ] ω | give | L v ( ω ) | ≤ Lip( ϕ ) | [ ϕ ∗ ] ω || v | m Y -a.e. ∀ ω ∈ L ( T ∗ X ) . The vector space { ϕ ∗ ω : ω ∈ L ( T ∗ X ) } generates ϕ ∗ L ( T ∗ X ) and the dual of this module is- by Theorem 1.30 and the separability assumption on L ( T X ) - the module ϕ ∗ L ( T X ), thusby Proposition 1.16 we deduce that there is a unique element in ϕ ∗ L ( T X ), which we will calld ϕ ( v ), for which (1.38) holds and such d ϕ ( v ) also satisfies (1.39).It is clear that the assignment v d ϕ ( v ) is L ∞ (Y)-linear and since the bound (1.39) alsoensures that such assignment is continuous, the proof is completed. (cid:3) Remark 1.36. If ϕ is invertible with inverse of bounded compression, then Remark 1.28 tellsthat the pullback module ϕ ∗ L ( T X ) can be identified with L ( T X ) via the pullback map. Oncethis identification is done, the differential d ϕ can be seen as a map from L ( T Y) to L ( T X )and (1.38) reads as ω (d ϕ ( v )) = ϕ ∗ ω ( v ) ◦ ϕ − . (cid:4) We shall now relate the differential just built with the notion of ‘speed of a test plan’ asgiven by Theorem 1.32 to see that in our setting we have an analogous of the standard chainrule ( ϕ ◦ γ ) ′ t = d ϕ ( γ ′ t )valid in the smooth world.As before, let ϕ : Y → X be of bounded deformation, denote by ˆ ϕ the induced map from C ([0 , , Y) to C ([0 , , X ) and let π be a test plan on Y. For t ∈ [0 ,
1] let us also denote bye X t , e Y t the evaluation maps on C ([0 , , X ) and C ([0 , , Y) respectively.22otice that [(e Y t ) ∗ ]d ϕ : L ( T Y) → (e Y t ) ∗ ϕ ∗ L ( T X ) satisfies | [(e Y t ) ∗ ]d ϕ ( v ) | ≤ Lip( ϕ ) | v | ◦ e Y t and thus by the universal property of the pullback given in Proposition 1.26 we see thatthere is a unique L ∞ ( π )-linear and continuous map, which we shall denote by c d ϕ , from L ( T Y , e Y t , π ) to (e Y t ) ∗ ϕ ∗ L ( T X ) such that c d ϕ ([(e Y t ) ∗ ]( v )) = [(e Y t ) ∗ ]d ϕ ( v ) ∀ v ∈ L ( T Y) . We observe that for such map it holds (cid:0) [(e Y t ) ∗ ]( ϕ ∗ ω ) (cid:1) ( V ) = (cid:0) [(e Y t ) ∗ ][ ϕ ∗ ]( ω ) (cid:1)(cid:0) c d ϕ ( V ) (cid:1) ∀ ω ∈ L ( T ∗ X ) , V ∈ L ( T Y , e Y t , π ) , (1.40)indeed for V of the form (e Y t ) ∗ v for v ∈ L ( T Y) this is a direct consequence of the definingproperty and the conclusion for general V ’s follows from the fact that both sides of (1.40) are L ∞ ( π )-linear and continuous in V .With this said, we have the following result, proved in [19]: Proposition 1.37 (Chain rule for speeds) . Assume that L ( T X ) is separable. Then for a.e. t we have c d ϕ ( π ′ t ) = [ ˆ ϕ ∗ ]( ˆ ϕ ∗ π ) ′ t . (1.41) proof Both sides of (1.41) define elements of (e Y t ) ∗ ϕ ∗ L ( T X ) ∼ ˆ ϕ ∗ (e X t ) ∗ L ( T X ), where the‘ ∼ ’ comes from the functoriality of the pull-back (Remark 1.27) and ϕ ◦ e Y t = e X t ◦ ˆ ϕ . Since(e Y t ) ∗ ϕ ∗ L ( T X ) is the dual of (e Y t ) ∗ ϕ ∗ L ( T ∗ X ) (by the separability assumption and Theorem1.30), to prove (1.41) it is sufficient to test both sides against forms of the kind [(e Y t ) ∗ ][ ϕ ∗ ](d f )for f ∈ S ( X ), as they generate (e Y t ) ∗ ϕ ∗ L ( T ∗ X ) (recall Proposition 1.16).Thus let f ∈ S ( X ) and notice that for a.e. t we have[(e Y t ) ∗ ][ ϕ ∗ ](d f ) (cid:0) c d ϕ ( π ′ t ) (cid:1) = [(e Y t ) ∗ ]( ϕ ∗ d f )( π ′ t ) by (1.40)= [(e Y t ) ∗ ](d( f ◦ ϕ ))( π ′ t ) by (1.35)= L ( π ) − lim h → f ◦ ϕ ◦ e Y t + h − f ◦ ϕ ◦ e Y t h by definition of π ′ t = (cid:16) L ( ˆ ϕ ∗ π ) − lim h → f ◦ e X t + h − f ◦ e X t h (cid:17) ◦ ˆ ϕ because ϕ ◦ e Y t = e X t ◦ ˆ ϕ = [(e X t ) ∗ ](d f )( ˆ ϕ ∗ π ) ′ t ◦ ˆ ϕ by definition of ( ˆ ϕ ∗ π ) ′ t = (cid:0) [ ˆ ϕ ∗ ][(e X t ) ∗ ](d f ) (cid:1)(cid:0) [ ˆ ϕ ∗ ]( ˆ ϕ ∗ π ) ′ t (cid:1) by (1.25)= (cid:0) [(e Y t ) ∗ ][ ϕ ∗ ](d f ) (cid:1)(cid:0) [ ˆ ϕ ∗ ]( ˆ ϕ ∗ π ) ′ t (cid:1) because ϕ ◦ e Y t = e X t ◦ ˆ ϕ having also used Remark 1.27 in the last step. This is sufficient to conclude. (cid:3) Remark 1.38. If ϕ is invertible with inverse of bounded compression we know from Re-mark 1.36 that d ϕ can be seen as a map from L ( T Y) to L ( T X ), thus in this case the liftof its composition with (e X t ) ∗ to L ( T Y , e Y t , π ) provides a map c d ϕ from L ( T Y , e Y t , π ) to L ( T X , e X t , ˆ ϕ ∗ π ) and in this case (1.41) reads as c d ϕ ( π ′ t ) = ( ˆ ϕ ∗ π ) ′ t . (cid:4) .6 Infinitesimally Hilbertian spaces and Laplacian Definition 1.39 (Infinitesimally Hilbertian spaces) . ( X , d , m ) is said to be infinitesimallyHilbertian provided L ( T ∗ X ) (and thus also L ( T X ) ) is a Hilbert module. Remark 1.40.
Since f ( f, d f ) is an isometry of W , ( X ) into L ( X ) × L ( T ∗ X ) endowedwith the norm k ( f, ω ) k := k f k L + k ω k L ( T ∗ X ) , we see that if X is infinitesimally Hilbertian,then W , ( X ) is a Hilbert space.It is possible, although not entirely trivial, to show that also the converse implication holds,i.e. if W , ( X ) is Hilbert, then so is L ( T ∗ X ). In fact, the original definition of infinitesimallyHilbertian spaces given in [26] adopted such ‘ W , ’ approach, but the for the purpose of thisnote we preferred to start with the seemingly more powerful definition above. (cid:4) By Proposition 1.15 we know that L ( T ∗ X ) and L ( T X ) are isomorphic as L ∞ -modules.For f ∈ S ( X ), the image of d f under such isomorphism is called gradient of f and denotedby ∇ f . Directly from (1.14) and (1.15) it follows that ∇ ( ϕ ◦ f ) = ϕ ′ ◦ f ∇ f, ∀ f ∈ S ( X ) , ϕ ∈ LIP ∩ C ( R ) , ∇ ( f g ) = f ∇ g + g ∇ f ∀ f, g ∈ L ∞ ∩ S ( X ) . Remark 1.41.
Remark 1.40 and (1.8) grant that W , ( X ) is separable. Hence by Remark 1.9we see that L ( T ∗ X ), and thus also L ( T X ), is separable. Thus all the results of the previoussections are applicable. (cid:4) Notice also that both L ( T ∗ X ) and L ( T X ) are endowed with a pointwise scalar product. Definition 1.42 (Laplacian) . The space D (∆) is the space of all functions f ∈ W , ( X ) suchthat there is h ∈ L ( X ) for which Z hg d m = − Z h∇ f, ∇ g i d m ∀ g ∈ W , ( X ) . In this case the function h is called Laplacian of f and denoted by ∆ f . In other words, ∆ is the infinitesimal generator associated to (as well as the opposite ofthe subdifferential of) the Dirichlet form E ( f ) := Z | d f | d m , if f ∈ W , ( X ) , + ∞ , otherwise. (1.42)in particular is a closed operator and from the density of { E < ∞} = W , ( X ) in L ( X ) itfollows that D (∆) is dense in W , ( X ). It is also clear from the definitions that f ∈ D (∆) ⇔ ∇ f ∈ D (div) and in this case ∆ f = div( ∇ f ) , and thus recalling (1.24) we see thaton infinitesimally Hilbertian spaces the space D (div) is dense in L ( T X ) . (1.43)24he following calculus rules are also easily established:∆( ϕ ◦ f ) = ϕ ′ ◦ f ∆ f + ϕ ′′ ◦ f |∇ f | , ∀ f ∈ LIP b ( X ) ∩ D (∆) , ϕ ∈ C ( R ) (1.44)∆( f g ) = f ∆ g + g ∆ f + 2 h∇ f, ∇ g i ∀ f, g ∈ LIP b ( X ) ∩ D (∆) . (1.45)For instance, for the second notice that for h ∈ W , ( X ) and f, g as stated, we have f h, gh ∈ W , ( X ) and thus the claim follows from Z h∇ h, ∇ ( f g ) i d m = Z h∇ ( f h ) , ∇ g i + h∇ ( gh ) , ∇ f i − h h∇ f, ∇ g i d m . Remark 1.43.
In [41] a different construction of ‘ L E admitting a Carr´e du champ Γ. Adapting a bit the original presentation,the construction starts defining a symmetric bilinear map from [ L ∞ ( X ) ⊗ D ( E )] to L ( X ) byputting h f ⊗ g, f ′ ⊗ g ′ i := f f ′ Γ( g, g ′ ) ∀ f, f ′ ∈ L ∞ ( X ) , g, g ′ ∈ D ( E )and extending it by bilinearity. Then one defines the seminorm k · k on L ∞ ( X ) ⊗ D ( E ) byputting k ω k := Z h ω, ω i d m ∀ ω ∈ L ∞ ( X ) ⊗ D ( E ) , then passes to the quotient and finally to the completion. Calling M the resulting Banach spaceit is easy to check that it comes with the structure of a L -normed module, the pointwise normbeing given by | ω | := p h ω, ω i and the product with L ∞ -functions as (the linear continuousextension of) h · ( f ⊗ g ) := ( hf ) ⊗ g .In particular, the space of forms of the kind ⊗ g , for g ∈ D ( E ), generates M and it holds | ⊗ g | = p Γ( g, g ).In the case of infinitesimally Hilbertian spaces, the form E defined in (1.42) is a Dirichletform whose Carr´e du champ is given (thanks to (1.45)) by Γ( f, g ) = h∇ f, ∇ g i and in particularΓ( g, g ) = | d g | . This and Theorem 1.8 (and Remark 1.9) show that the cotangent module L ( T ∗ X ) and the space M coincide, meaning that the map sending d g to ⊗ g , for g ∈ W , ( X ) = D ( E ), uniquely extends to an isomorphism of modules. (cid:4) We conclude with a proposition (which concentrates results from [7], [8], [26] and [24])which is crucial in the application of this theory to the study of geometry of
RCD spaces:it provides an explicit differentiation formula along (appropriate) W -geodesics. Both thestatement and the proof rely on notions of optimal transport, see e.g [46], [4], [40] for anintroduction to the topic. Notice that the result can be read as a purely metric version of theBrenier-McCann theorem about optimal maps and W -geodesics. Theorem 1.44 (Derivation along geodesics) . Let ( X , d , m ) be an infinitesimally Hilbertianspace and t µ t = ρ t m ⊂ P ( X ) a W -geodesic made os measures with uniformly boundedsupports and densities. Assume also that for some, and thus any, p ∈ [1 , ∞ ) , the map t ρ t ∈ L p ( m ) is continuous.Then for every f ∈ W , ( X ) the map t R f d µ t is C ([0 , and the formula dd t Z f d µ t = − Z h∇ f, ∇ ϕ t i d µ t , ∀ t ∈ [0 , , (1.46) where ϕ t is, for every t ∈ [0 , , Lipschitz and such that for some s = t the function ( s − t ) ϕ is a Kantorovich potential from µ t to µ s . RCD ( K, ∞ ) spaces every W -geodesic such that µ , µ have both bounded den-sities and support satisfy the assumptions (see [39]). Sketch of the proof
Step 1
Let ϕ be a Lipschitz Kantorovich potential from µ to µ and let π be a lifting of( µ t ), i.e. so that (e t ) ∗ π = µ t for every t ∈ [0 , π is concentrated on geodesics and (e , e ) ∗ π is an optimal plan. We claim thatlim t → Z ϕ ( γ ) − ϕ ( γ t ) t d π ( γ ) ≥ Z | d ϕ | d µ + 12 W ( µ , µ ) . (1.47)To see this, start noticing that γ ∈ ∂ c ϕ ( γ ) for π -a.e. γ and thus for π -a.e. γ we have ϕ ( z ) − ϕ ( γ ) ≤ d ( z, γ )2 − d ( γ , γ )2 ≤ d ( z, γ ) d ( z, γ ) + d ( γ , γ )2 , taking the positive part, dividing by d ( z, γ ) and letting z → γ we obtain | d ϕ | ( γ ) ≤ lim z → γ ( ϕ ( z ) − ϕ ( γ )) + d ( z, γ ) ≤ d ( γ , γ ) π -a.e. γ, (1.48)where the first inequality is an easy consequence of the definition of minimal weak uppergradient and the fact that ϕ is Lipschitz. On the other hand, still from γ ∈ ∂ c ϕ ( γ ) for π -a.e. γ we have ϕ ( γ ) − ϕ ( γ t ) ≥ d ( γ , γ )2 − d ( γ t , γ )2 = d ( γ , γ )( t − t / ∀ t ∈ (0 , π -a.e. γ. Thus lim t → Z ϕ ( γ ) − ϕ ( γ t ) t d π ( γ ) ≥ Z lim t → ϕ ( γ ) − ϕ ( γ t ) t d π ( γ ) ≥ Z d ( γ , γ ) d π ( γ )and since R d ( γ , γ ) d π ( γ ) = W ( µ , µ ), this inequality and (1.48) give (1.47). Step 2
Let π as before, notice that it is a test plan and let f ∈ W , ( X ). Then Z f ( γ t ) − f ( γ ) t d π ( γ ) ≤ t Z Z t | d f | ( γ s ) | ˙ γ s | d s d π ( γ ) ≤ t Z Z t | d f | ρ s d s d m + 12 W ( µ , µ ) , passing to the limit noticing that ( ρ t ) ⊂ L ∞ is weakly ∗ -continuous we conclude thatlim t → Z f ( γ t ) − f ( γ ) t d π ( γ ) ≤ Z | d f | d µ + 12 W ( µ , µ ) . Write this inequality with εf − ϕ in place of f and subtract (1.47) to deduce thatlim t → ε Z f ( γ t ) − f ( γ ) t d π ( γ ) ≤ Z | d( εf − ϕ ) | − | d ϕ | d µ . Dividing by ε > ε <
0) and letting ε ↓ ε ↑
0) and noticing that | d( εf − ϕ ) | −| d ϕ | ε = − h∇ f, ∇ ϕ i + ε | d f | we conclude thatdd t Z f d µ t | t =0 = − Z h∇ f, ∇ ϕ i d µ . (1.49)26 tep 3 By rescaling, we see from (1.49) that formula (1.46) holds for any t , so that to concludeit remains to prove that the right hand side is continuous in t . Notice also that we are freein the choice of the (rescaled) Kantorovich potentials in (1.46) and thus we may assume thatthey are equiLipschitz. Then since uniform limits of Kantorovich potentials are Kantorovichpotentials, it is easy to see that to conclude it is sufficient to prove that for t n → t and ( ϕ t n )uniformly Lipschitz and uniformly converging to some ϕ t we havelim n →∞ Z h∇ f, ∇ ϕ t n i ρ t n d m = Z h∇ f, ∇ ϕ t i ρ t d m . Since the ρ t ’s have uniformly bounded support, up to multiplying the ϕ ’s by an appropriatecut-off we can assume that the ϕ ’s are bounded in W , ( X ) and thus that the convergenceof ( ϕ t n ) to ϕ is weak in W , ( X ). Thus ( ∇ ϕ n ) weakly converges to ∇ ϕ t in L ( T X ) and, bythe assumptions on ρ t , ( ρ t n ∇ f ) strongly converges to ρ t ∇ f in L ( T X ). The thesis follows. (cid:3) Remark 1.45.
In connection with Theorem 1.32, the proof of this last proposition can beused to show that for π as in the proof, the vector fields π ′ t are defined for every t (and notjust for a.e. t ) and are given by π ′ t = e ∗ t ( ∇ ϕ t ) . This follows noticing that for A ⊂ C ([0 , , X ) Borel with π ( A ) >
0, the plan π A :=( π ( A )) − π | A is still a test plan and the curve t (e t ) ∗ π A still satisfies the assumptionswith the same functions ϕ ’s. (cid:4) RCD spaces
RCD spaces
From now on, we shall always assume that our space satisfies the Riemannian CurvatureDimension condition
RCD ( K, ∞ ), the definition being ([9]): Definition 2.1 ( RCD ( K, ∞ ) spaces) . Let K ∈ R . ( X , d , m ) is a RCD ( K, ∞ ) space provided:i) it is infinitesimally Hilbertianii) for some C > and x ∈ X it holds m ( B r ( x )) ≤ e Cr for every r > iii) every f ∈ W , ( X ) with | d f | ∈ L ∞ ( X ) admits a Lipschitz representative ˜ f with Lip( ˜ f ) ≤k| d f |k L ∞ iv) for every f ∈ D (∆) with ∆ f ∈ W , ( X ) and g ∈ L ∞ ( X ) ∩ D (∆) with g ≥ , ∆ g ∈ L ∞ ( X ) ,it holds the Bochner inequality: Z | d f | ∆ g d m ≥ Z g (cid:0) h∇ f, ∇ ∆ f i + K | d f | (cid:1) d m (2.1)In some sense the ‘truly defining’ properties are ( i ) and ( iv ), while ( ii ) , ( iii ) are more of atechnical nature: ( ii ) is necessary to ensure a priori that the heat flow - see below - preservesthe mass, while ( iii ) to grant that Sobolev functions determine the metric of the space (noticethat there are doubling spaces supporting a Poincar´e inequality for which ( iii ) fails).27he heat flow ( h t ) on X is the gradient flow of (=the flow associated to) the Dirichletform E , i.e. for f ∈ L ( X ) the map t h t f ∈ L ( X ) is the only continuous curve on [0 , ∞ )which is absolutely continuous on (0 , ∞ ) and such that h f = f anddd t h t f = ∆ h t f a.e. t > . It is possible to check, we omit the details, that the heat flow satisfies the weak maximumprinciple f ≤ C m -a.e. ⇒ h t f ≤ C m -a.e. ∀ t ≥ L + L ∞ ( X ). Then from (2.1) one gets the following important Bakry-´Emery estimate : for every f ∈ W , ( X ) and t ≥ | d h t f | ≤ e − Kt h t ( | d f | ) m -a.e. . (2.2)Formally, this comes noticing that the derivative of [0 , t ] ∋ s F ( s ) := h t − s ( | d h s f | ) is givenby h t − s (cid:16) − ∆( | d h s f | ) + 2 h∇ h s f, ∇ ∆ h s f i (cid:17) and this is ≤ − KF ( s ) by the Bochner inequality (2.1). Then one concludes with the Gron-wall’s Lemma.We shall also make use of the L ∞ − Lip regularization: for f ∈ L ∞ ( X ) and t > h t f ∈ LIP( X ) with s Z t e Ks d s Lip( h t f ) ≤ k f k L ∞ . (2.3)This, again formally, follows integrating in s ∈ [0 , t ] the bounddd s h s ( | h t − s f | ) = h s (cid:16) ∆ | h t − s f | − h t − s f ∆ h t − s f (cid:17) (1.45) = 2 h s ( | d h t − s f | ) (2.2) ≥ e Ks | d h t f | , then using the weak maximum principle and Property ( iii ) in the definition of RCD spaces.
A key tool that we shall use to develop second order calculus on
RCD spaces is the notion of‘test function’ introduced in [42]:Test( X ) := n f bounded, Lipschitz, in D (∆) with ∆ f ∈ W , ( X ) o . From (2.3) and general regularization properties of the heat flow we have that f ∈ L ∩ L ∞ ( X ) , f ≥ ⇒ h t f ∈ Test( X ) , h t f ≥ ∀ t > X ) is dense in W , ( X ). To analyze the properties of testfunctions it is useful to introduce the following notion, coming from [26]: Definition 2.2 (Measure-valued Laplacian) . Let f ∈ W , ( X ) . We say that f has a measure-valued Laplacian, and write f ∈ D ( ∆ ) , provided there exists a Borel measure µ on X finiteon bounded sets such that Z g d µ = − Z h∇ f, ∇ g i d m for every g ∈ LIP( X ) with bounded support.In this case the measure µ , which is clearly unique, will be denoted by ∆ f .
28t is readily verified that this concept is fully compatible with the one given in Definition1.42, in the sense that f ∈ D (∆) ⇔ f ∈ D ( ∆ ) with ∆ f ≪ m and d ∆ f d m ∈ L ( X ), and in this case ∆ f = ∆ f m , and one can check that f ∈ D ( ∆ ) , | d f | ∈ L ( X ) ⇒ ∆ f ( X ) = 0 (2.4)(this is trivial if m ( X ) < ∞ , for the general case one approximates the constant 1 withfunctions with uniformly bounded Laplacian).We then have the following crucial property, proved in [42], which is the first crucial steptowards second-order calculus in RCD spaces: among others, it provides Sobolev regularityfor | d f | for any f ∈ Test( X ) (in contrast, without any lower Ricci bound it seems impossibleto exhibit non-constant functions f for which | d f | has any kind of regularity). Theorem 2.3.
Let f ∈ Test( X ) . Then | d f | ∈ D ( ∆ ) ⊂ W , ( X ) and ∆ | d f | ≥ (cid:0) h∇ f, ∇ ∆ f i + K | d f | (cid:1) m . (2.5) Sketch of the proof
From the fact that | d f | , h∇ f, ∇ ∆ f i + K | d f | ∈ L ( X ) one can checkthat (2.1) holds for any g ∈ D (∆) non-negative. Picking g := h t ( | d f | ) we obtain Z | d h t/ ( | d f | ) | d m = − Z | d f | ∆ h t ( | d f | ) d m (2.1) ≤ − Z h t ( | d f | ) (cid:0) h∇ f, ∇ ∆ f i + K | d f | (cid:1) m ≤ k| d f | k L ∞ Z (cid:12)(cid:12) h∇ f, ∇ ∆ f i + K | d f | (cid:12)(cid:12) m , so that letting t ↓ | d f | ∈ W , ( X ). Now, at least if X is compact, | d f | ∈ D ( ∆ ) and (2.5) both follow noticing that from (2.1) we have that the linear operator C ( X ) ∩ D (∆) ∋ g L ( g ) := Z ∆ g | d f | − g (cid:0) h∇ f, ∇ ∆ f i + K | d f | (cid:1) d m is such that L ( g ) ≥ g ≥
0. Hence it must coincide with the integral of g w.r.t. anon-negative measure. (cid:3) A direct, and important, property that follows from the above is thatTest( X ) is an algebra.Indeed, in checking that f g ∈ Test( X ) for f, g ∈ Test( X ) the only non-trivial thing to prove isthat ∆( f g ) ∈ W , ( X ). Since it is clear that f ∆ g, g ∆ f ∈ W , ( X ), by the Leibniz rule for theLaplacian (1.45) to conclude it is sufficient to show that h∇ f, ∇ g i ∈ W , ( X ). This follows bypolarization from Theorem 2.3. 29 .3 The space W , ( X ) Let H , H be two Hilbert modules on X and denote by H ⊗ Alg H their tensor products as L ∞ -modules, so that H ⊗ Alg H can be seen as the space of formal finite sums of objects ofthe kind v ⊗ v with ( v , v ) v ⊗ v being L ∞ -bilinear.We define the L ∞ -bilinear and symmetric map : from [ H ⊗ Alg H ] to L ( X ) by putting( v ⊗ v ) : ( v ′ ⊗ v ′ ) := (cid:10) v , v ′ (cid:11) (cid:10) v , v ′ (cid:11) where h· , ·i i is the pointwise scalar product on H i , i = 1 ,
2, and extending it by L ∞ -bilinearity.It is readily verified that this definition is well posed and that the resulting map is positivelydefinite in the sense that for any A ∈ H ⊗ Alg H and E ⊂ X Borel it holds A : A ≥ m -a.e. A : A = 0 m -a.e. on E if and only if A = 0 m -a.e. on E. Then define the
Hilbert-Schimdt pointwise norm as | A | HS := √ A : A ∈ L ( X )and the tensor product norm as k A k H ⊗ H := sZ | A | d m ∈ [0 , + ∞ ] . We are now ready to give the following definition:
Definition 2.4 (Tensor product of Hilbert modules) . The space H ⊗ H is defined as thecompletion of n A ∈ H ⊗ Alg H : k A k H ⊗ H < ∞ o w.r.t. the tensor product norm k · k H ⊗ H . The multiplication by L ∞ functions in H ⊗ Alg H is easily seen to induce by continuitya multiplication by L ∞ -functions on H ⊗ H which together with the pointwise norm | · | HS show that H ⊗ H comes with the structure of L -normed module. Moreover, since | · | HS satisfies the pointwise parallelogram identity, H ⊗ H is in fact a Hilbert module.If H = H , the tensor product will be denoted H ⊗ . In this case the map v ⊗ v v ⊗ v on H ⊗ Alg H induces an automorphism A A t , called transposition, on H ⊗ and for ageneric A ∈ H ⊗ we put A Sym := A + A t A Asym := A − A t A , respectively. It is then clear that | A | = | A Sym | + | A Asym | m -a.e. ∀ A ∈ H ⊗ . (2.6)We shall write L (( T ∗ ) ⊗ X ) (resp. L ( T ⊗ X )) for the tensor product of L ( T ∗ X ) (resp. L ( T X )) with itself. These modules are one the dual of the other and we shall typicallywrite A ( X, Y ) in place of A ( X ⊗ Y ) for A ∈ L (( T ∗ ) ⊗ X ) and X ⊗ Y ∈ L ( T ⊗ X ).Notice that being L ( T ∗ X ) separable (Remark 1.41), so is L (( T ∗ ) ⊗ X ). Same for L ( T ⊗ X ).30 .3.2 Definition of W , ( X )Recall that on a smooth Riemannian manifold, the Hessian of the smooth function f ischaracterized by the validity of the identity2Hess( f )( ∇ g , ∇ g ) = h∇ ( h∇ f, ∇ g i ) , ∇ g i + h∇ ( h∇ f, ∇ g i ) , ∇ g i − h∇ f, ∇ ( h∇ g , ∇ g i ) i for any smooth functions g , g . This motivates the following definition: Definition 2.5 (The space W , ( X ) and the Hessian) . The space W , ( X ) is the set of allthe functions f ∈ W , ( X ) for which there exists A ∈ L (( T ∗ ) ⊗ X ) such that Z hA ( ∇ g , ∇ g ) d m = − Z h∇ f, ∇ g i div( h ∇ g ) + h∇ f, ∇ g i div( h ∇ g )+ h h∇ f, ∇ h∇ g , ∇ g ii d m (2.7) for every g , g ∈ Test( X ) and h ∈ LIP b ( X ) . Such A will be called Hessian of f and denotedby Hess f . The space W , ( X ) is equipped with the norm k f k W , ( X ) := k f k L ( X ) + k d f k L ( T ∗ X ) + k Hess f k L (( T ∗ ) ⊗ X ) . From the density of Test( X ) in W , ( X ) is easily follows that the Hessian, if it exists, isunique and thus in particular the W , -norm is well defined. Notice that in giving the abovedefinition we used in a crucial way Theorem 2.3 to grant that h∇ g , ∇ g i ∈ W , ( X ) so thatthe last addend in the integral in (2.7) is well defined.The following is easily verified: Theorem 2.6.
We have:i) W , ( X ) is a separable Hilbert space.ii) The Hessian is a closed operator, i.e. the set { ( f, Hess( f )) : f ∈ W , ( X ) } is a closedsubset of W , ( X ) × L (( T ∗ ) ⊗ X ) iii) For every f ∈ W , ( X ) the Hessian Hess( f ) is symmetric, i.e. Hess( f ) t = Hess( f ) .proof For given g , g , h ∈ Test( X ) the left (resp. right) hand side of (2.7) is continuous w.r.t. A ∈ L (( T ∗ ) ⊗ X ) (resp. f ∈ W , ( X )). Point ( ii ) and the completeness of W , follow. Thefact that the W , -norm satisfies the parallelogram rule is obvious. For the separability, noticethat L ( X ) × L ( T ∗ X ) × L (( T ∗ ) ⊗ X ) endowed with its natural Hilbert structure is separableand that the map W , ( X ) ∋ f ( f, d f, Hess f ) ∈ L ( X ) × L ( T ∗ X ) × L (( T ∗ ) ⊗ X )is an isometry. Point ( iii ) comes from the symmetry in g , g of (2.7). (cid:3) Remark 2.7.
As the example of weighted Riemannian manifold shows, in general the Lapla-cian is not the trace of the Hessian. (cid:4) .3.3 Existence of W , functions It is not at all obvious that W , ( X ) contains any non-constant function. This (and muchmore) is ensured by the following crucial Lemma which is about the self-improving of Bochnerinequality. Read in the smooth setting, the claim says that for the vector field X := P i g i ∇ f i and the 2-tensor A := P j ∇ h j ⊗ ∇ h j it holds |∇ X : A | ≤ (cid:16) ∆ | X | h X, ∆ H X i − K | X | − | ( ∇ X ) Asym | (cid:17) | A | , (2.8)see also Lemma 2.33. Given that for the moment we don’t have the covariant derivative andthe Hodge Laplacian, we have to state (2.8) by ‘unwrapping’ these operators.From now on, we shall denote by Meas ( X ) the space of finite Borel measures on X equippedwith the total variation norm. Then for f, g, h ∈ Test( X ) it will be useful to introduce Γ ( f, g ) ∈ Meas ( X ) and H [ f ]( g, h ) ∈ L ( X ) as Γ ( f, g ) := 12 (cid:16) ∆ ( h∇ f, ∇ g i ) − (cid:0) h∇ f, ∇ ∆ g i + h∇ g, ∇ ∆ f i (cid:1) m (cid:17) H [ f ]( g, h ) := 12 (cid:16) h∇ ( h∇ f, ∇ g i ) , ∇ h i + h∇ ( h∇ f, ∇ h i ) , ∇ g i − h∇ f, ∇ ( h∇ g, ∇ h i ) i (cid:17) We shall also write Γ ( f, g ) = γ ( f, g ) m + Γ s ( f, g ) , with Γ s ( f, g ) ⊥ m . We then have the following:
Lemma 2.8 (Key inequality) . Let n, m ∈ N and f i , g i , h j ∈ Test( X ) , i = 1 , . . . , n , j =1 , . . . , m . Define the measure µ = µ (cid:0) ( f i ) , ( g i ) (cid:1) ∈ Meas ( X ) as µ (cid:0) ( f i ) , ( g i ) (cid:1) := X i,i ′ g i g i ′ (cid:0) Γ ( f i , f i ′ ) − K h∇ f i , ∇ f i ′ i m (cid:1) + (cid:16) g i H [ f i ]( f i ′ , g i ′ ) + h∇ f i , ∇ f i ′ i h∇ g i , ∇ g i ′ i + h∇ f i , ∇ g i ′ i h∇ g i , ∇ f i ′ i (cid:17) m and write it as µ = ρ m + µ s with µ s ⊥ m .Then µ s ≥ and (cid:12)(cid:12)(cid:12)(cid:12) X i,j h∇ f i , ∇ h j i h∇ g i , ∇ h j i + g i H [ f i ]( h j , h j ) (cid:12)(cid:12)(cid:12)(cid:12) ≤ ρ X j,j ′ | (cid:10) ∇ h j , ∇ h j ′ (cid:11) | . (2.10) Sketch of the proof
We shall prove the thesis in the simplified case n = m = 1 and g ≡ RCD ( K, ∞ ) spaces in [42]): in this casethe measure µ is given by µ = Γ ( f, f ) − K h∇ f, ∇ f i m . Then (2.9) follows from (2.5) and(2.10) reads as (cid:12)(cid:12) H [ f ]( h, h ) (cid:12)(cid:12) ≤ (cid:16) γ ( f, f ) − K h∇ f, ∇ f i (cid:17) |∇ h | . (2.11)For λ, c ∈ R define Φ λ,c = Φ λ,c ( f, h ) := λf + h − ch ∈ Test( X ). It is only a matter ofcomputations to check that γ (Φ λ,c , Φ λ,c ) − K |∇ Φ λ,c | = λ (cid:0) γ ( f, f ) − K |∇ f | (cid:1) + 4 λH [ f ]( h, h ) + 4 |∇ h | + ( h − c ) F λ,c F λ,c ∈ L ( X , m ) so that c F λ,c ∈ L ( X , m ) is continuous. It follows that m -a.e.the inequality γ (Φ λ,c , Φ λ,c ) − K |∇ Φ λ,c | ≥ c ∈ R .Hence for m -a.e. x we can take c = h ( x ) and conclude that λ (cid:0) γ ( f, f ) − K |∇ f | (cid:1) + 4 λH [ f ]( h, h ) + 4 |∇ h | ≥ m -a.e.and (2.11) follows by the arbitrariness of λ ∈ R .The general case follows by a similar optimization argument using Φ( f i , g i , h j ) in place ofΦ( f, h ) for Φ given byΦ( x , . . . , x n , y , . . . , y n , z , . . . , z m ) := X i ( λx i y i + a i x i − b i y i ) + X j z j − c j z j , we omit the details. (cid:3) The first important consequence of this lemma is the following result, which shows inparticular that W , ( X ) is dense in W , ( X ). Theorem 2.9.
Let f ∈ Test( X ) . Then f ∈ W , ( X ) and | Hess f | ≤ γ ( f, f ) − K |∇ f | , m -a.e. , (2.12) and moreover for every g , g ∈ Test( X ) it holds H [ f ]( g , g ) = Hess f ( ∇ g , ∇ g ) , m -a.e. . (2.13) proof We apply Lemma (2.8) with n = 1 for given functions f, h j ∈ Test( X ), j = 1 , . . . , m and g ≡ m ( X ) < ∞ , in the general case an approximation argumentis required). In this case inequality (2.10) reads, also recalling the definition of pointwise normon L ( T ⊗ X ), as: (cid:12)(cid:12)(cid:12)(cid:12) X j H [ f ]( h j , h j ) (cid:12)(cid:12)(cid:12)(cid:12) ≤ p γ ( f, f ) − K |∇ f | (cid:12)(cid:12)(cid:12) X j ∇ h j ⊗ ∇ h j (cid:12)(cid:12)(cid:12) HS , m -a.e. . (2.14)Now notice that for arbitrary h j , h ′ j ∈ Test( X ), g j ∈ LIP b ( X ) we have g j H [ f ]( h j , h ′ j ) = 12 g j (cid:16) H [ f ]( h j + h ′ j , h j + h ′ j ) − H [ f ]( h j , h j ) − H [ f ]( h ′ j , h ′ j ) (cid:17) g j ∇ h j ⊗ ∇ h ′ j + ∇ h ′ j ⊗ ∇ h j g j ∇ ( h j + h ′ j ) ⊗ ∇ ( h j + h ′ j ) − ∇ h j ⊗ ∇ h j − ∇ h ′ j ⊗ ∇ h ′ j , hence taking into account the trivial inequality | A Sym | HS ≤ | A | HS m -a.e. (recall (2.6)) for A := P j g j ∇ h j ⊗ ∇ h ′ j , from (2.14) we obtain (cid:12)(cid:12)(cid:12) X j g j H [ f ]( h j , h ′ j ) (cid:12)(cid:12)(cid:12) ≤ p γ ( f, f ) − K |∇ f | (cid:12)(cid:12)(cid:12)(cid:12) X j g j ∇ h j ⊗ ∇ h ′ j + ∇ h ′ j ⊗ ∇ h j (cid:12)(cid:12)(cid:12)(cid:12) HS ≤ p γ ( f, f ) − K |∇ f | (cid:12)(cid:12)(cid:12) X j g j ∇ h j ⊗ ∇ h ′ j (cid:12)(cid:12)(cid:12) HS . (2.15)33ow let V ⊂ L ( T ⊗ X ) be the space of linear combinations of tensors of the form g ∇ h ⊗ ∇ h ′ for h, h ′ ∈ Test( X ), g ∈ LIP b ( X ) and define A : V → L ( X ) as A (cid:16) X j g j ∇ h j ⊗ ∇ h ′ j (cid:17) := X j g j H [ f ]( h j , h ′ j ) . From (2.15) we see that this is a good definition, i.e. that A ( T ) depends only on T . Moreover,recalling that by (2.9) we have Γ s ( f, f ) ≥
0, we obtain Z γ ( f, f ) − K |∇ f | d m ≤ Γ ( f, f )( X ) − K Z |∇ f | d m (2.4) = Z (∆ f ) − K |∇ f | d m (2.16)hence from (2.15) we deduce that k A ( T ) k L ( X ) ≤ sZ (∆ f ) − K |∇ f | d m k T k L ( T ⊗ X ) , ∀ T ∈ V. It is readily verified that V is dense in L ( T ⊗ X ), therefore A can be uniquely extended to acontinuous linear operator from L ( T ⊗ X ) to L ( X ) which is readily checked to be L ∞ -linear.In other words, A ∈ L (( T ∗ ) ⊗ X ).Now let h , h ∈ Test( X ), g ∈ LIP b ( X ) be arbitrary and notice that we have Z A ( g ∇ h ⊗ ∇ h ) d m = 2 Z gH [ f ]( h , h ) d m and, by the definition of H [ f ] and after an integration by parts, that2 Z gH [ f ]( h , h ) d m = Z − h∇ f, ∇ h i div( g ∇ h ) − h∇ f, ∇ h i div( g ∇ h ) − g (cid:10) ∇ f, ∇ h∇ h , ∇ h i (cid:11) d m . These show that f ∈ W , ( X ) with Hess f = A and that (2.13) holds. For (2.12) notice that(2.15) can be restated as | Hess f ( T ) | ≤ p γ ( f, f ) − K |∇ f | | T | HS , ∀ T ∈ V, and use once again the density of V in L ( T ⊗ X ) to conclude. (cid:3) In particular, we have the following important corollary:
Corollary 2.10.
We have D (∆) ⊂ W , ( X ) and Z | Hess f | d m ≤ Z (∆ f ) − K |∇ f | d m , ∀ f ∈ D (∆) . (2.17) Sketch of the proof
For f ∈ Test( X ) the claim follows integrating (2.12) and recalling (2.16).The general case is then achieved by approximation recalling that the Hessian is a closedoperator. (cid:3) Such corollary ensures that the following definition is meaningful:
Definition 2.11.
We define H , ( X ) as the W , -closure of D (∆) ⊂ W , ( X ) . It is not hard to check that H , ( X ) also coincides with the W , ( X ) closure of Test( X );on the other hand it is important to underline that it is not at all clear whether H , ( X )coincides with W , ( X ) or not. 34 .3.4 Calculus rulesProposition 2.12 (Product rule for functions) . Let f , f ∈ LIP b ∩ W , ( X ) . Then f f ∈ W , ( X ) and the formula Hess( f f ) = f Hess f + f Hess f + d f ⊗ d f + d f ⊗ d f , m -a.e. (2.18) holds.proof It is obvious that f f ∈ W , ( X ) and that the right hand side of (2.18) defines anobject in L (( T ∗ ) ⊗ X ). Now let g , g ∈ Test( X ), h ∈ LIP b ( X ) be arbitrary and notice that − h∇ ( f f ) , ∇ g i div( h ∇ g ) = − f h∇ f , ∇ g i div( h ∇ g ) − f h∇ f , ∇ g i div( h ∇ g )= − h∇ f , ∇ g i div( f h ∇ g ) + h h∇ f , ∇ g i h∇ f , ∇ g i− h∇ f , ∇ g i div( f h ∇ g ) + h h∇ f , ∇ g i h∇ f , ∇ g i . Exchanging the roles of g , g , noticing that − h (cid:10) ∇ ( f f ) , ∇ h∇ g , ∇ g i (cid:11) = − hf (cid:10) ∇ f , ∇ h∇ g , ∇ g i (cid:11) − hf (cid:10) ∇ f , ∇ h∇ g , ∇ g i (cid:11) , adding everything up, integrating and observing that f h, f h ∈ LIP b ( X ) we conclude. (cid:3) Proposition 2.13 (Chain rule) . Let f ∈ LIP ∩ W , ( X ) and ϕ : R → R a C function withuniformly bounded first and second derivative (and ϕ (0) = 0 if m ( X ) = + ∞ ).Then ϕ ◦ f ∈ W , ( X ) and the formula Hess( ϕ ◦ f ) = ϕ ′′ ◦ f d f ⊗ d f + ϕ ′ ◦ f Hess f, m -a.e. (2.19) holds.proof It is obvious that ϕ ◦ f ∈ W , ( X ) and that the right hand side of (2.19) defines anobject in L (( T ∗ ) ⊗ X ). Now let g , g ∈ Test( X ), h ∈ LIP b ( X ) be arbitrary and notice that − h∇ ( ϕ ◦ f ) , ∇ g i div( h ∇ g ) = − ϕ ′ ◦ f h∇ f, ∇ g i div( h ∇ g )= − h∇ f, ∇ g i div( ϕ ′ ◦ f h ∇ g ) + hϕ ′′ ◦ f h∇ f, ∇ g i h∇ f, ∇ g i . Similarly, − h∇ ( ϕ ◦ f ) , ∇ g i div( h ∇ g ) = − h∇ f, ∇ g i div( ϕ ′ ◦ f h ∇ g ) + hϕ ′′ ◦ f h∇ f, ∇ g i h∇ f, ∇ g i and − h (cid:10) ∇ ( ϕ ◦ f ) , ∇ h∇ g , ∇ g i (cid:11) = − hϕ ′ ◦ f (cid:10) ∇ f, ∇ h∇ g , ∇ g i (cid:11) . To conclude, add up these three identities, integrate and notice that hϕ ′ ◦ f ∈ LIP b ( X ). (cid:3) Proposition 2.14 (Product rule for gradients) . Let f , f ∈ LIP ∩ H , ( X ) . Then h∇ f , ∇ f i ∈ W , ( X ) and d h∇ f , ∇ f i = Hess f ( ∇ f , · ) + Hess f ( ∇ f , · ) , m -a.e. . (2.20)35 ketch of the proof For f , f ∈ Test( X ) the fact that h∇ f , ∇ f i ∈ W , ( X ) follows fromTheorem 2.3 by polarization. Also, by the very definition of H [ f ], we know that for any g ∈ Test( X ) it holds hh∇ f , ∇ f i , ∇ g i = H [ f ]( f , g ) + H [ f ]( f , g ) , hence in this case the conclusion comes from (2.13) and the arbitrariness of g . The generalcase follows by approximation by observing that with an argument based on truncation andregularization with the heat flow, we can approximate any f ∈ LIP ∩ H , ( X ) in the H , ( X )-topology with test functions which are uniformly Lipschitz. (cid:3) The definition of Sobolev vector fields is based on the identity h∇ ∇ g X, ∇ g i = h∇ ( h X, ∇ g i ) , ∇ g i − Hess( g )( X, ∇ g ) , valid in the smooth world for smooth functions g , g and a smooth vector field X . Definition 2.15 (The Sobolev space W , C ( T X )) . The Sobolev space W , C ( T X ) ⊂ L ( T X ) is the space of all X ∈ L ( T X ) for which there exists T ∈ L ( T ⊗ X ) such that for every g , g ∈ Test( X ) and h ∈ LIP b ( X ) it holds Z h T : ( ∇ g ⊗ ∇ g ) d m = Z − h X, ∇ g i div( h ∇ g ) − h Hess( g )( X, ∇ g ) d m . In this case we shall call the tensor T the covariant derivative of X and denote it by ∇ X .We endow W , C ( T X ) with the norm k · k W , C ( T X ) defined by k X k W , C ( T X ) := k X k L ( T X ) + k∇ X k L ( T ⊗ X ) . It will be useful to introduce the space of ‘test vector fields’ asTestV( X ) := n n X i =1 g i ∇ f i : n ∈ N , f i , g i ∈ Test( X ) o ⊂ L ( T X ) . It is easy to show that TestV( X ) is dense in L ( T X ). Theorem 2.16 (Basic properties of W , C ( T X )) . We have:i) W , C ( T X ) is a separable Hilbert space.ii) The covariant derivative is a closed operator, i.e. the set { ( X, ∇ X ) : X ∈ W , C ( T X ) } isa closed subset of L ( T X ) × L ( T ⊗ X ) .iii) Given f ∈ W , ( X ) we have ∇ f ∈ W , C ( T X ) with ∇ ( ∇ f ) = (Hess f ) ♯ , where · ♯ : L (( T ∗ ) ⊗ X ) → L ( T X ) is the Riesz (musical) isomorphims. v) We have TestV( X ) ⊂ W , C ( T X ) with ∇ X = X i ∇ g i ⊗ ∇ f i + g i (Hess f i ) ♯ , for X = X i g i ∇ f i . In particular, W , C ( T X ) is dense in L ( T X ) .Sketch of the proof ( i ) , ( ii ) are proved along the same lines of Theorem 2.6. ( iii ) follows fromProposition 2.14 and direct verification; then ( iv ) follows from ( iii ) and the definitions. (cid:3) We know that TestV( X ) is contained in W , C ( T X ), but not if it is dense. Thus the followingdefinition is meaningful: Definition 2.17.
We define H , C ( T X ) ⊂ W , C ( T X ) as the W , C ( T X ) -closure of TestV( X ) . We shall also denote by L ( T X ) the L -completion of L ( T X ) (Theorem 1.7) and by L ∞ ( T X ) its subspace made of X ’s such that | X | ∈ L ∞ ( X ). Proposition 2.18 (Leibniz rule) . Let X ∈ L ∞ ∩ W , C ( T X ) and f ∈ L ∞ ∩ W , ( X ) .Then f X ∈ W , C ( T X ) and ∇ ( f X ) = ∇ f ⊗ X + f ∇ X, m -a.e. . (2.21) proof Assume for the moment that f ∈ Test( X ) and let g , g ∈ Test( X ), h ∈ LIP b ( X ) bearbitrary. Then f h ∈ LIP b ( X ) and from the definition of ∇ X we see that Z f h ∇ X : ( ∇ g ⊗ ∇ g ) d m = Z − h X, ∇ g i div( f h ∇ g ) − f h Hess g ( X, ∇ g ) d m . Using the identity div( f h ∇ g ) = h h∇ f, ∇ g i + f div( h ∇ g ) (recall (1.24)), this gives Z h h∇ f, ∇ g i h X, ∇ g i + f h ∇ X : ( ∇ g ⊗ ∇ g ) d m = Z − h f X, ∇ g i div( h ∇ g ) − h Hess g ( f X, ∇ g ) d m , which is the thesis. The general case comes by approximation. (cid:3) It will be useful to introduce the following notation: for X ∈ W , C ( T X ) and Z ∈ L ∞ ( T X ),the vector field ∇ Z X ∈ L ( T X ) is defined by h∇ Z X, Y i := ∇ X : ( Z ⊗ Y ) , m -a.e. , ∀ Y ∈ L ( T X ) . Since L ( T X ) ∋ Y
7→ ∇ X : ( Z ⊗ Y ) ∈ L ( X ) is continuous and L ∞ -linear, we see fromProposition 1.15 that this is a good definition. Proposition 2.19 (Compatibility with the metric) . Let
X, Y ∈ L ∞ ∩ H , C ( T X ) . Then h X, Y i ∈ W , ( X ) and d h X, Y i ( Z ) = h∇ Z X, Y i + h∇ Z Y, X i , m -a.e. , for every Z ∈ L ( T X ) . ketch of the proof For
X, Y ∈ TestV( X ) the claim follows directly from (2.20) and (2.21).The general case then follows by approximation (to be done carefully, because for ( X n ) , ( Y n )converging to X, Y in H , C ( T X ) the differential of h X n , Y n i only converge in L ( T ∗ X ) so thatProposition 1.11 cannot be applied as it is). (cid:3) In the following proposition and below we shall write X ( f ) in place of d f ( X ). Proposition 2.20 (Torsion free identity) . Let f ∈ LIP ∩ H , ( X ) and X, Y ∈ L ∞ ∩ H , C ( T X ) .Then X ( f ) , Y ( f ) ∈ W , ( X ) and X ( Y ( f )) − Y ( X ( f )) = d f ( ∇ X Y − ∇ Y X ) , m -a.e. . (2.22) proof By the very definition of H , C ( T X ) we have ∇ f ∈ L ∞ ∩ H , C ( T X ), thus from Proposition2.19 we know that Y ( f ) ∈ W , ( X ) and X ( Y ( f )) = ∇ Y : ( X ⊗ ∇ f ) + Hess f ( X, Y ) = d f ( ∇ X Y ) + Hess f ( X, Y ) . Subtracting the analogous expression for Y ( X ( f )) and using the symmetry of the Hessian weconclude. (cid:3) Since Test( X ) ⊂ LIP ∩ H , ( X ), we have that { d f : f ∈ LIP ∩ H , ( X ) } generates L ( T ∗ X ),hence ∇ X Y − ∇ Y X is the only vector field for which the identity (2.22) holds. It is thereforemeaningful to define the Lie bracket of vector fields as:[ X, Y ] := ∇ X Y − ∇ Y X ∈ L ( T X ) ∀ X, Y ∈ H , C ( T X ) . In the smooth setting, the Cauchy-Lipschitz theorem provides existence and uniqueness forthe solution of γ ′ t = v t ( γ t ) γ given, (2.23)for a suitable family of Lipschitz vector fields v t on R d . The Ambrosio-Di Perna-Lions theory([21], [1]) provides an extension of this classical result to the case of Sobolev/BV vector fieldswith a one-sided bound on the divergence. As it turned out ([14]) such theory admits anextension to RCD spaces, which we very briefly recall here. We remark that [14] has beendeveloped independently from [25], and that the definitions and results in [14] cover casesmore general than those we recall below: here we just want to phrase the main result of [14]in the language we are proposing and in a set of assumptions which is usually relevant inapplications.The concept of solution of (2.23) is replaced by the following definition:
Definition 2.21 (Regular Lagrangian flow) . Let ( X t ) ∈ L ([0 , , L ( T X )) . We say that F : [0 , × X → X is a Regular Lagrangian Flow for ( X t ) provided:i) For some C > it holds ( F t ) ∗ m ≤ C m ∀ t ∈ [0 , . (2.24) ii) For m -a.e. x ∈ X the curve [0 , ∋ t F t ( x ) ∈ X is continuous and such that F ( x ) = x . ii) for every f ∈ W , ( X ) we have: for m -a.e. x ∈ X the function t f ( F t ( x )) belongs to W , (0 , and it holds dd t f ( F t ( x )) = d f ( X t )( F t ( x )) m × L | [0 , -a.e.( x, t ) (2.25) where the derivative at the left-hand-side is the distributional one. Notice that it is due to property ( i ) that property ( iii ) makes sense. Indeed, for given X t ∈ L ( T X ) and f ∈ W , ( X ) the function d f ( X t ) ∈ L ( X ) is only defined m -a.e., so that(part of) the role of (2.24) is to grant that d f ( X t ) ◦ F t is well defined m -a.e..Notice that by arguing as in the proof of the equality (1.29) we see that for m -a.e. x ∈ X the curve t F t ( x ) is absolutely continuous with | ˙ F t ( x ) | = | X t | ( F t ( x )) m × L | [0 , -a.e.( x, t ) . Taking into account the integrability condition on ( X t ) we then see that for every µ ∈ P ( X )with µ ≤ C m for some C >
0, the plan π := ( F · ) ∗ µ is a test plan, where F · : X → C ([0 , , X )is the m -a.e. defined map sending x to t F t ( x ). It is then clear from the defining properties(1.28) and (2.25) that the velocity vector fields π ′ t ∈ L ( T X , e t , π ) of π are given by π ′ t = e ∗ t X t , a . e . t. The main result of [14] can then be stated as:
Theorem 2.22.
Let ( X t ) ∈ L ([0 , , W , C ( T X )) ∩ L ∞ ([0 , , L ∞ ( T X )) be such that X t ∈ D (div) for a.e. t ∈ [0 , , with Z k div( X t ) k L ( X ) + k (cid:0) div( X t ) (cid:1) − k L ∞ ( X ) d t < ∞ . Then a Regular Lagrangian Flow F t for ( X t ) exists and is unique, in the sense that if ˜ F isanother flow, then for m -a.e. x ∈ X it holds F t ( x ) = ˜ F t ( x ) for every t ∈ [0 , . Moreover: ( F t ) ∗ m ≤ exp (cid:16) Z t k (cid:0) div( X t ) (cid:1) − k L ∞ ( X ) d t (cid:17) m ∀ t ∈ [0 , . It is outside the scope of this note to present the proof of this result, which is non-trivialeven in Euclidean setting; we rather refer to [3] and [13] for an overview of the theory in R n and RCD spaces respectively.
Let H be a Hilbert module and put H ⊗ k := H ⊗ · · · ⊗ H | {z } k times . The k -th exterior power H ∧ k of H is defined as the quotient of H ⊗ k w.r.t. the space of L ∞ -linear combinations of elementsof the form v ⊗ · · · ⊗ v k with v i = v j for some i = j .We denote by v ∧ · · · ∧ v k the image of v ⊗ · · · ⊗ v k under the quotient map and endow H ∧ k with the (rescaling of the) quotient pointwise scalar product given by h v ∧ · · · ∧ v k , w ∧ · · · ∧ w k i := det (cid:0) h v i , w j i (cid:1) m -a.e. . H ∧ k is a Hilbert module. For H = L ( T ∗ X ), we write L (Λ k T ∗ X ) for the k -th exterior power if k >
1, keeping the notation L ( T ∗ X ) and L ( X ) forthe cases k = 1 , L (Λ k T ∗ X ) as k -forms.It is readily checked that the duality relation between L ( T ∗ X ) and L ( T X ) induces a dual-ity relation between the respective k -th exterior powers; we shall typically write ω ( X , . . . , X k )in place of ω ( X ∧ · · · ∧ X k ). In the smooth setting the exterior differential of the k -form ω if characterized byd ω ( X , . . . , X k ) = X i ( − i d (cid:0) ω ( X , . . . , ˆ X i , . . . , X k ) (cid:1) ( X i )+ X i
These all follow from the definitions, the identity d f ∧ . . . ∧ d f k ( X , . . . , X k ) =det(d f i ( X j )) and routine computations based on the calculus rules obtained so far. (cid:3) This last proposition motivates the following definition:
Definition 2.26. H , (Λ k T ∗ X ) ⊂ W , (Λ k T ∗ X ) is the W , -closure of TestForm k ( X ) . Clearly, H , (Λ k T ∗ X ) is dense in L (Λ k T ∗ X ). Another crucial property of H , -forms is: Proposition 2.27 (d = 0 for forms in H , (Λ k T ∗ X )) . Let ω ∈ H , (Λ k T ∗ X ) . Then d ω ∈ H , (Λ k +1 T ∗ X ) and d(d ω ) = 0 . proof The identities (2.27) and (2.28) establish the claim for forms in TestForm k ( X ). Thegeneral case then follows by approximation taking into account the closure of the exteriordifferential. (cid:3) .5.3 de Rham cohomology and Hodge theorem Proposition 2.27 is the starting point for building de Rham cohomology. The definition ofclosed and exact k -forms is naturally given by:C k ( X ) := (cid:8) ω ∈ H , (Λ k T ∗ X ) : d ω = 0 (cid:9) , E k ( X ) := (cid:8) d ω : ω ∈ H , (Λ k − T ∗ X ) (cid:9) . Proposition 2.27 ensures that E k ( X ) ⊂ C k ( X ) and the closure of the differential that C k ( X ) isa closed subspace of L (Λ k T ∗ X ). Hence defining E k ( X ) asE k ( X ) := L (Λ k T ∗ X )-closure of E k ( X )we also have that E k ( X ) ⊂ C k ( X ). We can then give the following: Definition 2.28 (de Rham cohomology) . For k ∈ N the Hilbert space H kdR ( X ) is defined asthe quotient H k dR ( X ) := C k ( X )E k ( X ) , where C k ( X ) and E k ( X ) are endowed with the L (Λ k T ∗ X ) -norm. Cohomology as we just defined it is functorial in the following sense. Let ϕ : X → X beof bounded deformation and recall that in Theorem 1.34 we gave the definition of pullbackof 1-forms ϕ ∗ : L ( T ∗ X ) → L ( T ∗ X ). It is then not hard to see that for every k ∈ N thereis a unique linear map ϕ ∗ : L (Λ k T ∗ X ) → L (Λ k T ∗ X ) such that ϕ ∗ ( ω ∧ . . . ∧ ω k ) = ( ϕ ∗ ω ) ∧ . . . ∧ ( ϕ ∗ ω k ) ,ϕ ∗ ( f ω ) = f ◦ ϕ ϕ ∗ ω, | ϕ ∗ ω | ≤ Lip ( ϕ ) k | ω | ◦ ϕ, (2.29)for every ω , . . . , ω k ∈ L ∩ L ∞ ( T ∗ X ), ω ∈ L (Λ k T ∗ X ) and f ∈ L ∞ ( X ).Then we have: Proposition 2.29 (Functoriality) . Let ( X , d , m ) , ( X , d , m ) be two RCD ( K, ∞ ) spaces, K ∈ R , and ϕ : X → X of bounded deformation. Then for every k ∈ N and ω ∈ H , (Λ k T ∗ X ) we have ϕ ∗ ω ∈ H , (Λ k T ∗ X ) and d( ϕ ∗ ω ) = ϕ ∗ d ω. (2.30) In particular, ϕ ∗ passes to the quotient and induces a linear continuous map from H k dR ( X ) to H k dR ( X ) with norm bounded by Lip ( ϕ ) k .proof From the linearity and continuity of ϕ ∗ and of d : H , (Λ k T ∗ X ) → L (Λ k +1 T ∗ X ), itis sufficient to prove (2.30) for ω of the form ω = f d f ∧ . . . ∧ d f k , for f i ∈ Test( X ). In thiscase (2.29) gives that ϕ ∗ ω = f ◦ ϕ d( f ◦ ϕ ) ∧ . . . ∧ d( f k ◦ ϕ )and since f i ◦ ϕ ∈ L ∞ ∩ W , ( X ) with | d( f i ◦ ϕ ) | ∈ L ∞ ( X ), from point ( i ) of Proposition2.25 we deduce thatd ϕ ∗ ω = d( f ◦ ϕ ) ∧ d( f ◦ ϕ ) ∧ . . . ∧ d( f k ◦ ϕ ) = ϕ ∗ d ω, as desired.The fact that ϕ ∗ passes to the quotient is then a direct consequence of its linearity andcontinuity, and the bound on the norm comes directly from the last in (2.29). (cid:3)
42e now want to show that an analogue of Hodge theorem about representation of coho-mology classes via harmonic forms holds. We shall need a few definitions.We start with that of codifferential , defined as the adjoint of the exterior differential:for k ∈ N the space D ( δ ) ⊂ L (Λ k T ∗ X ) is the space of those forms ω for which there exists aform δω ∈ L (Λ k − T ∗ X ), called codifferential of ω , such that Z h δω, η i d m = Z h ω, d η i d m , ∀ η ∈ TestForm k − ( X ) . In the case k = 0 we put D ( δ ) := L ( X ) and define the δ operator to be identically 0 on it.It is not hard to check that δ is well-defined and closed, while some computations (whichwe omit) show that TestForm k ( X ) ⊂ D ( δ ). In particular, the following definitions of ‘Hodge’Sobolev spaces are meaningful: Definition 2.30.
For k ∈ N , we define W , (Λ k T ∗ X ) := W , (Λ k T ∗ X ) ∩ D ( δ ) with the norm k ω k W , (Λ k T ∗ X ) := k ω k L (Λ k T ∗ X ) + k d ω k L (Λ k +1 T ∗ X ) + k δω k L (Λ k − T ∗ X ) . The space H , (Λ k T ∗ X ) is the W , -closure of TestForm k ( X ) . In particular, H , (Λ k T ∗ X ) is a Hilbert space dense in L (Λ k T ∗ X ). Definition 2.31 (Hodge Laplacian and harmonic forms) . Given k ∈ N , the domain D (∆ H ) ⊂ H , (Λ k T ∗ X ) of the Hodge Laplacian is the set of ω ∈ H , (Λ k T ∗ X ) for which there exists α ∈ L (Λ k T ∗ X ) such that Z h α, η i d m = Z h d ω, d η i + h δω, δη i d m , ∀ η ∈ H , (Λ k T ∗ X ) . In this case, the form α (which is unique by the density of H , (Λ k T ∗ X ) in L (Λ k T ∗ X ) ) willbe called Hodge Laplacian of ω and denoted by ∆ H ω .The space Harm k ( X ) ⊂ D (∆ H ) is the space of forms ω ∈ D (∆ H ) such that ∆ H ω = 0 . In the case of functions, we have the usual unfortunate sign relation:∆ H f = − ∆ f ∀ f ∈ D (∆) = D (∆ H ) ⊂ L (Λ T ∗ X ) = L ( X ) . The Hodge Laplacian is a closed operator: this can be seen by noticing that it is the subdif-ferential of the convex and lower semicontinuous functional on L (Λ k T ∗ X ) defined by ω Z | d ω | + | δω | d m if ω ∈ H , (Λ k T ∗ X ), + ∞ otherwise . From such closure it follows that Harm k ( X ) is a closed subspace of L (Λ k T ∗ X ) and thus aHilbert space itself when endowed with the L (Λ k T ∗ X )-norm. We then have: Theorem 2.32 (Hodge theorem on
RCD spaces) . The map
Harm k ( X ) ∋ ω [ ω ] ∈ H k dR ( X ) is an isomorphism of Hilbert spaces. roof Start noticing that ω ∈ Harm k ( X ) ⇔ d ω = 0 and δω = 0 , indeed the ‘if’ is obvious by definition, while the ‘only if’ comes from the identity Z h ω, ∆ H ω i d m = Z | d ω | + | δω | d m . Recalling the definition of δ , we thus see that ω ∈ Harm k ( X ) ⇔ ω ∈ C k ( X ) and Z h ω, η i d m = 0 ∀ η ∈ E k ( X ) . The conclusion follows recalling that for every Hilbert space H and subspace V , the map V ⊥ ∋ w w + V ∈ H/V is an isomorphism of Hilbert spaces. (cid:3)
In the course of this section we shall abuse a bit the notation and identify vector and covectorfields, thus in for instance we shall write X ∈ D (∆ H ) and consider the vector field ∆ H X ∈ L ( T X ) when we should write X ♭ ∈ D (∆ H ) and (∆ H X ♭ ) ♯ ∈ L ( T X ), where · ♭ : L ( T X ) → L ( T ∗ X ) and · ♯ : L ( T ∗ X ) → L ( T X ) are the Riesz (musical) isomorphisms.We begin reinterpreting the key Lemma 2.8: the differential operators introduced so farallow to restate the key inequalities (2.9), (2.10) in a much more familiar way. Lemma 2.33.
Let X ∈ TestV( X ) . Then X ∈ D (∆ H ) , | X | ∈ D ( ∆ ) and the inequality ∆ | X | ≥ (cid:16) |∇ X | − h X, ∆ H X i + K | X | (cid:17) m (2.31) holdsSketch of the proof Let X = P i g i ∇ f i for f i , g i ∈ Test( X ). It is only a matter of computationsto see that | X | ∈ D ( ∆ ) and X ∈ D (∆ H ) with ∆ | X | X i,j g i g j ∆ h∇ f i , ∇ f j i + (cid:16) g j ∆ g i h∇ f i , ∇ f j i + h∇ g i , ∇ g j i h∇ f i , ∇ f j i (cid:17) m + (cid:16) g i Hess f i ( ∇ f j , ∇ g j ) + 2 g i Hess f j ( ∇ f i , ∇ g j ) (cid:17) m ∆ H X = X i − g i d∆ f i − ∆ g i d f i − f i ( ∇ g i , · )( ∇ X ) Asym = X i ∇ g i ⊗ ∇ f i − ∇ f i ⊗ ∇ g i µ (( f i ) , ( g i )) given in Lemma 2.8 we see that µ (cid:0) ( f i ) , ( g i ) (cid:1) = ∆ | X | (cid:16) h X, (∆ H X ) i − K | X | − | ( ∇ X ) Asym | (cid:17) m . ∆ | X | = ∆ ac | X | m + ∆ sing | X | , with ∆ sing | X | ⊥ m , inequality (2.9) inLemma 2.8 yields ∆ sing | X | ≥ m ∈ N and choice of h , . . . , h m ∈ Test( X ) we have (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∇ X : m X i =1 ∇ h i ⊗ ∇ h i (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ r ∆ ac | X | h X, ∆ H X i − K | X | − | ( ∇ X ) Asym | (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) m X i =1 ∇ h i ⊗ ∇ h i (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) HS m -a.e., which in turn implies2 ∇ X : m X i =1 ∇ h i ⊗ ∇ h i − (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) m X i =1 ∇ h i ⊗ ∇ h i (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ ∆ ac | X | h X, ∆ H X i − K | X | − | ( ∇ X ) Asym | m -a.e.. Noticing that L ∞ -linear combinations of objects of the form ∇ h ⊗ ∇ h for h ∈ Test( X ),are L -dense in the space of symmetric 2-tensors, taking the (essential) supremum in this lastinequality among m ∈ N and choices of h , . . . , h m ∈ Test( X ) we obtain | ( ∇ X ) Sym | ≤ ∆ ac | X | h X, ∆ H X i − K | X | − | ( ∇ X ) Asym | , m -a.e. , which, recalling (2.6) and (2.32), gives the conclusion. (cid:3) Let us introduce the ‘covariant energy’ and the ‘Hodge energy’ functionals on L ( T X ) as E C ( X ) := 12 Z |∇ X | d m if X ∈ H , C ( T X ) , + ∞ otherwise , E H ( X ) := 12 Z | d X | + | δX | d m if X ∈ H , ( T X ) , + ∞ otherwise . Notice that the closure of the differential operators involved grant that these are L ( T X )-lowersemicontinuous. Then the last lemma has the following useful corollary (which generalizesCorollary 2.10): Corollary 2.34.
We have H , ( T X ) ⊂ H , C ( T X ) and E C ( X ) ≤ E H ( X ) − K2 k X k ( TX ) , ∀ X ∈ H , H ( TX ) . (2.33) proof For X ∈ TestV( X ) the bound (2.33) comes integrating (2.31) recalling (2.4). The generalcase then follows approximating X ∈ H , ( T X ) with vector fields in TestV( X ) and using the L -lower semicontinuity of E C . (cid:3) We are now ready to introduce the Ricci curvature operator:
Theorem/Definition 2.35 (Ricci curvature) . There exists a unique continuous map, calledRicci curvature,
Ric : [ H , ( T X )] → Meas ( X ) such that for every X, Y ∈ TestV( X ) it holds Ric ( X, Y ) = ∆ h X, Y i (cid:16) h X, ∆ H Y i + 12 h Y, ∆ H X i − ∇ X : ∇ Y (cid:17) m . (2.34)45 uch map is bilinear, symmetric and satisfies Ric ( X, X ) ≥ K | X | m (2.35) Ric ( X, Y )( X ) = Z h d X, d Y i + δX δY − ∇ X : ∇ Y d m (2.36) k Ric ( X, Y ) k TV ≤ q E H ( X ) + K − k X k ( TX ) q E H ( Y ) + K − k Y k ( TX ) (2.37) for every X, Y ∈ H , ( T X ) , where K − := max { , − K } .Sketch of the proof The fact that the right hand side of (2.34) is well defined for
X, Y ∈ TestV( X ) is a direct consequence of Lemma 2.33. That such right hand side is bilinear, sym-metric and satisfies (2.36) is obvious, while (2.35) is a restatement of (2.31). Thanks to thedensity of TestV( X ) in H , ( T X ), to conclude it is therefore sufficient to prove (2.37) for X, Y ∈ TestV( X ): we shall do so for the case K = 0 only.Let X, Y ∈ TestV( X ), choose µ ∈ Meas ( X ), µ ≥
0, such that
Ric ( X, X ) , Ric ( X, Y )and
Ric ( Y, Y ) are all absolutely continuous w.r.t. µ and let f, g, h be the respective Radon-Nikodym derivatives. Then (2.35) grants that f, h ≥ µ -a.e. and that for any λ ∈ R we have Ric ( λX + Y, λX + Y ) ≥
0. Hence λ f + 2 λg + h ≥ , µ -a.e. , which easily implies | g | ≤ √ f h µ -a.e. and therefore k Ric ( X, Y ) k TV = Z | g | d µ ≤ sZ f d µ Z h d µ = p k Ric ( X, X ) k TV k Ric ( Y, Y ) k TV . The conclusion then follows noticing that k Ric ( X, X ) k TV (2.35) = Ric ( X, X )( X ) (2.36) = 2 E H ( X ) − C ( X ) ≤ H ( X ) ∀ X ∈ TestV( X ) . (cid:3) The Ricci curvature operator as defined in the last theorem is a tensor in the sense thatit holds:
Ric ( f X, Y ) = f Ric ( X, Y ) ∀ X, Y ∈ H , ( T X ) , f ∈ Test( X ) , as can be showed with some algebraic manipulations based on the calculus rules developedso far (we omit the details). Moreover, directly from the definitions we get( X , d , m ) is a RCD ( K ′ , ∞ ) space with Ric ( X, X ) ≥ K | X | m ∀ X ∈ H , ( T X ) (cid:27) ⇒ ( X , d , m ) is a RCD ( K, ∞ ) space. Remark 2.36.
Directly from the definition it is easy to see that the Ricci measure gives 0mass to sets with 0 capacity. It follows that, for instance, on a two dimensional space with aconical singularity, the Ricci curvature as we defined it does not see any ‘delta’ at the vertex:this also implies that we cannot hope for such measure to have any kind of Gauss-Bonnetformula.If the space is sufficiently regular ( C , manifold is enough), then one can detect thesingularity of the curvature at the vertex of such a cone by computing the curvature alongobjects more regular than Sobolev vector fields, namely Lipschitz half densities (see [36]). (cid:4) .7 Some properties in the finite dimensional case Here we briefly present, without proofs, some related results about analysis and geometry offinite dimensional
RCD spaces ([26], [9], [22], [11]).
Definition 2.37 ( RCD ∗ ( K, N ) spaces) . Let K ∈ R , N ∈ [1 , ∞ ) . ( X , d , m ) is a RCD ∗ ( K, N ) space provided it is a RCD ∗ ( K, ∞ ) space and the Bochner inequality holds in the followingform: Z ∆ g | d f | d m ≥ Z g (cid:16) (∆ f ) N + h∇ f, ∇ ∆ f i + K | d f | (cid:17) d m for every f ∈ D (∆) with ∆ f ∈ W , ( X ) and g ∈ L ∞ ( X ) ∩ D (∆) with g ≥ , ∆ g ∈ L ∞ ( X ) . On compact finite-dimensional
RCD spaces, the following natural second-order differenti-ation formula holds (proved in [31]), which links the Hessian as we defined it to the secondderivative along geodesics, compare with Theorem 1.44.
Theorem 2.38 (Second order differentiation formula) . Let ( X , d , m ) be a compact RCD ∗ ( K, N ) space, N < ∞ and ( µ t ) ⊂ P ( X ) a W -geodesic such that µ , µ ≤ C m for some C > .Then for every f ∈ H , ( X ) the map t R f d µ t is C ([0 , and it holds d d t Z f d µ t = Z Hess f ( ∇ ϕ t , ∇ ϕ t ) d µ t ∀ t ∈ [0 , , where ϕ t is, for every t ∈ [0 , , such that for some s = t the function ( s − t ) ϕ is a Kantorovichpotential from µ t to µ s . The proof of this theorem relies upon an approximation of W -geodesics with the so-called entropic interpolation (see [35] for an overview on the topic). The result requires finite-dimensionality because is based, among other things, on the Li-Yau inequality. Compactnessis likely not needed, but so far the general result is unknown.A better understanding of the structure of RCD spaces can be achieved by introducingthe concept of local dimension of a module: we say that M has dimension n ∈ N on theBorel set E ⊂ X provided there are v , . . . , v n ∈ M such that X i f i v i = 0 ⇒ f i = 0 m -a.e. on E for every i = 1 , . . . , n,L ∞ -linear combinations of the v i ’s are dense in { v ∈ M : χ E c v = 0 } . It is then not hard to see that for any given module there exists a (unique up to negligiblesets) Borel partition ( E i ) i ∈ N ∪{∞} of X such that M has dimension i on E i for every i ∈ N andhas not finite dimension on any F ⊂ E ∞ with positive measure.When the module under consideration is the tangent one, we call the resulting partitionthe dimensional decomposition of X . This also allows to m -a.e. define the ‘analytic localdimension’ function dim loc : X → N ∪ {∞} which sends E i to i for every i ∈ N ∪ {∞} . It isconjectured that such function is actually constant (after [18] and Theorem 2.39 below thisis known to hold at least for Ricci-limit spaces), but so far this is unknown.The results in [38] grant that the pointed rescaled spaces ( X , d /r, m ( B r ( x )) − m , x ) con-verge, for m -a.e. x ∈ X , to the Euclidean space ( R n ( x ) , d Eucl , L n ( x ) ,
0) in the pointed-measured-Gromov-Hausdorff sense for some n ( x ) ∈ N , n ( x ) ≤ N . In particular, the number n ( x ) pro-vides a ‘geometric’ notion of dimension at x . It turns out ([29]) that this notion is equivalentto the analytic one: 47 heorem 2.39. With the above notation, we have dim loc ( x ) = n ( x ) for m -a.e. x ∈ X . Inparticular, m -a.e. we have dim loc ≤ N . In fact, something stronger holds: the tangent module L ( T X ) is isomorphic to the spaceof ‘ L sections’ of the bundle on X made of the collections of the pmGH-limits of rescaledspaces. The proof of this fact uses the charts built in [38], along with the improvements givenin [34] and [28], to produce the desired isomorphism.In a different direction, the properties of the cohomology groups reflect on the geometryof the space, as shown by the following result (proved in [30]) which generalizes a classicalresult of Bochner to the setting of RCD spaces.
Theorem 2.40.
Let ( X , d , m ) be a RCD (0 , ∞ ) space. Then dim( H dR ( X )) ≤ min X dim loc .Moreover, if ( X , d , m ) is RCD (0 , N ) and dim( H dR ( X )) = N (so that in particular N ∈ N ),then X is the flat N -dimensional torus. The first part of the statement follows noticing that, much like in the smooth case, har-monic 1-forms must be parallel (because of (2.33)). The second claim is harder, because theclassical proof which passes via universal cover can’t be adapted; instead, the desired iso-morphism is built from scratch by considering the Regular Lagrangian Flows of a basis ofharmonic forms.In the smooth setting of weighted Riemannian manifolds, it is well known that the validityof a curvature dimension condition is linked to the fact that the N -Ricci tensor is boundedfrom below by K and that N is equal to the geometric dimension of the manifold if and onlyif the trace of the Hessian is equal to the Laplacian.Something similar holds on RCD ∗ ( K, N ) spaces, as proved in [32] by adapting the computa-tions done in [45] to the non-smooth setting. Let us introduce the function R N : [ H , ( T X )] → L ( X ) as R N ( X, Y ) := (cid:0) tr( ∇ X ) − div X (cid:1)(cid:0) tr( ∇ Y ) − div Y (cid:1) N − dim loc if dim loc < N, N -Ricci tensor Ric N : [ H , ( T X )] → Meas ( X ) as Ric N ( X, Y ) :=
Ric ( X, Y ) − R N ( X, Y ) m . It is easy to see that |∇ X | + R N ( X, X ) ≥ (div X ) N , and
Ric N ( f X, Y ) = f Ric N ( X, Y ) , for every X, Y ∈ H , ( T X ) and f ∈ Test( X ).The main results in [32] can then be summarized as: Theorem 2.41.
Let ( X , d , m ) be a RCD ∗ ( K ′ , ∞ ) space. Then it is a RCD ∗ ( K, N ) space if andonly ifi) dim loc ≤ N m -a.e.ii) For any X ∈ H , ( T X ) we have tr( ∇ X ) = div X m -a.e. on { dim loc = N } iii) For any X ∈ H , ( T X ) we have Ric N ( X, X ) ≥ K | X | m . eferences [1] L. Ambrosio , Transport equation and Cauchy problem for BV vector fields , Invent.Math., 158 (2004), pp. 227–260.[2] L. Ambrosio, M. Colombo, and S. Di Marino , Sobolev spaces in metric measurespaces: reflexivity and lower semicontinuity of slope . Accepted at Adv. St. in Pure Math.,arXiv:1212.3779, 2014.[3]
L. Ambrosio and G. Crippa , Existence, uniqueness, stability and differentiability prop-erties of the flow associated to weakly differentiable vector fields , in Transport equa-tions and multi-D hyperbolic conservation laws, vol. 5 of Lect. Notes Unione Mat. Ital.,Springer, Berlin, 2008, pp. 3–57.[4]
L. Ambrosio and N. Gigli , A user’s guide to optimal transport , in Modelling andOptimisation of Flows on Networks, Lecture Notes in Mathematics, Springer Berlin Hei-delberg, 2013, pp. 1–155.[5]
L. Ambrosio, N. Gigli, and G. Savar´e , Gradient flows in metric spaces and in thespace of probability measures , Lectures in Mathematics ETH Z¨urich, Birkh¨auser Verlag,Basel, second ed., 2008.[6] ,
Heat flow and calculus on metric measure spaces with Ricci curvature boundedbelow—the compact case , Boll. Unione Mat. Ital. (9), 5 (2012), pp. 575–629.[7] ,
Calculus and heat flow in metric measure spaces and applications to spaces withRicci bounds from below , Invent. Math., 195 (2014), pp. 289–391.[8] ,
Metric measure spaces with Riemannian Ricci curvature bounded from below , DukeMath. J., 163 (2014), pp. 1405–1490.[9] ,
Bakry- ´Emery curvature-dimension condition and Riemannian Ricci curvaturebounds , The Annals of Probability, 43 (2015), pp. 339–404.[10]
L. Ambrosio and S. Honda , New stability results for sequences of metric measurespaces with uniform Ricci bounds from below . To appear in “Measure Theory in Non-Smooth Spaces”, Partial Differential Equations and Measure Theory. De Gruyter Open,2017. ArXiv:1605.07908.[11]
L. Ambrosio, A. Mondino, and G. Savar´e , Nonlinear diffusion equations and cur-vature conditions in metric measure spaces . Preprint, arXiv:1509.07273, 2015.[12]
L. Ambrosio, F. Stra, and D. Trevisan , Weak and strong convergence of derivationsand stability of flows with respect to MGH convergence , J. Funct. Anal., 272 (2017),pp. 1182–1229.[13]
L. Ambrosio and D. Trevisan , Lecture notes on the DiPerna-Lions theory in abstractmeasure spaces . Accepted at Annales Fac. Sc. de Toulouse, arXiv:1505.05292, 2015.[14] ,
Well posedness of Lagrangian flows and continuity equations in metric measurespaces , Anal. PDE, 7 (2014), pp. 1179–1234.4915]
D. Bakry , Transformations de Riesz pour les semi-groupes sym´etriques. II. ´Etude sousla condition Γ ≥
0, in S´eminaire de probabilit´es, XIX, 1983/84, vol. 1123 of LectureNotes in Math., Springer, Berlin, 1985, pp. 145–174.[16]
J. Cheeger , Differentiability of Lipschitz functions on metric measure spaces , Geom.Funct. Anal., 9 (1999), pp. 428–517.[17]
J. Cheeger and T. H. Colding , On the structure of spaces with Ricci curvaturebounded below. III , J. Differential Geom., 54 (2000), pp. 37–74.[18]
T. H. Colding and A. Naber , Sharp H¨older continuity of tangent cones for spaceswith a lower Ricci curvature bound and applications , Ann. of Math. (2), 176 (2012),pp. 1173–1229.[19]
G. De Philippis and N. Gigli , From volume cone to metric cone in the nonsmoothsetting , Geom. Funct. Anal., 26 (2016), pp. 1526–1587.[20]
J. Diestel and J. J. Uhl, Jr. , Vector measures , American Mathematical Society,Providence, R.I., 1977. With a foreword by B. J. Pettis, Mathematical Surveys, No. 15.[21]
R. J. DiPerna and P.-L. Lions , Ordinary differential equations, transport theory andSobolev spaces , Invent. Math., 98 (1989), pp. 511–547.[22]
M. Erbar, K. Kuwada, and K.-T. Sturm , On the equivalence of the entropiccurvature-dimension condition and Bochner’s inequality on metric measure spaces , In-ventiones mathematicae, 201 (2014), pp. 1–79.[23]
N. Gigli , On the heat flow on metric measure spaces: existence, uniqueness and stability ,Calc. Var. PDE, 39 (2010), pp. 101–120.[24] ,
The splitting theorem in non-smooth context . Preprint, arXiv:1302.5555, 2013.[25] ,
Nonsmooth differential geometry - an approach tailored for spaces with Ricci cur-vature bounded from below . Accepted at Mem. Amer. Math. Soc., arXiv:1407.0809, 2014.[26] ,
On the differential structure of metric measure spaces and applications , Mem.Amer. Math. Soc., 236 (2015), pp. vi+91.[27]
N. Gigli, K. Kuwada, and S.-i. Ohta , Heat flow on Alexandrov spaces , Communica-tions on Pure and Applied Mathematics, 66 (2013), pp. 307–331.[28]
N. Gigli and E. Pasqualetto , Behaviour of the reference measure on
RCD spacesunder charts . Preprint, arxiv: 1607.05188, 2016.[29] ,
Equivalence of two different notions of tangent bundle on rectifiable metric measurespaces . Preprint, arXiv:1611.09645, 2016.[30]
N. Gigli and C. Rigoni , Recognizing the flat torus among
RCD ∗ (0 , N ) spaces via thestudy of the first cohomology group . Work in progress.[31] N. Gigli and L. Tamanini , Second order differentiation formula on compact
RCD ∗ ( K, N ) spaces . Preprint, arXiv:1701.03932.5032] B. Han , Ricci tensor on
RCD ∗ ( K, N ) spaces . Preprint, arXiv: 1412.0441.[33] S. Honda , Spectral convergence under bounded Ricci curvature . Preprint,arXiv:1510.05349.[34]
M. Kell and A. Mondino , On the volume measure of non-smooth spaces with riccicurvature bounded below . Preprint, arXiv:1607.02036.[35]
C. L´eonard , A survey of the Schr¨odinger problem and some of its connections withoptimal transport , Discrete Contin. Dyn. Syst., 34 (2014), pp. 1533–1574.[36]
J. Lott , Ricci measure for some singular riemannian metrics , Mathematische Annalen,365 (2016), pp. 449–471.[37]
J. Lott and C. Villani , Ricci curvature for metric-measure spaces via optimal trans-port , Ann. of Math. (2), 169 (2009), pp. 903–991.[38]
A. Mondino and A. Naber , Structure Theory of Metric-Measure Spaces with LowerRicci Curvature Bounds I . Preprint, arXiv:1405.2222, 2014.[39]
T. Rajala and K.-T. Sturm , Non-branching geodesics and optimal maps in strong CD ( K, ∞ ) -spaces , Calc. Var. Partial Differential Equations, 50 (2012), pp. 831–846.[40] F. Santambrogio , Optimal transport for applied mathematicians , vol. 87 of Progressin Nonlinear Differential Equations and their Applications, Birkh¨auser/Springer, Cham,2015. Calculus of variations, PDEs, and modeling.[41]
J.-L. Sauvageot , Tangent bimodule and locality for dissipative operators on C ∗ -algebras , in Quantum probability and applications, IV (Rome, 1987), vol. 1396 of LectureNotes in Math., Springer, Berlin, 1989, pp. 322–338.[42] G. Savar´e , Self-improvement of the Bakry- ´Emery condition and Wasserstein contrac-tion of the heat flow in
RCD( K, ∞ ) metric measure spaces , Discrete Contin. Dyn. Syst.,34 (2014), pp. 1641–1661.[43] N. Shanmugalingam , Newtonian spaces: an extension of Sobolev spaces to metric mea-sure spaces , Rev. Mat. Iberoamericana, 16 (2000), pp. 243–279.[44]
K.-T. Sturm , On the geometry of metric measure spaces. I , Acta Math., 196 (2006),pp. 65–131.[45] ,
Ricci Tensor for Diffusion Operators and Curvature-Dimension Inequalities underConformal Transformations and Time Changes . Preprint, arXiv:1401.0687, 2014.[46]
C. Villani , Optimal transport. Old and new , vol. 338 of Grundlehren der Mathematis-chen Wissenschaften, Springer-Verlag, Berlin, 2009.[47]