Introduction
Lower bounds on the Ricci curvature of metric measure spaces were introduced, in a synthetic form based on optimal transport, by Sturm [26 , 27 ] and Lott–Villani [22 ] , through the curvature-dimension condition CD(K , N K,N K , N ); see also [9 , 12 ] for the reduced condition CD∗ ( K , N ) ^{*}(K,N) ∗ ( K , N ) and its globalisation. The first-order calculus underlying the theory goes back to Cheeger [11 ] and was developed, in the form used here, by Ambrosio, Gigli and Savaré [2 , 3 ] , who also identified the heat flow as the L 2 L^2 L 2 -gradient flow of the Cheeger energy and as the Wasserstein gradient flow of the entropy (see also [14 ] ). Requiring in addition that the Cheeger energy be a quadratic form, that is, that the space be infinitesimally Hilbertian [15 ] , singles out the class of RCD(K , ∞ K,\infty K , ∞ ) spaces [4 ] , extended to σ \sigma σ -finite reference measures in [1 ] ; the finite-dimensional class RCD(K , N K,N K , N ) was introduced in [15 ] and studied in [13 , 7 , 12 ] . The Eulerian characterisation of these classes, via the weak Bochner inequality and the Sobolev-to-Lipschitz property, was obtained in [5 ] for N = ∞ N=\infty N = ∞ and in [13 , 7 ] for N < ∞ N<\infty N < ∞ ; the classes are stable under (pointed) measured Gromov–Hausdorff convergence [4 , 17 ] . For the second-order calculus on these spaces we refer to [16 ] and to the monograph [18 ] , whose notation we follow, and to the survey [6 ] for a general account of the theory.
In the Eulerian approach, RCD(K , ∞ K,\infty K , ∞ ) spaces can be defined defined as follows (along the lines of [5 ] ). Here W 1 , 2 ( X ) W^{1,2}(X) W 1 , 2 ( X ) , ∣ D f ∣ |Df| ∣ D f ∣ , ∇ f ⋅ ∇ g \nabla f \cdot\nabla g ∇ f ⋅ ∇ g , Δ \Delta Δ and ( h t ) t ≥ 0 (h_t)_{t\geq0} ( h t ) t ≥ 0 denote the Sobolev space, the minimal weak upper gradient, the pointwise scalar product on the tangent module, the Laplacian and the heat flow; we recall them in § 2 .
Definition 1.1 (RCD(K , ∞ K,\infty K , ∞ ) spaces). Let ( X , d ) (X,d) ( X , d ) be a complete and separable metric space, let m \mathfrak{m} m be a Borel measure on X X X which is finite on bounded sets and satisfies supp ( m ) = X \operatorname{supp}(\mathfrak{m})=X supp ( m ) = X , and let K ∈ R K\in\mathbb{R} K ∈ R . The space ( X , d , m ) (X,d,\mathfrak{m}) ( X , d , m ) is an RCD(K , ∞ K,\infty K , ∞ ) space if the following hold:
(G) Gaussian volume growth: there exist x ˉ ∈ X \bar{x}\in X x ˉ ∈ X and C , c > 0 C,c>0 C , c > 0 such that m ( B r ( x ˉ ) ) ≤ C e c r 2 \mathfrak{m}(B_r(\bar{x}))\leq Ce^{cr^2} m ( B r ( x ˉ )) ≤ C e c r 2 for every r > 0 r>0 r > 0 ;
(SL) Sobolev-to-Lipschitz property [4 , 19 ] : every f ∈ W 1 , 2 ( X ) f\in W^{1,2}(X) f ∈ W 1 , 2 ( X ) with ∣ D f ∣ ∈ L ∞ ( m ) |Df|\in L^\infty(\mathfrak{m}) ∣ D f ∣ ∈ L ∞ ( m ) admits a Lipschitz representative f ~ ∈ Lip ( X ) \tilde{f}\in\operatorname{Lip}(X) f ~ ∈ Lip ( X ) such that f ~ = f \tilde{f}=f f ~ = f m \mathfrak{m} m -a.e. and Lip ( f ~ ) ≤ ∥ D f ∥ L ∞ ( m ) \operatorname{Lip}(\tilde{f})\leq\lVert Df\rVert_{L^\infty(\mathfrak{m})} Lip ( f ~ ) ≤ ∥ D f ∥ L ∞ ( m ) ;
(IH) Infinitesimal Hilbertianity [15 ] : the Sobolev space W 1 , 2 ( X ) W^{1,2}(X) W 1 , 2 ( X ) endowed with the norm ∥ f ∥ W 1 , 2 ( X ) 2 : = ∥ f ∥ L 2 ( m ) 2 + 2 Ch ( f ) \lVert f\rVert_{W^{1,2}(X)}^2:=\lVert f\rVert_{L^2(\mathfrak{m})}^2+2\operatorname{Ch}(f) ∥ f ∥ W 1 , 2 ( X ) 2 := ∥ f ∥ L 2 ( m ) 2 + 2 Ch ( f ) , where Ch ( f ) : = 1 2 ∫ X ∣ D f ∣ 2 d m \operatorname{Ch}(f):=\frac{1}{2}\int_X |Df|^2\mathrm{d}\mathfrak{m} Ch ( f ) := 2 1 ∫ X ∣ D f ∣ 2 d m , is a Hilbert space; (BE) Weak Bochner inequality BE ( K , ∞ ) \operatorname{BE}(K,\infty) BE ( K , ∞ ) : for every f ∈ D ( Δ ) f \in D(\Delta) f ∈ D ( Δ ) with Δ f ∈ W 1 , 2 ( X ) \Delta f \in W^{1,2}(X) Δ f ∈ W 1 , 2 ( X ) and every φ ∈ D ( Δ ) ∩ L ∞ ( m ) \varphi\in D(\Delta) \cap L^{\infty}(\mathfrak{m}) φ ∈ D ( Δ ) ∩ L ∞ ( m ) with φ ≥ 0 \varphi\ge0 φ ≥ 0 and Δ φ ∈ L ∞ ( m ) \Delta\varphi\in L^{\infty}(\mathfrak{m}) Δ φ ∈ L ∞ ( m ) ,
1 2 ∫ X ∣ D f ∣ 2 Δ φ d m ≥ ∫ X φ ( ∇ f ⋅ ∇ Δ f + K ∣ D f ∣ 2 ) d m . (1) \frac{1}{2}\int_X |Df|^2\Delta\varphi\,\mathrm{d}\mathfrak{m} \ge\int_X \varphi(\nabla f\cdot\nabla\Delta f+K|Df|^2)\,\mathrm{d}\mathfrak{m}. \tag*{(1)} 2 1 ∫ X ∣ D f ∣ 2 Δ φ d m ≥ ∫ X φ ( ∇ f ⋅ ∇Δ f + K ∣ D f ∣ 2 ) d m . (1)
Remark 1.2. In the Lagrangian approach, based on the K K K -convexity of the Boltzmann–Shannon entropy along Wasserstein geodesics, (G) is a consequence of the curvature condition: Sturm proved that every CD ( K , ∞ ) \operatorname{CD}(K,\infty) CD ( K , ∞ ) space satisfies ∫ X e − c d 2 ( x 0 , ⋅ ) d m < ∞ \int_X e^{-cd^2(x_0,\cdot)}\,\mathrm{d}\mathfrak{m}<\infty ∫ X e − c d 2 ( x 0 , ⋅ ) d m < ∞ for some c > 0 c>0 c > 0 [26 Thm. 4.24] , which gives (G) since m ( B r ( x 0 ) ) e − c r 2 ≤ ∫ B r ( x 0 ) e − c d 2 ( x 0 , ⋅ ) d m \mathfrak{m}(B_r(x_0))e^{-cr^2}\le\int_{B_r(x_0)}e^{-cd^2(x_0,\cdot)}\,\mathrm{d}\mathfrak{m} m ( B r ( x 0 )) e − c r 2 ≤ ∫ B r ( x 0 ) e − c d 2 ( x 0 , ⋅ ) d m (see also [1 Section 2] ). In the Eulerian approach, (G) is part of the definition, and it is used in [5 ] to prove the equivalence with the Lagrangian approach.
The purpose of this note is to show that, in Definition 1.1, (G) follows from (SL), (IH) and (BE), and that the finiteness of m \mathfrak{m} m on bounded sets can be weakened to the requirement that m \mathfrak{m} m be a Radon measure.
Definition 1.3 (Radon measures). Let ( X , d ) (X,\mathrm{d}) ( X , d ) be a metric space. A Borel measure m \mathfrak{m} m on X X X is a Radon measure if m ( K ) < ∞ \mathfrak{m}(K)<\infty m ( K ) < ∞ for every compact set K ⊆ X K\subseteq X K ⊆ X , m \mathfrak{m} m is outer regular and m \mathfrak{m} m is inner regular on open sets, that is,
m ( B ) = inf { m ( U ) : U ⊇ B , U open } for every Borel set B ⊆ X , \mathfrak{m}(B)=\inf\{\mathfrak{m}(U):U\supseteq B,\ U\text{ open}\}\qquad\text{for every Borel set }B\subseteq X, m ( B ) = inf { m ( U ) : U ⊇ B , U open } for every Borel set B ⊆ X , m ( U ) = sup { m ( K ) : K ⊆ U , K compact } for every open set U ⊆ X . \mathfrak{m}(U)=\sup\{\mathfrak{m}(K):K\subseteq U,\ K\text{ compact}\}\qquad\text{for every open set }U\subseteq X. m ( U ) = sup { m ( K ) : K ⊆ U , K compact } for every open set U ⊆ X .
(S) ( X , d ) (X,\mathrm{d}) ( X , d ) is a complete and separable metric space, m \mathfrak{m} m is a Radon measure on X X X with supp ( m ) = X \operatorname{supp}(\mathfrak{m})=X supp ( m ) = X , and (SL), (IH) and (BE) hold for some K ∈ R K\in\mathbb{R} K ∈ R .
The objects in (SL), (IH) and (BE) make sense under (S); see § 2 . The regularity in Definition 1.3 cannot be dropped: in Remark 2.2 we give a Borel measure on a separable Hilbert space, finite on compact sets and with full support, for which (SL), (IH) and (BE) hold trivially and every ball has infinite measure. For K ∈ R K\in\mathbb{R} K ∈ R and t > 0 t>0 t > 0 we set, following the notation of [5 ]
I K ( t ) : = ∫ 0 t e 2 K s d s = e 2 K t − 1 2 K (with c 0 ( t ) : = t ) . (2) I_K(t):=\int_0^t e^{2Ks}\,\mathrm{d}s=\frac{e^{2Kt}-1}{2K}\qquad\text{(with }c_0(t):=t\text{)}. \tag*{(2)} I K ( t ) := ∫ 0 t e 2 K s d s = 2 K e 2 K t − 1 (with c 0 ( t ) := t ) . (2)
Theorem 1.4 (Gaussian volume growth). Assume (S). Let x 0 ∈ X x_0\in X x 0 ∈ X , let r 0 > 0 r_0>0 r 0 > 0 be such that m ( B r 0 ( x 0 ) ) < ∞ \mathfrak{m}(B_{r_0}(x_0))<\infty m ( B r 0 ( x 0 )) < ∞ , let t > 0 t>0 t > 0 , and set f : = 1 2 ( 1 − d ( ⋅ , x 0 ) / r 0 ) + f:=\frac{1}{2}(1-\mathrm{d}(\cdot,x_0)/r_0)^+ f := 2 1 ( 1 − d ( ⋅ , x 0 ) / r 0 ) + . Then the continuous representative of h t f h_tf h t f satisfies h t f ( x 0 ) > 0 h_tf(x_0)>0 h t f ( x 0 ) > 0 , and for every r > 0 r>0 r > 0
m ( B r ( x 0 ) ) ≤ ∥ f ∥ L 1 ( m ) ( h t f ( x 0 ) ) 2 exp ( r 2 2 I K ( t ) ) . (3) \mathfrak{m}(B_r(x_0))\le\frac{\|f\|_{L^1(\mathfrak{m})}}{(h_tf(x_0))^2}\exp\left(\frac{r^2}{2I_K(t)}\right). \tag*{(3)} m ( B r ( x 0 )) ≤ ( h t f ( x 0 ) ) 2 ∥ f ∥ L 1 ( m ) exp ( 2 I K ( t ) r 2 ) . (3) In particular, m \mathfrak{m} m is finite on bounded sets, and ( X , d , m ) (X,\mathrm{d},\mathfrak{m}) ( X , d , m ) is an RCD ( K , ∞ ) \operatorname{RCD}(K,\infty) RCD ( K , ∞ ) space.
A radius r 0 r_0 r 0 as in Theorem 1.4 exists for every x 0 ∈ X x_0\in X x 0 ∈ X by Lemma 2.1.
Remark 1.5. (Dependence of the constants). Choosing t = 1 t=1 t = 1 in (1.3) yields m ( B r ( x 0 ) ) ≤ C e c r 2 \mathfrak{m}(B_r(x_0))\le Ce^{cr^2} m ( B r ( x 0 )) ≤ C e c r 2 with c = ( 2 I K ( 1 ) ) − 1 c=(2I_K(1))^{-1} c = ( 2 I K ( 1 ) ) − 1 and C = ∥ f ∥ L 1 ( m ) ( h 1 f ( x 0 ) ) − 2 C=\|f\|_{L^1(\mathfrak{m})}(h_1f(x_0))^{-2} C = ∥ f ∥ L 1 ( m ) ( h 1 f ( x 0 ) ) − 2 . A quantitative upper bound on C C C requires a lower bound on h t f ( x 0 ) h_tf(x_0) h t f ( x 0 ) , which can be obtained from the short-time estimate ∥ h t f − f ∥ L 2 ( m ) 2 ≤ t ∥ D f ∥ L 2 ( m ) 2 \|h_tf-f\|_{L^2(\mathfrak{m})}^2\le t\|Df\|_{L^2(\mathfrak{m})}^2 ∥ h t f − f ∥ L 2 ( m ) 2 ≤ t ∥ D f ∥ L 2 ( m ) 2 combined with the Harnack inequality of Lemma 4.14; this brings in m ( B r 0 / 2 ( x 0 ) ) \mathfrak{m}(B_{r_0/2}(x_0)) m ( B r 0 /2 ( x 0 )) and makes the choice of t t t (hence c c c ) depend on K K K , r 0 r_0 r 0 and the ratio m ( B r 0 ( x 0 ) ) / m ( B r 0 / 2 ( x 0 ) ) \mathfrak{m}(B_{r_0}(x_0))/\mathfrak{m}(B_{r_0/2}(x_0)) m ( B r 0 ( x 0 )) / m ( B r 0 /2 ( x 0 )) .
Strategy of the proof and related results
The proof follows a route that is classical for smooth weighted manifolds; the point is to check that, in the present nonsmooth setting, none of its steps requires the growth of m \mathfrak{m} m , nor its finiteness on balls.
The starting point is the L 2 L^2 L 2 Bakry–Émery estimate ∣ D h t f ∣ 2 ≤ e − 2 K t h t ( ∣ D f ∣ 2 ) |Dh_tf|^2\le e^{-2Kt}h_t(|Df|^2) ∣ D h t f ∣ 2 ≤ e − 2 K t h t ( ∣ D f ∣ 2 ) , which is proved in [5 Cor. 2.3] starting from the weak Bochner inequality; its proof uses neither (G) nor the finiteness of m \mathfrak{m} m on balls (Lemma 2.5).
The second ingredient is the L 1 L^{1} L 1 Bakry–Émery estimate ∣ D h t f ∣ ≤ e − K t h t ( ∣ D f ∣ ) \lvert Dh_{t}f\rvert\le e^{-Kt}h_{t}(\lvert Df\rvert) ∣ D h t f ∣ ≤ e − K t h t (∣ D f ∣) (Proposition 4.13), due to Bakry for smooth diffusions and to Savaré [25 ] for abstract Dirichlet forms. We follow [25 ] , but without requiring that the heat flow preserves mass. Savaré uses this assumption to construct the iterated carré du champ as a finite measure [25 Lemma 2.6] ; we represent instead the nonnegative functional Γ 2 K [ f ] \Gamma_{2}^{K}[f] Γ 2 K [ f ] by a locally finite measure (Lemma 4.9 ). Moreover, the quasi-regularity of the Cheeger energy, proved in [25 Thm. 4.1] by exhausting X X X with balls, is obtained by exhausting X X X with superlevel sets of test functions (Lemma 4.1 ).
The third ingredient is a dimension-free Harnack inequality (Lemma 4.14): for 0 ≤ f ≤ 1 / 2 0 \le f \le1/2 0 ≤ f ≤ 1/2 ,
( h t f ( x ) ) 2 ≤ h t f ( y ) exp ( d ( x , y ) 2 2 I K ( t ) ) \left(h_{t}f(x)\right)^{2} \le h_{t}f(y)\exp\left(\frac{d(x,y)^{2}}{2I_{K}(t)}\right) ( h t f ( x ) ) 2 ≤ h t f ( y ) exp ( 2 I K ( t ) d ( x , y ) 2 )
for all x , y ∈ X x,y \in X x , y ∈ X .
Wang [28 ] proved on Riemannian manifolds the stronger inequality
( h t f ( x ) ) 2 ≤ h t ( f 2 ) ( y ) exp ( d ( x , y ) 2 2 I K ( t ) ) , \left(h_{t}f(x)\right)^{2} \le h_{t}(f^{2})(y)\exp\left(\frac{d(x,y)^{2}}{2I_{K}(t)}\right), ( h t f ( x ) ) 2 ≤ h t ( f 2 ) ( y ) exp ( 2 I K ( t ) d ( x , y ) 2 ) ,
and Li [21 ] extended it to RCD( K , ∞ ) (K,\infty) ( K , ∞ ) spaces defined via optimal transport. Li’s proof uses the W 2 W_{2} W 2 -contractivity of the heat flow, which relies on the identification of the heat flow with the Wasserstein gradient flow of the entropy and is not available under (S). This is why we prove the weaker inequality above: it only requires the L 1 L^{1} L 1 Bakry–Émery estimate and the Sobolev-to-Lipschitz property, and it suffices to bound the volume of balls.
The L 1 L^{1} L 1 Bakry–Émery estimate is essential here. Indeed, for every δ ∈ ( 0 , 1 / 4 ) \delta\in(0,1/4) δ ∈ ( 0 , 1/4 ) , the proof of Lemma 4.14 shows that − log ( h t f + δ ) \sqrt{-\log(h_{t}f+\delta)} − log ( h t f + δ ) is ( 2 I K ( t ) ) − 1 (2\sqrt{I_{K}(t)})^{-1} ( 2 I K ( t ) ) − 1 -Lipschitz; this bounds h t f ( y ) h_{t}f(y) h t f ( y ) from below at every y ∈ X y \in X y ∈ X in terms of h t f ( x ) h_{t}f(x) h t f ( x ) and d ( x , y ) d(x,y) d ( x , y ) . The key step is the pointwise bound
∣ D h t f ∣ 2 ≤ e − 2 K s ( h t f + δ ) h s ( ∣ D h t − s f ∣ 2 h t − s f + δ ) , \lvert Dh_{t}f\rvert^{2} \le e^{-2Ks}(h_{t}f+\delta)h_{s}\left(\frac{\lvert Dh_{t-s}f\rvert^{2}}{h_{t-s}f+\delta}\right), ∣ D h t f ∣ 2 ≤ e − 2 K s ( h t f + δ ) h s ( h t − s f + δ ∣ D h t − s f ∣ 2 ) ,
which follows from ∣ D h t f ∣ ≤ e − K s h s ( ∣ D h t − s f ∣ ) \lvert Dh_{t}f\rvert\le e^{-Ks}h_{s}(\lvert Dh_{t-s}f\rvert) ∣ D h t f ∣ ≤ e − K s h s (∣ D h t − s f ∣) and the Cauchy–Schwarz inequality for h s h_{s} h s . The L 2 L^{2} L 2 Bakry–Émery estimate ∣ D h t f ∣ 2 ≤ e − 2 K s h s ( ∣ D h t − s f ∣ 2 ) \lvert Dh_{t}f\rvert^{2} \le e^{-2Ks}h_{s}(\lvert Dh_{t-s}f\rvert^{2}) ∣ D h t f ∣ 2 ≤ e − 2 K s h s (∣ D h t − s f ∣ 2 ) only gives the same bound with the factor h t f + δ h_{t}f+\delta h t f + δ replaced by the constant ∥ h t − s f + δ ∥ L ∞ ( m ) ≤ 1 / 2 + δ \lVert h_{t-s}f+\delta\rVert_{L^{\infty}(m)} \le1/2+\delta ∥ h t − s f + δ ∥ L ∞ ( m ) ≤ 1/2 + δ . Following the rest of the argument, this yields that G ( h t f + δ ) G(h_{t}f+\delta) G ( h t f + δ ) is I K ( t ) − 1 / 2 I_{K}(t)^{-1/2} I K ( t ) − 1/2 -Lipschitz, where G ( u ) : = ∫ 0 u ( v log ( 1 / v ) ) − 1 / 2 d v G(u):=\int_{0}^{u}(v\log(1/v))^{-1/2}\,dv G ( u ) := ∫ 0 u ( v log ( 1/ v ) ) − 1/2 d v for u ∈ [ 0 , 3 / 4 ] u \in[0,3/4] u ∈ [ 0 , 3/4 ] . Since G G G is bounded, such an estimate gives no information on h t f ( y ) h_{t}f(y) h t f ( y ) once d ( x , y ) ≥ I K ( t ) 1 / 2 G ( 3 / 4 ) d(x,y) \ge I_{K}(t)^{1/2}G(3/4) d ( x , y ) ≥ I K ( t ) 1/2 G ( 3/4 ) , and therefore no lower bound at every point.
Finally, Theorem 1.4 follows by integrating the Harnack inequality in y y y over B r ( x 0 ) B_{r}(x_{0}) B r ( x 0 ) and using the L 1 L^{1} L 1 -contractivity of h t h_{t} h t .
For smooth weighted manifolds ( M , g , e − V vol g ) (M,g,e^{-V}\operatorname{vol}_{g}) ( M , g , e − V vol g ) whose Bakry–Émery Ricci tensor satisfies
Ric V : = Ric g + Hess g V ≥ K , \operatorname{Ric}_{V}:=\operatorname{Ric}_{g}+\operatorname{Hess}_{g}V \ge K, Ric V := Ric g + Hess g V ≥ K ,
with no assumption on the potential V V V , the Gaussian growth ( e − V vol g ) ( B r ( x 0 ) ) ≤ C e c r 2 (e^{-V}\operatorname{vol}_{g})(B_{r}(x_{0})) \le Ce^{cr^{2}} ( e − V vol g ) ( B r ( x 0 )) ≤ C e c r 2 is due to Morgan [24 ] and Wei–Wylie [WW09, Thm. 3.1], and it is sharp on the Gaussian soliton. Their proof relies on comparison geometry along geodesics, which has no direct counterpart in the Eulerian nonsmooth setting; the argument above replaces it by the heat flow. A discrete counterpart, for graphs of bounded degree satisfying the dimension-free Bakry–Émery condition with K = 0 K=0 K = 0 , has recently been obtained in [20 ] .
Preliminaries
Radon measures
We will use the following elementary lemma.
Lemma 2.1 (Local finiteness). Let ( X , d ) (X,\mathrm{d}) ( X , d ) be a metric space and let m \mathfrak{m} m be a Radon measure on X X X . Then every x ∈ X x \in X x ∈ X has a neighbourhood of finite measure. If moreover ( X , d ) (X,\mathrm{d}) ( X , d ) is separable, then there are x i ∈ X x_{i} \in X x i ∈ X and r i > 0 r_{i}>0 r i > 0 , i ∈ N i \in\mathbb{N} i ∈ N , such that m ( B 3 r i ( x i ) ) < ∞ \mathfrak{m}(B_{3r_{i}}(x_{i}))<\infty m ( B 3 r i ( x i )) < ∞ for every i i i and X = ⋃ i B r i ( x i ) X=\bigcup_{i}B_{r_{i}}(x_{i}) X = ⋃ i B r i ( x i ) ; in particular, m \mathfrak{m} m is σ \sigma σ -finite and L 2 ( m ) L^{2}(\mathfrak{m}) L 2 ( m ) is separable.
Proof. Let x ∈ X x \in X x ∈ X . Since { x } \{x\} { x } is compact, m ( { x } ) < ∞ \mathfrak{m}(\{x\})<\infty m ({ x }) < ∞ , and by outer regularity there is an open set U ∋ x U \ni x U ∋ x with m ( U ) < ∞ \mathfrak{m}(U)<\infty m ( U ) < ∞ ; U U U contains a ball centred at x x x . If ( X , d ) (X,\mathrm{d}) ( X , d ) is separable, it is Lindelöf, so the open cover { B r x ( x ) } x ∈ X \{B_{r_{x}}(x)\}_{x\in X} { B r x ( x ) } x ∈ X , where r x > 0 r_{x}>0 r x > 0 is such that m ( B 3 r x ( x ) ) < ∞ \mathfrak{m}(B_{3r_{x}}(x))<\infty m ( B 3 r x ( x )) < ∞ , has a countable subcover. The separability of L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) follows from the σ \sigma σ -finiteness of m \mathfrak{m} m and the separability of ( X , d ) (X,d) ( X , d ) . □
Under (S), we call a ball B r ( x ) B_r(x) B r ( x ) admissible if m ( B 3 r ( x ) ) < ∞ \mathfrak{m}(B_{3r}(x)) < \infty m ( B 3 r ( x )) < ∞ . By Lemma 2.1 , every point is the centre of admissible balls, and X X X is covered by countably many admissible balls.
Remark 2.2. On a complete and separable metric space, a Borel measure is a Radon measure if and only if every point has a neighbourhood of finite measure. One implication is Lemma 2.1 . Conversely, if X X X is covered by countably many open sets U i U_i U i of finite measure, then m \mathfrak{m} m is finite on compact sets, and outer regularity and inner regularity on open sets follow from the corresponding properties of the finite measures m ↾ U i \mathfrak{m}\mathord{\upharpoonright} U_i m ↾ U i , the latter by Ulam’s theorem.
The finiteness on compact sets alone is not enough for Theorem 1.4 . Let X X X be a separable infinite-dimensional Hilbert space. Compact subsets of X X X have empty interior, so σ \sigma σ -compact subsets of X X X are meagre, and no nonempty open set is contained in a σ \sigma σ -compact set by the Baire category theorem. Set m ( B ) : = 0 \mathfrak{m}(B) := 0 m ( B ) := 0 if the Borel set B B B is contained in a σ \sigma σ -compact set and m ( B ) : = ∞ \mathfrak{m}(B) := \infty m ( B ) := ∞ otherwise. Then m \mathfrak{m} m is a Borel measure (if one of the disjoint sets B k B_k B k is not contained in a σ \sigma σ -compact set, neither is ⋃ k B k \bigcup_k B_k ⋃ k B k ), it vanishes on compact sets, and m ( U ) = ∞ \mathfrak{m}(U) = \infty m ( U ) = ∞ for every nonempty open set U U U , so supp ( m ) = X \operatorname{supp}(\mathfrak{m}) = X supp ( m ) = X . Every f ∈ L 2 ( m ) f \in L^2(\mathfrak{m}) f ∈ L 2 ( m ) vanishes m \mathfrak{m} m -a.e., since the sets { ∣ f ∣ > 1 / k } \{|f| > 1/k\} { ∣ f ∣ > 1/ k } have finite, hence zero, measure. Thus W 1 , 2 ( X ) = { 0 } W^{1,2}(X) = \{0\} W 1 , 2 ( X ) = { 0 } , and (SL), (IH) and (BE) hold trivially, while every ball has infinite measure. This measure is neither outer regular, since m ( { 0 } ) = 0 \mathfrak{m}(\{0\}) = 0 m ({ 0 }) = 0 while every open set containing 0 0 0 has infinite measure, nor inner regular on open sets.
We denote by LIP f m ( X ) \operatorname{LIP}_{\mathrm{fm}}(X) LIP fm ( X ) the space of bounded Lipschitz functions f : X → R f : X \to\mathbb{R} f : X → R such that m ( supp ( f ) ) < ∞ \mathfrak{m}(\operatorname{supp}(f)) < \infty m ( supp ( f )) < ∞ . Under (S), this space replaces the space LIP b s ( X ) \operatorname{LIP}_{\mathrm{bs}}(X) LIP bs ( X ) of Lipschitz functions with bounded support, which in general is not contained in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) .
Lemma 2.3. Assume that ( X , d ) (X,d) ( X , d ) is complete and separable and that m \mathfrak{m} m is a Radon measure on X X X . Then:
(i) LIP f m ( X ) ⊆ W 1 , 2 ( X ) \operatorname{LIP}_{\mathrm{fm}}(X) \subseteq W^{1,2}(X) LIP fm ( X ) ⊆ W 1 , 2 ( X ) , and ∣ D f ∣ ≤ lip ( f ) ≤ Lip ( f ) χ supp ( f ) |Df| \le\operatorname{lip}(f) \le\operatorname{Lip}(f)\chi_{\operatorname{supp}(f)} ∣ D f ∣ ≤ lip ( f ) ≤ Lip ( f ) χ supp ( f ) m \mathfrak{m} m -a.e. for every f ∈ LIP f m ( X ) f \in\operatorname{LIP}_{\mathrm{fm}}(X) f ∈ LIP fm ( X ) ;
(ii) if B r ( x ) B_r(x) B r ( x ) is admissible, then β x , r : = ( 1 − d ( ⋅ , B r ( x ) ) / r ) + \beta_{x,r} := (1-d(\cdot,B_r(x))/r)^+ β x , r := ( 1 − d ( ⋅ , B r ( x )) / r ) + belongs to LIP f m ( X ) \operatorname{LIP}_{\mathrm{fm}}(X) LIP fm ( X ) , 0 ≤ β x , r ≤ 1 0 \le\beta_{x,r} \le1 0 ≤ β x , r ≤ 1 , β x , r = 1 \beta_{x,r} = 1 β x , r = 1 on B r ( x ) B_r(x) B r ( x ) and β x , r = 0 \beta_{x,r} = 0 β x , r = 0 outside B 2 r ( x ) B_{2r}(x) B 2 r ( x ) ;
(iii) LIP f m ( X ) \operatorname{LIP}_{\mathrm{fm}}(X) LIP fm ( X ) is dense in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) .
Proof. (i). Since f f f is bounded and m ( supp ( f ) ) < ∞ \mathfrak{m}(\operatorname{supp}(f)) < \infty m ( supp ( f )) < ∞ , f ∈ L 2 ( m ) f \in L^2(\mathfrak{m}) f ∈ L 2 ( m ) . The local Lipschitz constant lip ( f ) \operatorname{lip}(f) lip ( f ) vanishes on the open set X ∖ supp ( f ) X \setminus\operatorname{supp}(f) X ∖ supp ( f ) and is bounded by Lip ( f ) \operatorname{Lip}(f) Lip ( f ) , so lip ( f ) ∈ L 2 ( m ) \operatorname{lip}(f) \in L^2(\mathfrak{m}) lip ( f ) ∈ L 2 ( m ) . Moreover, lip ( f ) \operatorname{lip}(f) lip ( f ) is an upper gradient of f f f : for every absolutely continuous curve γ : [ 0 , 1 ] → X \gamma: [0,1] \to X γ : [ 0 , 1 ] → X , the function t ↦ f ( γ t ) t \mapsto f(\gamma_t) t ↦ f ( γ t ) is absolutely continuous and ∣ f ( γ 1 ) − f ( γ 0 ) ∣ ≤ ∫ 0 1 lip ( f ) ( γ t ) ∣ γ ˙ t ∣ d t |f(\gamma_1)-f(\gamma_0)| \le\int_0^1 \operatorname{lip}(f)(\gamma_t)|\dot{\gamma}_t|\,dt ∣ f ( γ 1 ) − f ( γ 0 ) ∣ ≤ ∫ 0 1 lip ( f ) ( γ t ) ∣ γ ˙ t ∣ d t . In particular, lip ( f ) \operatorname{lip}(f) lip ( f ) is a weak upper gradient of f f f , so f ∈ W 1 , 2 ( X ) f \in W^{1,2}(X) f ∈ W 1 , 2 ( X ) and ∣ D f ∣ ≤ lip ( f ) |Df| \le\operatorname{lip}(f) ∣ D f ∣ ≤ lip ( f ) m \mathfrak{m} m -a.e.
(ii). β x , r \beta_{x,r} β x , r is r − 1 r^{-1} r − 1 -Lipschitz, and its support is contained in B 2 r ( x ) ⊆ B 3 r ( x ) B_{2r}(x) \subseteq B_{3r}(x) B 2 r ( x ) ⊆ B 3 r ( x ) .
(iii). Let ( B r i ( x i ) ) i ∈ N (B_{r_i}(x_i))_{i\in\mathbb{N}} ( B r i ( x i ) ) i ∈ N be admissible balls covering X X X , and set E n : = ⋃ i ≤ n B r i ( x i ) E_n := \bigcup_{i\le n} B_{r_i}(x_i) E n := ⋃ i ≤ n B r i ( x i ) . Then E n E_n E n is open, E n ↑ X E_n \uparrow X E n ↑ X , and m ( E n ‾ ) ≤ ∑ i ≤ n m ( B 3 r i ( x i ) ) < ∞ \mathfrak{m}(\overline{E_n}) \le\sum_{i\le n} \mathfrak{m}(B_{3r_i}(x_i)) < \infty m ( E n ) ≤ ∑ i ≤ n m ( B 3 r i ( x i )) < ∞ . Let g ∈ L 2 ( m ) g \in L^2(\mathfrak{m}) g ∈ L 2 ( m ) and ε > 0 \varepsilon> 0 ε > 0 . By dominated convergence there is n n n with ∥ g − g χ E n ∥ L 2 ( m ) < ε \|g-g\chi_{E_n}\|_{L^2(\mathfrak{m})} < \varepsilon ∥ g − g χ E n ∥ L 2 ( m ) < ε . Since bounded Lipschitz functions are dense in L 2 ( ν ) L^2(\nu) L 2 ( ν ) for every finite Borel measure ν \nu ν on a metric space, applying this to ν : = m ↾ E n \nu:= \mathfrak{m}\mathord{\upharpoonright} E_n ν := m ↾ E n we find a bounded Lipschitz function u u u with ∥ u − g ∥ L 2 ( m ↾ E n ) < ε \|u-g\|_{L^2(\mathfrak{m}\mathord{\upharpoonright} E_n)} < \varepsilon ∥ u − g ∥ L 2 ( m ↾ E n ) < ε . The functions ζ k : = min { 1 , k d ( ⋅ , X ∖ E n ) } \zeta_k := \min\{1,k\,d(\cdot,X\setminus E_n)\} ζ k := min { 1 , k d ( ⋅ , X ∖ E n )} are Lipschitz, vanish outside E n E_n E n and satisfy ζ k ↑ χ E n \zeta_k \uparrow\chi_{E_n} ζ k ↑ χ E n , so ζ k u ∈ LIP f m ( X ) \zeta_k u \in\operatorname{LIP}_{\mathrm{fm}}(X) ζ k u ∈ LIP fm ( X ) , because supp ( ζ k u ) ⊆ E n ‾ \operatorname{supp}(\zeta_k u) \subseteq\overline{E_n} supp ( ζ k u ) ⊆ E n , and ∥ ζ k u − u ∥ L 2 ( m ↾ E n ) → 0 \|\zeta_k u-u\|_{L^2(\mathfrak{m}\mathord{\upharpoonright} E_n)} \to0 ∥ ζ k u − u ∥ L 2 ( m ↾ E n ) → 0 by dominated convergence. Hence ∥ ζ k u − g ∥ L 2 ( m ) < 3 ε \|\zeta_k u-g\|_{L^2(\mathfrak{m})} < 3\varepsilon ∥ ζ k u − g ∥ L 2 ( m ) < 3 ε for k k k large. □
Calculus and heat flow
Under (S), the Sobolev space W 1 , 2 ( X ) W^{1,2}(X) W 1 , 2 ( X ) and the minimal weak upper gradient ∣ D f ∣ |Df| ∣ D f ∣ are defined via test plans; under (IH), ( f , g ) ↦ ∇ f ⋅ ∇ g ∈ L 1 ( m ) (f,g) \mapsto\nabla f \cdot\nabla g \in L^1(\mathfrak{m}) ( f , g ) ↦ ∇ f ⋅ ∇ g ∈ L 1 ( m ) is the pointwise scalar product on the Hilbert tangent module L 2 ( T X ) L^2(TX) L 2 ( T X ) , Δ : D ( Δ ) ⊆ W 1 , 2 ( X ) → L 2 ( m ) \Delta: D(\Delta) \subseteq W^{1,2}(X) \to L^2(\mathfrak{m}) Δ : D ( Δ ) ⊆ W 1 , 2 ( X ) → L 2 ( m ) is the Laplacian associated with C h \mathrm{Ch} Ch , and ( h t ) t ≥ 0 (h_t)_{t\ge0} ( h t ) t ≥ 0 is the heat flow, i.e. the gradient flow of C h \mathrm{Ch} Ch in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) .
Proposition 2.4. Let ( X , d ) (X,d) ( X , d ) be a complete and separable metric space, let m \mathfrak{m} m be a Radon measure on X X X with supp ( m ) = X \operatorname{supp}(\mathfrak{m}) = X supp ( m ) = X , and assume (IH). Then:
(a) Calculus rules: the minimal weak upper gradient is local ; for f ∈ W 1 , 2 ( X ) f \in W^{1,2}(X) f ∈ W 1 , 2 ( X ) and η : R → R \eta:\mathbb{R}\to\mathbb{R} η : R → R Lipschitz with η ( 0 ) = 0 \eta(0)=0 η ( 0 ) = 0 , η ∘ f ∈ W 1 , 2 ( X ) \eta\circ f\in W^{1,2}(X) η ∘ f ∈ W 1 , 2 ( X ) and ∣ D ( η ∘ f ) ∣ = ∣ η ′ ∣ ∘ f ∣ D f ∣ |D(\eta\circ f)|=|\eta'|\circ f\,|Df| ∣ D ( η ∘ f ) ∣ = ∣ η ′ ∣ ∘ f ∣ D f ∣ m-a.e.; ∇ f ⋅ ∇ g \nabla f\cdot\nabla g ∇ f ⋅ ∇ g is bilinear, symmetric and continuous from W 1 , 2 ( X ) × W 1 , 2 ( X ) W^{1,2}(X)\times W^{1,2}(X) W 1 , 2 ( X ) × W 1 , 2 ( X ) to L 1 ( m ) L^1(\mathfrak{m}) L 1 ( m ) , and satisfies the chain rule and the Leibniz rule on W 1 , 2 ( X ) ∩ L ∞ ( m ) W^{1,2}(X)\cap L^\infty(\mathfrak{m}) W 1 , 2 ( X ) ∩ L ∞ ( m ) ; if f n → f f_n\to f f n → f in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) , f n → f f_n\to f f n → f m-a.e. and ∣ D f n ∣ ⇀ G |Df_n|\rightharpoonup G ∣ D f n ∣ ⇀ G weakly in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) , then G G G is a weak upper gradient of f f f .
(b) Laplacian: Δ \Delta Δ is self-adjoint, and
∫ X g Δ f d m = − ∫ X ∇ f ⋅ ∇ g d m for f ∈ D ( Δ ) , g ∈ W 1 , 2 ( X ) ; \int_X g\,\Delta f\,\mathrm{d}\mathfrak{m}=-\int_X\nabla f\cdot\nabla g\,\mathrm{d}\mathfrak{m}\qquad\text{for }f\in D(\Delta),\ g\in W^{1,2}(X); ∫ X g Δ f d m = − ∫ X ∇ f ⋅ ∇ g d m for f ∈ D ( Δ ) , g ∈ W 1 , 2 ( X ) ;
for f , g ∈ LIP b ( X ) ∩ D ( Δ ) f,g\in\operatorname{LIP}_{b}(X)\cap D(\Delta) f , g ∈ LIP b ( X ) ∩ D ( Δ ) and ϕ ∈ C 2 ( R ) \phi\in C^2(\mathbb{R}) ϕ ∈ C 2 ( R ) with ϕ ′ ′ \phi'' ϕ ′′ bounded,
Δ ( f g ) = f Δ g + g Δ f + 2 ∇ f ⋅ ∇ g , Δ ( ϕ ∘ f ) = ϕ ′ ∘ f Δ f + ϕ ′ ′ ∘ f ∣ D f ∣ 2 . \Delta(fg)=f\Delta g+g\Delta f+2\nabla f\cdot\nabla g,\qquad\Delta(\phi\circ f)=\phi'\circ f\,\Delta f+\phi''\circ f\,|Df|^2. Δ ( f g ) = f Δ g + g Δ f + 2∇ f ⋅ ∇ g , Δ ( ϕ ∘ f ) = ϕ ′ ∘ f Δ f + ϕ ′′ ∘ f ∣ D f ∣ 2 .
(c) Heat flow: h t h_t h t is linear and self-adjoint, Ch ( h t f ) ≤ Ch ( f ) \operatorname{Ch}(h_tf)\leq\operatorname{Ch}(f) Ch ( h t f ) ≤ Ch ( f ) , h t Δ f = Δ h t f h_t\Delta f=\Delta h_tf h t Δ f = Δ h t f for f ∈ D ( Δ ) f\in D(\Delta) f ∈ D ( Δ ) ; for f ∈ L 2 ( m ) f\in L^2(\mathfrak{m}) f ∈ L 2 ( m ) , t ↦ h t f t\mapsto h_tf t ↦ h t f belongs to C ( [ 0 , ∞ ) ; L 2 ( m ) ) ∩ C ∞ ( ( 0 , ∞ ) ; W 1 , 2 ( X ) ) C([0,\infty);L^2(\mathfrak{m}))\cap C^\infty((0,\infty);W^{1,2}(X)) C ([ 0 , ∞ ) ; L 2 ( m )) ∩ C ∞ (( 0 , ∞ ) ; W 1 , 2 ( X )) , with h t f ∈ D ( Δ ) h_tf\in D(\Delta) h t f ∈ D ( Δ ) and d d t h t f = Δ h t f \frac{\mathrm{d}}{\mathrm{d}t}h_tf=\Delta h_tf d t d h t f = Δ h t f for t > 0 t>0 t > 0 ; ( h t ) t ≥ 0 (h_t)_{t\geq0} ( h t ) t ≥ 0 is positivity preserving, satisfies the weak maximum principle, and extends to a semigroup of contractions on L p ( m ) L^p(\mathfrak{m}) L p ( m ) for every p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] .
Proof. Originally proved in [4 , 15 ] , all these statements are also presented in [18 Chapters 2, 4, and 5] under the standing assumption (1.21) therein that m \mathfrak{m} m is finite on bounded sets; we shall follow the latter presentation throughout. In Chapters 1–5 of [18 ] , this assumption is used explicitly only in Prop. 1.1.18 (density of LIP b s ( X ) \operatorname{LIP}_{bs}(X) LIP b s ( X ) in L p L^p L p ), Prop. 1.1.19 (the topology of L p ( m ) L^p(\mathfrak{m}) L p ( m ) , where σ \sigma σ -finiteness suffices), Prop. 2.1.26 (LIP b s ( X ) ⊆ W 1 , 2 ( X ) \operatorname{LIP}_{bs}(X)\subseteq W^{1,2}(X) LIP b s ( X ) ⊆ W 1 , 2 ( X ) ) and Thm. 2.2.28 (density in energy of Lipschitz functions). In our setting, m \mathfrak{m} m is σ \sigma σ -finite by Lemma 2.1 , and wherever the inclusion LIP b s ( X ) ⊆ W 1 , 2 ( X ) \operatorname{LIP}_{bs}(X)\subseteq W^{1,2}(X) LIP b s ( X ) ⊆ W 1 , 2 ( X ) or the density of LIP b s ( X ) \operatorname{LIP}_{bs}(X) LIP b s ( X ) in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) is used, Lemma 2.3 provides the same statements for LIP f m ( X ) \operatorname{LIP}_{fm}(X) LIP f m ( X ) ; in particular W 1 , 2 ( X ) W^{1,2}(X) W 1 , 2 ( X ) is dense in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) , which is needed for the heat flow to be defined on the whole L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) . Thm. 2.2.28 is not used in this note. □
Growth-free results on the heat flow
The results of this subsection are proved in [5 , 25 ] . Following the proofs collected in [18 Chapters 1–6] for RCD( K , ∞ ) (K,\infty) ( K , ∞ ) spaces, one can check that they only use the results of Chapters 1–5, which hold in our setting by Proposition 2.4 , together with (SL), (IH) and (BE). In particular, they use neither (G) nor the finiteness of m \mathfrak{m} m on balls of arbitrary radius.
We use the space of test functions
Test ∞ ( X ) : = { f ∈ LIP ( X ) ∩ L ∞ ( m ) ∩ D ( Δ ) : Δ f ∈ W 1 , 2 ( X ) ∩ L ∞ ( m ) } , \operatorname{Test}^{\infty}(X):=\{f\in\operatorname{LIP}(X)\cap L^\infty(\mathfrak{m})\cap D(\Delta):\Delta f\in W^{1,2}(X)\cap L^\infty(\mathfrak{m})\}, Test ∞ ( X ) := { f ∈ LIP ( X ) ∩ L ∞ ( m ) ∩ D ( Δ ) : Δ f ∈ W 1 , 2 ( X ) ∩ L ∞ ( m )} ,
its subset Test + ∞ ( X ) \operatorname{Test}^{\infty}_{+}(X) Test + ∞ ( X ) of m-a.e. nonnegative functions, and, for f ∈ L 2 ( m ) f\in L^2(\mathfrak{m}) f ∈ L 2 ( m ) and φ ∈ C c ∞ ( 0 , ∞ ) \varphi\in C^\infty_c(0,\infty) φ ∈ C c ∞ ( 0 , ∞ ) , the mollified heat flow
h φ f : = ∫ 0 ∞ h s f φ ( s ) d s . h_\varphi f:=\int_0^\infty h_s f\,\varphi(s)\,\mathrm{d}s. h φ f := ∫ 0 ∞ h s f φ ( s ) d s .
Elements of Test ∞ ( X ) \operatorname{Test}^{\infty}(X) Test ∞ ( X ) are identified with their Lipschitz representatives, which are unique since supp ( m ) = X \operatorname{supp}(\mathfrak{m})=X supp ( m ) = X .
Lemma 2.5 (Growth-free results). Assume (S). Then:
(i) for every f ∈ W 1 , 2 ( X ) f\in W^{1,2}(X) f ∈ W 1 , 2 ( X ) and every t ≥ 0 t\geq0 t ≥ 0 ,
∣ D h t f ∣ 2 ≤ e − 2 K t h t ( ∣ D f ∣ 2 ) m-a.e. in X ; (4) |D h_t f|^2\leq e^{-2Kt}h_t(|Df|^2)\qquad\text{m-a.e. in }X; \tag*{(4)} ∣ D h t f ∣ 2 ≤ e − 2 K t h t ( ∣ D f ∣ 2 ) m-a.e. in X ; (4) (ii) for every f ∈ L 2 ( m ) ∩ L ∞ ( m ) f\in L^2(\mathfrak{m})\cap L^\infty(\mathfrak{m}) f ∈ L 2 ( m ) ∩ L ∞ ( m ) and every t > 0 t>0 t > 0 , ∣ D h t f ∣ ∈ L ∞ ( m ) |D h_t f|\in L^\infty(\mathfrak{m}) ∣ D h t f ∣ ∈ L ∞ ( m ) and h t f h_t f h t f admits a Lipschitz representative;
(iii) for every f ∈ L 2 ( m ) ∩ L ∞ ( m ) f\in L^2(\mathfrak{m})\cap L^\infty(\mathfrak{m}) f ∈ L 2 ( m ) ∩ L ∞ ( m ) and every φ ∈ C c ∞ ( 0 , ∞ ) \varphi\in C^\infty_c(0,\infty) φ ∈ C c ∞ ( 0 , ∞ ) , h φ f ∈ Test ∞ ( X ) h_\varphi f\in\operatorname{Test}^{\infty}(X) h φ f ∈ Test ∞ ( X ) ;
(iv) Test + ∞ ( X ) \operatorname{Test}^{\infty}_{+}(X) Test + ∞ ( X ) is dense in W 1 , 2 ( X ) + W^{1,2}(X)_{+} W 1 , 2 ( X ) + and Test ∞ ( X ) \operatorname{Test}^{\infty}(X) Test ∞ ( X ) is dense in W 1 , 2 ( X ) W^{1,2}(X) W 1 , 2 ( X ) ;
(v) ∣ D f ∣ 2 ∈ W 1 , 2 ( X ) |Df|^2\in W^{1,2}(X) ∣ D f ∣ 2 ∈ W 1 , 2 ( X ) for every f ∈ Test ∞ ( X ) f\in\operatorname{Test}^{\infty}(X) f ∈ Test ∞ ( X ) ;
(vi) Test ∞ ( X ) \operatorname{Test}^{\infty}(X) Test ∞ ( X ) is an algebra.
Dirichlet forms
We collect the notions from the theory of Dirichlet forms used in § 4 , following [25 Sections 2–3] and [5 Section 2] . Throughout this section, ( X , τ ) (X,\tau) ( X , τ ) is a Polish space, m \mathfrak{m} m is a σ \sigma σ -finite Borel measure on X X X with supp ( m ) = X \operatorname{supp}(\mathfrak{m})=X supp ( m ) = X , and functions are identified when they agree m \mathfrak{m} m -a.e.
Definition 3.1 (Symmetric Dirichlet form, generator, semigroup). A symmetric Dirichlet form on L 2 ( X , m ) L^2(X,\mathfrak{m}) L 2 ( X , m ) is a lower semicontinuous quadratic form E : L 2 ( X , m ) → [ 0 , ∞ ] \mathcal{E}:L^2(X,\mathfrak{m})\to[0,\infty] E : L 2 ( X , m ) → [ 0 , ∞ ] whose domain V : = { f ∈ L 2 ( X , m ) : E ( f ) < ∞ } \mathbb{V}:=\{f\in L^2(X,\mathfrak{m}):\mathcal{E}(f)<\infty\} V := { f ∈ L 2 ( X , m ) : E ( f ) < ∞ } is dense in L 2 ( X , m ) L^2(X,\mathfrak{m}) L 2 ( X , m ) and which satisfies the Markov property
E ( η ∘ f ) ≤ E ( f ) \mathcal{E}(\eta\circ f)\leq\mathcal{E}(f) E ( η ∘ f ) ≤ E ( f ) for every f ∈ L 2 ( X , m ) f\in L^2(X,\mathfrak{m}) f ∈ L 2 ( X , m ) and every 1 1 1 -Lipschitz η : R → R \eta:\mathbb{R}\to\mathbb{R} η : R → R with η ( 0 ) = 0 \eta(0)=0 η ( 0 ) = 0 . We denote by E ( f , g ) : = 1 4 ( E ( f + g ) − E ( f − g ) ) \mathcal{E}(f,g):=\frac{1}{4}\bigl(\mathcal{E}(f+g)-\mathcal{E}(f-g)\bigr) E ( f , g ) := 4 1 ( E ( f + g ) − E ( f − g ) ) the associated symmetric bilinear form on V \mathbb{V} V , by ∥ f ∥ V 2 : = ∥ f ∥ L 2 ( m ) 2 + E ( f ) \|f\|_{\mathbb{V}}^2:=\|f\|_{L^2(\mathfrak{m})}^2+\mathcal{E}(f) ∥ f ∥ V 2 := ∥ f ∥ L 2 ( m ) 2 + E ( f ) the Hilbert norm of V \mathbb{V} V , and set V ∞ : = V ∩ L ∞ ( m ) \mathbb{V}_\infty:=\mathbb{V}\cap L^\infty(\mathfrak{m}) V ∞ := V ∩ L ∞ ( m ) and V + : = { f ∈ V : f ≥ 0 } \mathbb{V}_+:=\{f\in\mathbb{V}:f\geq0\} V + := { f ∈ V : f ≥ 0 } . The generator of E \mathcal{E} E is the nonpositive self-adjoint operator L : D ( L ) ⊆ V → L 2 ( X , m ) L:D(L)\subseteq\mathbb{V}\to L^2(X,\mathfrak{m}) L : D ( L ) ⊆ V → L 2 ( X , m ) characterised by
f ∈ D ( L ) , L f = h ⟺ f ∈ V and E ( f , g ) = − ∫ X h g d m for every g ∈ V , f\in D(L),\ Lf=h\Longleftrightarrow f\in\mathbb{V}\ \text{and}\ \mathcal{E}(f,g)=-\int_X hg\,\mathrm{d}\mathfrak{m}\quad\text{for every }g\in\mathbb{V}, f ∈ D ( L ) , L f = h ⟺ f ∈ V and E ( f , g ) = − ∫ X h g d m for every g ∈ V , and ( P t ) t ≥ 0 (P_t)_{t\geq0} ( P t ) t ≥ 0 , P t : = e t L P_t:=e^{tL} P t := e t L , is the associated Markov semigroup .
Remark 3.2. By the Markov property, ( P t ) t ≥ 0 (P_t)_{t\geq0} ( P t ) t ≥ 0 is positivity preserving and sub-Markovian, that is, 0 ≤ f ≤ c 0\leq f\leq c 0 ≤ f ≤ c m \mathfrak{m} m -a.e. implies 0 ≤ P t f ≤ c 0\leq P_tf\leq c 0 ≤ P t f ≤ c m \mathfrak{m} m -a.e.; it extends to a semigroup of contractions on L p ( X , m ) L^p(X,\mathfrak{m}) L p ( X , m ) for every p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] , strongly continuous for p < ∞ p<\infty p < ∞ and weakly* continuous for p = ∞ p=\infty p = ∞ , and it is an analytic semigroup on L 2 ( X , m ) L^2(X,\mathfrak{m}) L 2 ( X , m ) [5 Section 2.1] .
Definition 3.3 (Strong locality and carré du champ). Let E \mathcal{E} E be a symmetric Dirichlet form on L 2 ( X , m ) L^2(X,\mathfrak{m}) L 2 ( X , m ) .
(i) E \mathcal{E} E is strongly local if E ( f , g ) = 0 \mathcal{E}(f,g)=0 E ( f , g ) = 0 for all f , g ∈ V f,g\in\mathbb{V} f , g ∈ V such that ( f + a ) g = 0 (f+a)g=0 ( f + a ) g = 0 m \mathfrak{m} m -a.e. for some a ∈ R a\in\mathbb{R} a ∈ R .
(ii) E \mathcal{E} E admits a carré du champ if there is a symmetric, bilinear and continuous map Γ : V × V → L 1 ( X , m ) \Gamma:\mathbb{V}\times\mathbb{V}\to L^1(X,\mathfrak{m}) Γ : V × V → L 1 ( X , m ) such that
2 ∫ X Γ ( f , g ) φ d m = E ( f , g φ ) + E ( g , f φ ) − E ( f g , φ ) 2\int_X\Gamma(f,g)\varphi\,\mathrm{d}\mathfrak{m}=\mathcal{E}(f,g\varphi)+\mathcal{E}(g,f\varphi)-\mathcal{E}(fg,\varphi) 2 ∫ X Γ ( f , g ) φ d m = E ( f , g φ ) + E ( g , f φ ) − E ( f g , φ ) for all f , g , φ ∈ V ∞ f,g,\varphi\in\mathbb{V}_\infty f , g , φ ∈ V ∞ . We write Γ ( f ) : = Γ ( f , f ) \Gamma(f):=\Gamma(f,f) Γ ( f ) := Γ ( f , f ) .
Definition 3.4 (Capacity, nests, quasi-continuity). Let E \mathcal{E} E be a symmetric Dirichlet form on L 2 ( X , m ) L^2(X,\mathfrak{m}) L 2 ( X , m ) . The capacity of an open set A ⊆ X A\subseteq X A ⊆ X is
Cap ( A ) : = inf { ∥ u ∥ V 2 : u ∈ V , u ≥ 1 m -a.e. on A } ( inf ∅ : = + ∞ ) , \operatorname{Cap}(A):=\inf\{\|u\|_{\mathbb{V}}^2:u\in\mathbb{V},\ u\geq1\ \mathfrak{m}\text{-a.e. on }A\}\qquad(\inf\varnothing:=+\infty), Cap ( A ) := inf { ∥ u ∥ V 2 : u ∈ V , u ≥ 1 m -a.e. on A } ( inf ∅ := + ∞ ) , and the capacity of an arbitrary set B ⊆ X B\subseteq X B ⊆ X is Cap ( B ) : = inf { Cap ( A ) : A ⊇ B , A open } \operatorname{Cap}(B):=\inf\{\operatorname{Cap}(A):A\supseteq B,\ A\text{ open}\} Cap ( B ) := inf { Cap ( A ) : A ⊇ B , A open } . For a closed set F ⊆ X F\subseteq X F ⊆ X let V F : = { f ∈ V : f = 0 m -a.e. on X ∖ F } \mathbb{V}_F:=\{f\in\mathbb{V}:f=0\ \mathfrak{m}\text{-a.e. on }X\setminus F\} V F := { f ∈ V : f = 0 m -a.e. on X ∖ F } .
(i) An E \mathcal{E} E -nest is a nondecreasing sequence ( F k ) k ∈ N (F_k)_{k\in\mathbb{N}} ( F k ) k ∈ N of closed subsets of X X X such that ⋃ k V F k \bigcup_k\mathbb{V}_{F_k} ⋃ k V F k is dense in V \mathbb{V} V .
(ii) A set N ⊆ X N\subseteq X N ⊆ X is E \mathcal{E} E -polar if N ⊆ X ∖ ⋃ k F k N\subseteq X\setminus\bigcup_kF_k N ⊆ X ∖ ⋃ k F k for some E \mathcal{E} E -nest ( F k ) k ∈ N (F_k)_{k\in\mathbb{N}} ( F k ) k ∈ N ; a property holds quasi-everywhere (q.e.) if it holds outside an E \mathcal{E} E -polar set.
(iii) A function f : X → R f:X\to\mathbb{R} f : X → R is E \mathcal{E} E -quasi-continuous if there is an E \mathcal{E} E -nest ( F k ) k ∈ N (F_k)_{k\in\mathbb{N}} ( F k ) k ∈ N such that the restriction f ∣ F k f|_{F_k} f ∣ F k is continuous on F k F_k F k , endowed with the topology induced by X X X , for every k k k .
Lemma 3.5. Let E \mathcal{E} E be a symmetric Dirichlet form on L 2 ( X , m ) L^2(X,\mathfrak{m}) L 2 ( X , m ) . Then V \mathbb{V} V is separable, and Cap ( ⋃ k A k ) ≤ ∑ k Cap ( A k ) \operatorname{Cap}(\bigcup_k A_k)\leq\sum_k\operatorname{Cap}(A_k) Cap ( ⋃ k A k ) ≤ ∑ k Cap ( A k ) for every sequence ( A k ) k ∈ N (A_k)_{k\in\mathbb{N}} ( A k ) k ∈ N of subsets of X X X .
Proof. The countable subadditivity of the capacity holds for every symmetric Dirichlet form [23 Ch. III, Thm. 2.8] . The space V \mathbb{V} V , endowed with ∥ ⋅ ∥ V \|\cdot\|_{\mathbb{V}} ∥ ⋅ ∥ V , is a Hilbert space, since E \mathcal{E} E is lower semicontinuous, and the inclusion ι : V → L 2 ( m ) \iota:\mathbb{V}\to L^2(\mathfrak{m}) ι : V → L 2 ( m ) is linear, continuous and injective; hence its adjoint ι ∗ : L 2 ( m ) → V \iota^*:L^2(\mathfrak{m})\to\mathbb{V} ι ∗ : L 2 ( m ) → V has dense range, since ( ran ι ∗ ) ⊥ = ker ι = { 0 } (\operatorname{ran}\iota^*)^\perp=\ker\iota=\{0\} ( ran ι ∗ ) ⊥ = ker ι = { 0 } . As X X X is Polish and m \mathfrak{m} m is σ \sigma σ -finite, L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) is separable, hence so are ran ι ∗ \operatorname{ran}\iota^* ran ι ∗ and V \mathbb{V} V . □ \square □
Definition 3.6 (Quasi-regularity). A symmetric Dirichlet form E \mathcal{E} E on L 2 ( X , m ) L^2(X,\mathfrak{m}) L 2 ( X , m ) is quasi-regular if:
⟨ Q R .1 ⟩ \langle\mathrm{QR.1}\rangle ⟨ QR.1 ⟩ there exists an E \mathcal{E} E -nest consisting of compact sets;
⟨ Q R .2 ⟩ \langle\mathrm{QR.2}\rangle ⟨ QR.2 ⟩ there exists a dense subset of V \mathbb{V} V whose elements admit E \mathcal{E} E -quasi-continuous representatives;
⟨ Q R .3 ⟩ \langle\mathrm{QR.3}\rangle ⟨ QR.3 ⟩ there exist an E \mathcal{E} E -polar set N ⊆ X N \subseteq X N ⊆ X and a countable family of E \mathcal{E} E -quasi-continuous functions in V \mathbb{V} V separating the points of X ∖ N X \setminus N X ∖ N .
Lemma 3.7 ([25 Lemma 2.4] ). Let E \mathcal{E} E be a symmetric Dirichlet form on L 2 ( X , m ) L^2(X,\mathfrak{m}) L 2 ( X , m ) . Assume that there is a nondecreasing sequence ( X n ) n ≥ 1 (X_n)_{n\geq1} ( X n ) n ≥ 1 of open subsets of X X X such that:
(a) Cap ( X n ) < ∞ \operatorname{Cap}(X_n) < \infty Cap ( X n ) < ∞ and X n ‾ ⊆ X n + 1 \overline{X_n} \subseteq X_{n+1} X n ⊆ X n + 1 for every n n n , and ( X n ‾ ) n ≥ 1 (\overline{X_n})_{n\geq1} ( X n ) n ≥ 1 is an E \mathcal{E} E -nest;
(b) ⟨ Q R .1 ′ ⟩ \langle\mathrm{QR.1}'\rangle ⟨ QR.1 ′ ⟩ : for all n , m ≥ 1 n,m \geq1 n , m ≥ 1 there is a compact set K n , m ⊆ X K_{n,m} \subseteq X K n , m ⊆ X with Cap ( X n ∖ K n , m ) ≤ 1 / m \operatorname{Cap}(X_n \setminus K_{n,m}) \leq1/m Cap ( X n ∖ K n , m ) ≤ 1/ m ;
(c) ⟨ Q R .2 ′ ⟩ \langle\mathrm{QR.2}'\rangle ⟨ QR.2 ′ ⟩ : the space of elements of V \mathbb{V} V admitting a continuous representative is dense in V \mathbb{V} V and separates the points of X X X .
Then E \mathcal{E} E is quasi-regular.
Proposition 3.8. Let E \mathcal{E} E be a quasi-regular symmetric Dirichlet form on L 2 ( X , m ) L^2(X,\mathfrak{m}) L 2 ( X , m ) . Then:
(i) every u ∈ V u \in\mathbb{V} u ∈ V admits an E \mathcal{E} E -quasi-continuous representative u ~ \widetilde{u} u , unique up to E \mathcal{E} E -polar sets;
(ii) for every linear continuous ℓ : V → R \ell:\mathbb{V} \to\mathbb{R} ℓ : V → R with ℓ ( u ) ≥ 0 \ell(u) \geq0 ℓ ( u ) ≥ 0 for all u ∈ V + u \in\mathbb{V}_+ u ∈ V + , there is a unique σ \sigma σ -finite positive Borel measure μ \mu μ on X X X , not charging E \mathcal{E} E -polar sets, such that u ~ ∈ L 1 ( μ ) \widetilde{u} \in L^1(\mu) u ∈ L 1 ( μ ) and ℓ ( u ) = ∫ X u ~ d μ \ell(u) = \int_X \widetilde{u}\,d\mu ℓ ( u ) = ∫ X u d μ for every u ∈ V u \in\mathbb{V} u ∈ V .
Proof. (i) is [10 Rem. 1.3.9 (ii)] , which holds for every quasi-regular Dirichlet form on a σ \sigma σ -finite measure space.
(ii). The bilinear form E 1 ( u , v ) : = E ( u , v ) + ∫ X u v d m \mathcal{E}_1(u,v) := \mathcal{E}(u,v) + \int_X uv\,d\mathfrak{m} E 1 ( u , v ) := E ( u , v ) + ∫ X uv d m is the scalar product of V \mathbb{V} V , so by the Riesz representation theorem there is u ℓ ∈ V u_\ell\in\mathbb{V} u ℓ ∈ V with E 1 ( u ℓ , v ) = ℓ ( v ) \mathcal{E}_1(u_\ell,v) = \ell(v) E 1 ( u ℓ , v ) = ℓ ( v ) for every v ∈ V v \in\mathbb{V} v ∈ V ; in particular, E 1 ( u ℓ , v ) ≥ 0 \mathcal{E}_1(u_\ell,v) \geq0 E 1 ( u ℓ , v ) ≥ 0 for every v ∈ V + v \in\mathbb{V}_+ v ∈ V + . We claim that u ℓ u_\ell u ℓ is 1 1 1 -excessive, that is, u ℓ ≥ 0 u_\ell\geq0 u ℓ ≥ 0 and e − t P t u ℓ ≤ u ℓ e^{-t}P_tu_\ell\leq u_\ell e − t P t u ℓ ≤ u ℓ m \mathfrak{m} m -a.e. for every t > 0 t > 0 t > 0 .
For u ∈ V u \in\mathbb{V} u ∈ V , the Markov property gives u ± ∈ V u^\pm\in\mathbb{V} u ± ∈ V and E ( ∣ u ∣ ) ≤ E ( u ) \mathcal{E}(|u|) \leq\mathcal{E}(u) E ( ∣ u ∣ ) ≤ E ( u ) ; expanding ∣ u ∣ = u + + u − |u| = u^+ + u^- ∣ u ∣ = u + + u − and u = u + − u − u = u^+ - u^- u = u + − u − , this reads E ( u + , u − ) ≤ 0 \mathcal{E}(u^+,u^-) \leq0 E ( u + , u − ) ≤ 0 . Since u ℓ + u ℓ − = 0 u_\ell^+u_\ell^- = 0 u ℓ + u ℓ − = 0 , we get
0 ≤ E 1 ( u ℓ , u ℓ − ) = E ( u ℓ + , u ℓ − ) − E 1 ( u ℓ − ) ≤ − E 1 ( u ℓ − ) , 0 \leq\mathcal{E}_1(u_\ell,u_\ell^-) = \mathcal{E}(u_\ell^+,u_\ell^-) - \mathcal{E}_1(u_\ell^-) \leq-\mathcal{E}_1(u_\ell^-), 0 ≤ E 1 ( u ℓ , u ℓ − ) = E ( u ℓ + , u ℓ − ) − E 1 ( u ℓ − ) ≤ − E 1 ( u ℓ − ) ,
so u ℓ − = 0 u_\ell^- = 0 u ℓ − = 0 . Fix now v ∈ L 2 ( m ) v \in L^2(\mathfrak{m}) v ∈ L 2 ( m ) with v ≥ 0 v \geq0 v ≥ 0 , and set ϕ ( s ) : = e − s ∫ X u ℓ P s v d m \phi(s) := e^{-s}\int_X u_\ell P_sv\,d\mathfrak{m} ϕ ( s ) := e − s ∫ X u ℓ P s v d m . By Remark 3.2, ϕ \phi ϕ is continuous on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) and of class C 1 C^1 C 1 on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , and for s > 0 s > 0 s > 0 we have P s v ∈ D ( L ) P_sv \in D(L) P s v ∈ D ( L ) and P s v ≥ 0 P_sv \geq0 P s v ≥ 0 , so that
ϕ ′ ( s ) = e − s ∫ X u ℓ ( L P s v − P s v ) d m = − e − s E 1 ( u ℓ , P s v ) ≤ 0. \phi'(s) = e^{-s}\int_X u_\ell(LP_sv-P_sv)\,d\mathfrak{m} = -e^{-s}\mathcal{E}_1(u_\ell,P_sv) \leq0. ϕ ′ ( s ) = e − s ∫ X u ℓ ( L P s v − P s v ) d m = − e − s E 1 ( u ℓ , P s v ) ≤ 0.
Hence, by the symmetry of P t P_t P t , ∫ X e − t P t u ℓ v d m = ϕ ( t ) ≤ ϕ ( 0 ) = ∫ X u ℓ v d m \int_X e^{-t}P_tu_\ell\,v\,d\mathfrak{m} = \phi(t) \leq\phi(0) = \int_X u_\ell v\,d\mathfrak{m} ∫ X e − t P t u ℓ v d m = ϕ ( t ) ≤ ϕ ( 0 ) = ∫ X u ℓ v d m , and the claim follows since v v v is arbitrary.
Since E \mathcal{E} E is quasi-regular and u ℓ ∈ V u_\ell\in\mathbb{V} u ℓ ∈ V is 1 1 1 -excessive, the transfer method of [23 Ch. VI, Prop. 2.1] gives a unique σ \sigma σ -finite positive Borel measure μ \mu μ on X X X , not charging E \mathcal{E} E -polar sets, such that E 1 ( u ℓ , v ) = ∫ X v ~ d μ \mathcal{E}_1(u_\ell,v) = \int_X \widetilde{v}\,d\mu E 1 ( u ℓ , v ) = ∫ X v d μ for every v ∈ V + v \in\mathbb{V}_+ v ∈ V + . For v ∈ V v \in\mathbb{V} v ∈ V , the functions ( v ~ ) ± (\widetilde{v})^\pm ( v ) ± are quasi-continuous representatives of v ± ∈ V + v^\pm\in\mathbb{V}_+ v ± ∈ V + , hence coincide q.e., and so μ \mu μ -a.e., with v ± v^\pm v ± by the uniqueness in (i). Therefore ∫ X ( v ~ ) ± d μ = ℓ ( v ± ) < ∞ \int_X(\widetilde{v})^\pm d\mu= \ell(v^\pm) < \infty ∫ X ( v ) ± d μ = ℓ ( v ± ) < ∞ , so v ~ ∈ L 1 ( μ ) \widetilde{v} \in L^1(\mu) v ∈ L 1 ( μ ) and ∫ X v ~ d μ = ℓ ( v + ) − ℓ ( v − ) = ℓ ( v ) \int_X\widetilde{v}\,d\mu= \ell(v^+) - \ell(v^-) = \ell(v) ∫ X v d μ = ℓ ( v + ) − ℓ ( v − ) = ℓ ( v ) .
This is the argument of [25 (2.7) and Prop. 2.5] , where the statement is given under the standing assumptions of [25 Section 2] : strong locality, existence of a carré du champ and mass preservation of the semigroup. The proof above uses none of them. □ \square □
Remark 3.9. Let E \mathcal{E} E be quasi-regular, let u ∈ V u \in\mathbb{V} u ∈ V , and let ( F k ) k (F_k)_k ( F k ) k be an E \mathcal{E} E -nest such that u ~ ∣ F k \widetilde{u}|_{F_k} u ∣ F k is continuous for every k k k . Continuous functions are quasi-continuous with respect to the trivial nest F k = X F_k = X F k = X . By the uniqueness in Proposition 3.8 (i):
(i) if u u u is continuous, then u ~ = u \widetilde{u} = u u = u q.e.; hence ℓ ( u ) = ∫ X u d μ \ell(u) = \int_X u\,d\mu ℓ ( u ) = ∫ X u d μ for ℓ , μ \ell,\mu ℓ , μ as in Proposition 3.8 (ii);
(ii) if ζ ∈ C ( X ) ∩ L ∞ ( m ) \zeta\in C(X) \cap L^\infty(\mathfrak{m}) ζ ∈ C ( X ) ∩ L ∞ ( m ) and ζ u ∈ V \zeta u \in\mathbb{V} ζ u ∈ V , then ( ζ u ) ~ = ζ ~ u ~ (\zeta u)^{\widetilde{\ }} = \widetilde{\zeta}\widetilde{u} ( ζ u ) = ζ u q.e.;
(iii) if η : R → R \eta:\mathbb{R} \to\mathbb{R} η : R → R is Lipschitz with η ( 0 ) = 0 \eta(0)=0 η ( 0 ) = 0 and η ∘ u ∈ V \eta\circ u \in\mathbb{V} η ∘ u ∈ V , then ( η ∘ u ) ~ = η ∘ u ~ (\eta\circ u)^{\widetilde{\ }} = \eta\circ\widetilde{u} ( η ∘ u ) = η ∘ u q.e.
Indeed, in (ii) and (iii) the right-hand sides are representatives of ζ u \zeta u ζ u and η ∘ u \eta\circ u η ∘ u whose restrictions to each F k F_k F k are continuous.
Proposition 3.10 (The Cheeger energy as a Dirichlet form). Assume (S), and let E ( f ) : = ∫ X ∣ D f ∣ 2 d m \mathcal{E}(f) := \int_X |Df|^2\mathrm{d}m E ( f ) := ∫ X ∣ D f ∣ 2 d m for f ∈ W 1 , 2 ( X ) ∗ a n d ∗ f \in W^{1,2}(X) *and* f ∈ W 1 , 2 ( X ) ∗ an d ∗ \mathcal{E}(f) := +\infty$ for f ∈ L 2 ( m ) ∖ W 1 , 2 ( X ) f \in L^2(m) \setminus W^{1,2}(X) f ∈ L 2 ( m ) ∖ W 1 , 2 ( X ) . Then E = 2 C h \mathcal{E} = 2\mathrm{Ch} E = 2 Ch is a strongly local symmetric Dirichlet form on L 2 ( X , m ) L^2(X,m) L 2 ( X , m ) with domain V = W 1 , 2 ( X ) \mathcal{V} = W^{1,2}(X) V = W 1 , 2 ( X ) , it admits the carré du champ Γ ( f , g ) = ∇ f ⋅ ∇ g \Gamma(f,g) = \nabla f \cdot\nabla g Γ ( f , g ) = ∇ f ⋅ ∇ g , its generator is Δ \Delta Δ , and its Markov semigroup is ( h t ) t ≥ 0 (h_t)_{t\geq0} ( h t ) t ≥ 0 .
Proof. By (S) and Lemma 2.1 , ( X , d ) (X,\mathrm{d}) ( X , d ) is Polish and m \mathbf{m} m is σ \sigma σ -finite with full support.
Dirichlet form. E \mathcal{E} E is a quadratic form by (IH). It is lower semicontinuous on L 2 ( m ) L^2(m) L 2 ( m ) : if f n → f f_n \to f f n → f in L 2 ( m ) L^2(m) L 2 ( m ) and lim inf n E ( f n ) < ∞ \liminf_n \mathcal{E}(f_n) < \infty lim inf n E ( f n ) < ∞ , we extract a subsequence realising the liminf, converging m \mathbf{m} m -a.e., and along which ∣ D f n ∣ ⇀ G |Df_n| \rightharpoonup G ∣ D f n ∣ ⇀ G weakly in L 2 ( m ) L^2(m) L 2 ( m ) ; by Proposition 2.4 (a), G G G is a weak upper gradient of f f f , hence f ∈ W 1 , 2 ( X ) f \in W^{1,2}(X) f ∈ W 1 , 2 ( X ) and E ( f ) ≤ ∫ X G 2 d m ≤ lim inf n E ( f n ) \mathcal{E}(f) \leq\int_X G^2\mathrm{d}m \leq\liminf_n \mathcal{E}(f_n) E ( f ) ≤ ∫ X G 2 d m ≤ lim inf n E ( f n ) . For the Markov property, there is nothing to prove if E ( f ) = ∞ \mathcal{E}(f) = \infty E ( f ) = ∞ ; if f ∈ W 1 , 2 ( X ) f \in W^{1,2}(X) f ∈ W 1 , 2 ( X ) and η \eta η is 1-Lipschitz with η ( 0 ) = 0 \eta(0) = 0 η ( 0 ) = 0 , then ∣ η ∘ f ∣ ≤ ∣ f ∣ |\eta\circ f| \leq|f| ∣ η ∘ f ∣ ≤ ∣ f ∣ and ∣ D ( η ∘ f ) ∣ ≤ ∣ D f ∣ |D(\eta\circ f)| \leq|Df| ∣ D ( η ∘ f ) ∣ ≤ ∣ D f ∣ m \mathbf{m} m -a.e. by the chain rule. The domain W 1 , 2 ( X ) W^{1,2}(X) W 1 , 2 ( X ) is dense in L 2 ( m ) L^2(m) L 2 ( m ) by Lemma 2.3 .
Carré du champ and strong locality. By (IH) and Proposition 2.4 (a), ( f , g ) ↦ ∇ f ⋅ ∇ g (f,g) \mapsto\nabla f \cdot\nabla g ( f , g ) ↦ ∇ f ⋅ ∇ g is continuous, symmetric and bilinear, and satisfies the Leibniz rule; expanding with the Leibniz rule gives 2 ∫ X Γ ( f , g ) φ d m = E ( f , g φ ) + E ( g , f φ ) − E ( f g , φ ) 2\int_X \Gamma(f,g)\varphi\mathrm{d}m = \mathcal{E}(f,g\varphi) + \mathcal{E}(g,f\varphi) - \mathcal{E}(fg,\varphi) 2 ∫ X Γ ( f , g ) φ d m = E ( f , g φ ) + E ( g , f φ ) − E ( f g , φ ) for f , g , φ ∈ V ∞ f,g,\varphi\in\mathcal{V}_\infty f , g , φ ∈ V ∞ . Let f , g ∈ W 1 , 2 ( X ) f,g \in W^{1,2}(X) f , g ∈ W 1 , 2 ( X ) with ( f + a ) g = 0 (f+a)g = 0 ( f + a ) g = 0 m \mathbf{m} m -a.e. for some a ∈ R a \in\mathbb{R} a ∈ R . Then f = − a f = -a f = − a m \mathbf{m} m -a.e. on { g ≠ 0 } \{g \neq0\} { g = 0 } , so ∣ D f ∣ = 0 |Df| = 0 ∣ D f ∣ = 0 m \mathbf{m} m -a.e. on { g ≠ 0 } \{g \neq0\} { g = 0 } by locality, while ∣ D g ∣ = 0 |Dg| = 0 ∣ D g ∣ = 0 m \mathbf{m} m -a.e. on { g = 0 } \{g = 0\} { g = 0 } . Hence ∣ ∇ f ⋅ ∇ g ∣ ≤ ∣ D f ∣ ∣ D g ∣ = 0 |\nabla f \cdot\nabla g| \leq|Df||Dg| = 0 ∣∇ f ⋅ ∇ g ∣ ≤ ∣ D f ∣∣ D g ∣ = 0 m \mathbf{m} m -a.e. and E ( f , g ) = 0 \mathcal{E}(f,g) = 0 E ( f , g ) = 0 .
Generator and semigroup. By Proposition 2.4 (b), ∫ X g Δ f d m = − E ( f , g ) \int_X g\Delta f\mathrm{d}m = -\mathcal{E}(f,g) ∫ X g Δ f d m = − E ( f , g ) for f ∈ D ( Δ ) f \in D(\Delta) f ∈ D ( Δ ) and g ∈ W 1 , 2 ( X ) g \in W^{1,2}(X) g ∈ W 1 , 2 ( X ) , so Δ \Delta Δ is the generator of E \mathcal{E} E . Fix f ∈ L 2 ( m ) f \in L^2(m) f ∈ L 2 ( m ) and set u ( t ) : = h t f u(t) := h_tf u ( t ) := h t f , v ( t ) : = P t f v(t) := P_tf v ( t ) := P t f . Both curves belong to C ( [ 0 , ∞ ) ; L 2 ( m ) ) ∩ C 1 ( ( 0 , ∞ ) ; L 2 ( m ) ) C([0,\infty);L^2(m))\cap C^1((0,\infty);L^2(m)) C ([ 0 , ∞ ) ; L 2 ( m )) ∩ C 1 (( 0 , ∞ ) ; L 2 ( m )) , take values in D ( Δ ) D(\Delta) D ( Δ ) for t > 0 t>0 t > 0 , start from f f f , and satisfy u ′ = Δ u u' = \Delta u u ′ = Δ u , v ′ = Δ v v' = \Delta v v ′ = Δ v on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , by Proposition 2.4 (c) and Remark 3.2. Hence z : = u − v z := u-v z := u − v satisfies
d d t ∥ z ( t ) ∥ L 2 ( m ) 2 = 2 ∫ X z ( t ) Δ z ( t ) d m = − 2 ∫ X ∣ D z ( t ) ∣ 2 d m ≤ 0 for t > 0 , \frac{\mathrm{d}}{\mathrm{d}t}\lVert z(t)\rVert_{L^2(m)}^2 = 2\int_X z(t)\Delta z(t)\mathrm{d}m = -2\int_X |Dz(t)|^2\mathrm{d}m \leq0 \qquad\text{for }t>0, d t d ∥ z ( t ) ∥ L 2 ( m ) 2 = 2 ∫ X z ( t ) Δ z ( t ) d m = − 2 ∫ X ∣ D z ( t ) ∣ 2 d m ≤ 0 for t > 0 ,
and since z z z is continuous on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) with z ( 0 ) = 0 z(0)=0 z ( 0 ) = 0 , we get z ≡ 0 z\equiv0 z ≡ 0 . □
Proof of Theorem 1.4
Throughout this section we assume (S) and use the notation of § 3 with the identifications of Proposition 3.10 : V = W 1 , 2 ( X ) \mathcal{V} = W^{1,2}(X) V = W 1 , 2 ( X ) , E ( u , v ) = ∫ X Γ ( u , v ) d m \mathcal{E}(u,v) = \int_X \Gamma(u,v)\mathrm{d}m E ( u , v ) = ∫ X Γ ( u , v ) d m with Γ ( u , v ) = ∇ u ⋅ ∇ v \Gamma(u,v) = \nabla u \cdot\nabla v Γ ( u , v ) = ∇ u ⋅ ∇ v , Γ ( u ) = ∣ D u ∣ 2 \Gamma(u) = |Du|^2 Γ ( u ) = ∣ D u ∣ 2 , L = Δ L = \Delta L = Δ and P t = h t P_t = h_t P t = h t . The proof is organised as follows:
L 2 L^2 L 2 -BE ⟹ \Longrightarrow ⟹ Quasi-regularity (Lemma 4.1 )
⟹ \Longrightarrow ⟹ Improved Bochner inequality (Proposition 4.11 ) ⟹ \Longrightarrow ⟹ L 1 L^1 L 1 -BE (Proposition 4.13)
⟹ \Longrightarrow ⟹ Power-Harnack (Lemma 4.14) ⟹ \Longrightarrow ⟹ Theorem 1.4.
Quasi-regularity
We now show that, under (S), the Cheeger energy is a quasi-regular Dirichlet form. Quasi-regularity is what makes Proposition 3.8 available: in Subsection 4.3 it is used to represent the nonnegative functional Γ 2 K [ u ] \Gamma^K_2[u] Γ 2 K [ u ] by a measure, and to evaluate it on Lipschitz functions. We verify Savaré’s criterion (Lemma 3.7 ) on the open sets X n X_n X n where finitely many elements of a dense sequence of test functions are not too small. Since ( X , d ) (X,\mathrm{d}) ( X , d ) need not be locally compact, the main point is condition (b): compact sets of almost full capacity in X n X_n X n are obtained by removing from X n X_n X n an open set of small capacity, outside of which X n X_n X n is totally bounded.
Lemma 4.1 (Quasi-regularity). Assume (S). Then E = 2 C h \mathcal{E} = 2\mathrm{Ch} E = 2 Ch is a quasi-regular Dirichlet form on L 2 ( X , m ) L^2(X,\mathbf{m}) L 2 ( X , m ) .
Proof. We verify the conditions of Lemma 3.7 .
Step 1: condition (c). By Lemma 2.5(iv), Test ∞ ( X ) ⊆ V ∩ LIP ( X ) \operatorname{Test}^{\infty}(X) \subseteq\mathcal{V}\cap\operatorname{LIP}(X) Test ∞ ( X ) ⊆ V ∩ LIP ( X ) is dense in V \mathcal{V} V . Given x ≠ y x\neq y x = y , let B r ( x ) B_r(x) B r ( x ) be admissible with 2 r < d ( x , y ) 2r<\mathrm{d}(x,y) 2 r < d ( x , y ) ; the function β x , r \beta_{x,r} β x , r of Lemma 2.3 (ii) belongs to V ∩ C ( X ) \mathcal{V}\cap C(X) V ∩ C ( X ) and β x , r ( x ) = 1 ≠ 0 = β x , r ( y ) \beta_{x,r}(x)=1\neq0=\beta_{x,r}(y) β x , r ( x ) = 1 = 0 = β x , r ( y ) . Step 2: condition (a). By Lemma 3.5 , V \mathrm{V} V is separable, so there are test functions f j ∈ Test ∞ ( X ) f_j \in\operatorname{Test}^{\infty}(X) f j ∈ Test ∞ ( X ) , j ≥ 1 j \ge1 j ≥ 1 , forming a dense subset of V \mathrm{V} V . Set
X n : = ⋃ j = 1 n { ∣ f j ∣ > 1 n + 1 } , n ≥ 1. X_n := \bigcup_{j=1}^{n}\left\{ |f_j| > \frac{1}{n+1} \right\}, \qquad n \ge1. X n := j = 1 ⋃ n { ∣ f j ∣ > n + 1 1 } , n ≥ 1.
Each X n X_n X n is open, the sequence is nondecreasing, and by continuity of the f j f_j f j
X n ‾ ⊆ ⋃ j = 1 n { ∣ f j ∣ ≥ 1 n + 1 } ⊆ ⋃ j = 1 n + 1 { ∣ f j ∣ > 1 n + 2 } = X n + 1 . \overline{X_n} \subseteq\bigcup_{j=1}^{n}\left\{ |f_j| \ge\frac{1}{n+1} \right\} \subseteq\bigcup_{j=1}^{n+1}\left\{ |f_j| > \frac{1}{n+2} \right\} = X_{n+1}. X n ⊆ j = 1 ⋃ n { ∣ f j ∣ ≥ n + 1 1 } ⊆ j = 1 ⋃ n + 1 { ∣ f j ∣ > n + 2 1 } = X n + 1 .
Since ( n + 1 ) ∣ f j ∣ ∈ V (n+1)|f_j| \in\mathrm{V} ( n + 1 ) ∣ f j ∣ ∈ V is at least 1 on { ∣ f j ∣ > ( n + 1 ) − 1 } \{|f_j| > (n+1)^{-1}\} { ∣ f j ∣ > ( n + 1 ) − 1 } , Lemma 3.5 gives
Cap ( X n ) ≤ ( n + 1 ) 2 ∑ j ≤ n ∥ f j ∥ V 2 < ∞ . \operatorname{Cap}(X_n) \le(n+1)^2 \sum_{j\le n} \lVert f_j\rVert_{\mathrm{V}}^2 < \infty. Cap ( X n ) ≤ ( n + 1 ) 2 j ≤ n ∑ ∥ f j ∥ V 2 < ∞.
To prove that ( X n ‾ ) n (\overline{X_n})_n ( X n ) n is a nest, fix j j j and, for n ≥ j n \ge j n ≥ j , let η n ( s ) : = sgn ( s ) ( ∣ s ∣ − 1 n + 1 ) + \eta_n(s) := \operatorname{sgn}(s)\left(|s|-\frac{1}{n+1}\right)^+ η n ( s ) := sgn ( s ) ( ∣ s ∣ − n + 1 1 ) + . Then η n ∘ f j ∈ V \eta_n\circ f_j \in\mathrm{V} η n ∘ f j ∈ V vanishes outside X n X_n X n , so η n ∘ f j ∈ V X n \eta_n\circ f_j \in\mathrm{V}_{X_n} η n ∘ f j ∈ V X n ; moreover ∣ f j − η n ∘ f j ∣ ≤ ∣ f j ∣ ∧ 1 n + 1 |f_j-\eta_n\circ f_j| \le|f_j|\wedge\frac{1}{n+1} ∣ f j − η n ∘ f j ∣ ≤ ∣ f j ∣ ∧ n + 1 1 and, by the chain rule and locality, ∣ D ( f j − η n ∘ f j ) ∣ = ∣ D f j ∣ χ { 0 < ∣ f j ∣ ≤ ( n + 1 ) − 1 } |D(f_j-\eta_n\circ f_j)|=|Df_j|\chi_{\{0<|f_j|\le(n+1)^{-1}\}} ∣ D ( f j − η n ∘ f j ) ∣ = ∣ D f j ∣ χ { 0 < ∣ f j ∣ ≤ ( n + 1 ) − 1 } m \mathfrak{m} m -a.e. By dominated convergence η n ∘ f j → f j \eta_n\circ f_j\to f_j η n ∘ f j → f j in V \mathrm{V} V , so ⋃ n V X n ‾ \bigcup_n \mathrm{V}_{\overline{X_n}} ⋃ n V X n is dense in V \mathrm{V} V .
Step 3: condition (b). Fix n ≥ 1 n\ge1 n ≥ 1 and set
Ψ n : = 1 ∧ ( 2 ( n + 1 ) max 1 ≤ j ≤ n ( ∣ f j ∣ − 1 2 ( n + 1 ) ) ) + , Ω n : = ⋃ j = 1 n { ∣ f j ∣ ≥ 1 2 ( n + 1 ) } . \Psi_n := 1\wedge\left(2(n+1)\max_{1\le j\le n}\left(|f_j|-\frac{1}{2(n+1)}\right)\right)^+,\qquad\Omega_n := \bigcup_{j=1}^{n}\left\{|f_j|\ge\frac{1}{2(n+1)}\right\}. Ψ n := 1 ∧ ( 2 ( n + 1 ) 1 ≤ j ≤ n max ( ∣ f j ∣ − 2 ( n + 1 ) 1 ) ) + , Ω n := j = 1 ⋃ n { ∣ f j ∣ ≥ 2 ( n + 1 ) 1 } .
Ψ n \Psi_n Ψ n is Lipschitz, 0 ≤ Ψ n ≤ 1 0\le\Psi_n\le1 0 ≤ Ψ n ≤ 1 , Ψ n = 1 \Psi_n=1 Ψ n = 1 on X n ‾ \overline{X_n} X n , Ψ n = 0 \Psi_n=0 Ψ n = 0 outside Ω n \Omega_n Ω n , and, by Chebyshev’s inequality,
m ( Ω n ) ≤ 4 ( n + 1 ) 2 ∑ j ≤ n ∥ f j ∥ L 2 ( m ) 2 < ∞ . \mathfrak{m}(\Omega_n)\le4(n+1)^2\sum_{j\le n}\lVert f_j\rVert_{L^2(\mathfrak{m})}^2<\infty. m ( Ω n ) ≤ 4 ( n + 1 ) 2 j ≤ n ∑ ∥ f j ∥ L 2 ( m ) 2 < ∞.
Let ( x i ) i ≥ 1 (x_i)_{i\ge1} ( x i ) i ≥ 1 be dense in X X X and, for k ≥ 1 k\ge1 k ≥ 1 ,
w k : = Ψ n ∧ min 1 ≤ i ≤ k d ( ⋅ , x i ) . w_k:=\Psi_n\wedge\min_{1\le i\le k}d(\cdot,x_i). w k := Ψ n ∧ 1 ≤ i ≤ k min d ( ⋅ , x i ) .
Each w k w_k w k is Lipschitz with Lip ( w k ) ≤ L n : = max { Lip ( Ψ n ) , 1 } \operatorname{Lip}(w_k)\le L_n:=\max\{\operatorname{Lip}(\Psi_n),1\} Lip ( w k ) ≤ L n := max { Lip ( Ψ n ) , 1 } , vanishes outside the closed set Ω n \Omega_n Ω n and satisfies 0 ≤ w k ≤ Ψ n ≤ 1 0\le w_k\le\Psi_n\le1 0 ≤ w k ≤ Ψ n ≤ 1 , so w k ∈ LIP f m ( X ) w_k\in\operatorname{LIP}_{\mathrm{fm}}(X) w k ∈ LIP fm ( X ) and, by Lemma 2.3 (i), ∥ w k ∥ V 2 ≤ ( 1 + L n 2 ) m ( Ω n ) \lVert w_k\rVert_{\mathrm{V}}^2\le(1+L_n^2)\mathfrak{m}(\Omega_n) ∥ w k ∥ V 2 ≤ ( 1 + L n 2 ) m ( Ω n ) . The sequence ( w k ) k (w_k)_k ( w k ) k is nonincreasing and w k → 0 w_k\to0 w k → 0 pointwise, since ( x i ) (x_i) ( x i ) is dense; hence w k → 0 w_k\to0 w k → 0 in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) and w k ⇀ 0 w_k\rightharpoonup0 w k ⇀ 0 weakly in V \mathrm{V} V . By the Banach–Saks theorem there is an increasing sequence ( k i ) i (k_i)_i ( k i ) i such that v h : = h − 1 ∑ i = 1 h w k i → 0 v_h:=h^{-1}\sum_{i=1}^{h}w_{k_i}\to0 v h := h − 1 ∑ i = 1 h w k i → 0 strongly in V \mathrm{V} V ; we extract ( v h l ) l ≥ 1 (v_{h_l})_{l\ge1} ( v h l ) l ≥ 1 with ∥ v h l ∥ V ≤ 2 − l \lVert v_{h_l}\rVert_{\mathrm{V}}\le2^{-l} ∥ v h l ∥ V ≤ 2 − l . The functions v h l v_{h_l} v h l are continuous, so G m : = ⋃ l ≥ m { v h l > 2 − l } G_m:=\bigcup_{l\ge m}\{v_{h_l}>2^{-l}\} G m := ⋃ l ≥ m { v h l > 2 − l } is open; the series W m : = ∑ l ≥ m 2 l v h l W_m:=\sum_{l\ge m}2^lv_{h_l} W m := ∑ l ≥ m 2 l v h l converges in V \mathrm{V} V with ∥ W m ∥ V ≤ 2 − m + 1 \lVert W_m\rVert_{\mathrm{V}}\le2^{-m+1} ∥ W m ∥ V ≤ 2 − m + 1 , and W m ≥ 1 W_m\ge1 W m ≥ 1 m \mathfrak{m} m -a.e. on G m G_m G m . Hence the closed set F m : = X ∖ G m F_m:=X\setminus G_m F m := X ∖ G m satisfies Cap ( X ∖ F m ) ≤ 2 − 2 m + 2 \operatorname{Cap}(X\setminus F_m)\le2^{-2m+2} Cap ( X ∖ F m ) ≤ 2 − 2 m + 2 . On F m F_m F m we have w k h l ≤ v h l ≤ 2 − l w_{k_{h_l}}\le v_{h_l}\le2^{-l} w k h l ≤ v h l ≤ 2 − l for l ≥ m l\ge m l ≥ m , because ( w k ) k (w_k)_k ( w k ) k is nonincreasing. Given δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , choose l ≥ m l\ge m l ≥ m with 2 − l < δ 2^{-l}<\delta 2 − l < δ and set p : = k h l p:=k_{h_l} p := k h l : since Ψ n = 1 \Psi_n=1 Ψ n = 1 on X n ‾ \overline{X_n} X n , every x ∈ X n ‾ ∩ F m x\in\overline{X_n}\cap F_m x ∈ X n ∩ F m satisfies min i ≤ p d ( x , x i ) = w p ( x ) < δ \min_{i\le p}d(x,x_i)=w_p(x)<\delta min i ≤ p d ( x , x i ) = w p ( x ) < δ , that is, X n ‾ ∩ F m ⊆ ⋃ i ≤ p B δ ( x i ) \overline{X_n}\cap F_m\subseteq\bigcup_{i\le p}B_\delta(x_i) X n ∩ F m ⊆ ⋃ i ≤ p B δ ( x i ) . Thus K ~ n , m : = X n ‾ ∩ F m \widetilde K_{n,m}:=\overline{X_n}\cap F_m K n , m := X n ∩ F m is closed and totally bounded, hence compact by completeness, and Cap ( X n ∖ K ~ n , m ) ≤ Cap ( X ∖ F m ) ≤ 2 − 2 m + 2 \operatorname{Cap}(X_n\setminus\widetilde K_{n,m})\le\operatorname{Cap}(X\setminus F_m)\le2^{-2m+2} Cap ( X n ∖ K n , m ) ≤ Cap ( X ∖ F m ) ≤ 2 − 2 m + 2 . The sets K n , m : = K ~ n , m + 2 K_{n,m}:=\widetilde K_{n,m+2} K n , m := K n , m + 2 satisfy Cap ( X n ∖ K n , m ) ≤ 2 − 2 m − 2 < 1 / m \operatorname{Cap}(X_n\setminus K_{n,m})\le2^{-2m-2}<1/m Cap ( X n ∖ K n , m ) ≤ 2 − 2 m − 2 < 1/ m . □
Test functions and the heat flow
In this subsection we collect the properties of test functions used in Subsections 4.3–4.4. Lemma 4.2 shows that every u ∈ V ∞ u\in\mathrm{V}_\infty u ∈ V ∞ is approximated in V \mathrm{V} V by test functions, with uniform L ∞ L^\infty L ∞ bounds and preserving the sign; this allows us to define Γ 2 \Gamma_2 Γ 2 on test functions only and to test the resulting inequalities against general functions in V ∞ \mathrm{V}_\infty V ∞ . Lemma 4.3 shows that Test ∞ ( X ) \operatorname{Test}^{\infty}(X) Test ∞ ( X ) is invariant under the heat flow, with a Lipschitz bound coming from (4) and (SL). Lemma 4.4 gives the time regularity of r ↦ Γ ( h r f ) r\mapsto\Gamma(h_rf) r ↦ Γ ( h r f ) as an L 1 ( m ) L^1(\mathfrak{m}) L 1 ( m ) -valued map, needed to differentiate the function A A A of Subsection 4.4.
Lemma 4.2 (Approximation by test functions). Assume (S). Let ( φ k ) k ≥ 1 ⊆ C c ∞ ( 0 , ∞ ) (\varphi_k)_{k\geq1} \subseteq C^\infty_c(0,\infty) ( φ k ) k ≥ 1 ⊆ C c ∞ ( 0 , ∞ ) be nonnegative with ∫ 0 ∞ φ k ( s ) d s = 1 \int_0^\infty\varphi_k(s)\,\mathrm{d}s=1 ∫ 0 ∞ φ k ( s ) d s = 1 and supp ( φ k ) ⊆ ( 0 , 1 / k ) \operatorname{supp}(\varphi_k)\subseteq(0,1/k) supp ( φ k ) ⊆ ( 0 , 1/ k ) , and let u ∈ V ∞ u\in V_\infty u ∈ V ∞ . Then h φ k u ∈ Test ∞ ( X ) h_{\varphi_k}u\in\operatorname{Test}^\infty(X) h φ k u ∈ Test ∞ ( X ) , ∣ h φ k u ∣ ≤ ∥ u ∥ L ∞ ( m ) |h_{\varphi_k}u|\leq\|u\|_{L^\infty(\mathfrak{m})} ∣ h φ k u ∣ ≤ ∥ u ∥ L ∞ ( m ) m \mathfrak{m} m -a.e., h φ k u ≥ 0 h_{\varphi_k}u\geq0 h φ k u ≥ 0 m \mathfrak{m} m -a.e. whenever u ≥ 0 u\geq0 u ≥ 0 m \mathfrak{m} m -a.e., and h φ k u → u h_{\varphi_k}u\to u h φ k u → u in V V V as k → ∞ k\to\infty k → ∞ . Consequently, Γ ( h φ k u ) → Γ ( u ) \Gamma(h_{\varphi_k}u)\to\Gamma(u) Γ ( h φ k u ) → Γ ( u ) and Γ ( h φ k u , v ) → Γ ( u , v ) \Gamma(h_{\varphi_k}u,v)\to\Gamma(u,v) Γ ( h φ k u , v ) → Γ ( u , v ) in L 1 ( m ) L^1(\mathfrak{m}) L 1 ( m ) for every v ∈ V v\in V v ∈ V .
Proof. The fact that h φ k u h_{\varphi_k}u h φ k u belongs to Test ∞ ( X ) \operatorname{Test}^\infty(X) Test ∞ ( X ) is Lemma 2.5(iii), and the two pointwise bounds follow from the weak maximum principle applied to ± u \pm u ± u and ∥ u ∥ L ∞ ( m ) ∓ u \|u\|_{L^\infty(\mathfrak{m})}\mp u ∥ u ∥ L ∞ ( m ) ∓ u . For s ↓ 0 s\downarrow0 s ↓ 0 we have h s u → u h_su\to u h s u → u in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) and Ch ( h s u ) ≤ Ch ( u ) \operatorname{Ch}(h_su)\leq\operatorname{Ch}(u) Ch ( h s u ) ≤ Ch ( u ) , so the lower semicontinuity of Ch \operatorname{Ch} Ch gives Ch ( h s u ) → Ch ( u ) \operatorname{Ch}(h_su)\to\operatorname{Ch}(u) Ch ( h s u ) → Ch ( u ) ; since V V V is a Hilbert space, h s u → u h_su\to u h s u → u in V V V . As s ↦ h s u s\mapsto h_su s ↦ h s u is continuous from [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) to V V V ,
∥ h φ k u − u ∥ V = ∥ ∫ 0 ∞ ( h s u − u ) φ k ( s ) d s ∥ V ≤ sup s ∈ ( 0 , 1 / k ) ∥ h s u − u ∥ V ⟶ 0. \|h_{\varphi_k}u-u\|_V=\left\|\int_0^\infty(h_su-u)\varphi_k(s)\,\mathrm{d}s\right\|_V\leq\sup_{s\in(0,1/k)}\|h_su-u\|_V\longrightarrow0. ∥ h φ k u − u ∥ V = ∫ 0 ∞ ( h s u − u ) φ k ( s ) d s V ≤ s ∈ ( 0 , 1/ k ) sup ∥ h s u − u ∥ V ⟶ 0.
The L 1 L^1 L 1 convergences follow from the m \mathfrak{m} m -a.e. inequalities
∣ Γ ( w ) − Γ ( z ) ∣ ≤ Γ ( w − z ) 1 / 2 ( Γ ( w ) 1 / 2 + Γ ( z ) 1 / 2 ) , ∣ Γ ( w , v ) − Γ ( z , v ) ∣ ≤ Γ ( w − z ) 1 / 2 Γ ( v ) 1 / 2 , |\Gamma(w)-\Gamma(z)|\leq\Gamma(w-z)^{1/2}\left(\Gamma(w)^{1/2}+\Gamma(z)^{1/2}\right),\qquad|\Gamma(w,v)-\Gamma(z,v)|\leq\Gamma(w-z)^{1/2}\Gamma(v)^{1/2}, ∣Γ ( w ) − Γ ( z ) ∣ ≤ Γ ( w − z ) 1/2 ( Γ ( w ) 1/2 + Γ ( z ) 1/2 ) , ∣Γ ( w , v ) − Γ ( z , v ) ∣ ≤ Γ ( w − z ) 1/2 Γ ( v ) 1/2 ,
integrated with the Cauchy–Schwarz inequality.
:::
Lemma 4.3. Assume (S). Let f ∈ Test ∞ ( X ) f\in\operatorname{Test}^\infty(X) f ∈ Test ∞ ( X ) and r ≥ 0 r\geq0 r ≥ 0 . Then h r f ∈ Test ∞ ( X ) h_rf\in\operatorname{Test}^\infty(X) h r f ∈ Test ∞ ( X ) , Δ h r f = h r Δ f \Delta h_rf=h_r\Delta f Δ h r f = h r Δ f , and Γ ( h r f ) ≤ e − 2 K r Lip ( f ) 2 \Gamma(h_rf)\leq e^{-2Kr}\operatorname{Lip}(f)^2 Γ ( h r f ) ≤ e − 2 K r Lip ( f ) 2 m \mathfrak{m} m -a.e.
Proof. By (4) , the bound ∣ D f ∣ ≤ Lip ( f ) |Df|\leq\operatorname{Lip}(f) ∣ D f ∣ ≤ Lip ( f ) (which follows from Lemma 2.3 (i) by locality) and the weak maximum principle,
Γ ( h r f ) ≤ e − 2 K r h r ( Γ ( f ) ) ≤ e − 2 K r Lip ( f ) 2 . \Gamma(h_rf)\leq e^{-2Kr}h_r\left(\Gamma(f)\right)\leq e^{-2Kr}\operatorname{Lip}(f)^2. Γ ( h r f ) ≤ e − 2 K r h r ( Γ ( f ) ) ≤ e − 2 K r Lip ( f ) 2 .
so h r f h_rf h r f has a Lipschitz representative by (SL). The weak maximum principle gives ∥ h r f ∥ L ∞ ( m ) ≤ ∥ f ∥ L ∞ ( m ) \|h_rf\|_{L^\infty(\mathfrak{m})}\leq\|f\|_{L^\infty(\mathfrak{m})} ∥ h r f ∥ L ∞ ( m ) ≤ ∥ f ∥ L ∞ ( m ) . By Proposition 2.4 (c), h r f ∈ D ( Δ ) h_rf\in D(\Delta) h r f ∈ D ( Δ ) with Δ h r f = h r Δ f \Delta h_rf=h_r\Delta f Δ h r f = h r Δ f , which belongs to V V V because Ch ( h r v ) ≤ Ch ( v ) \operatorname{Ch}(h_rv)\leq\operatorname{Ch}(v) Ch ( h r v ) ≤ Ch ( v ) , and to L ∞ ( m ) L^\infty(\mathfrak{m}) L ∞ ( m ) by the weak maximum principle.
:::
Lemma 4.4. Assume (S). Let f ∈ Test ∞ ( X ) f\in\operatorname{Test}^\infty(X) f ∈ Test ∞ ( X ) , set f r : = h r f f_r:=h_rf f r := h r f and G ( r ) : = Γ ( f r ) G(r):=\Gamma(f_r) G ( r ) := Γ ( f r ) for r ≥ 0 r\geq0 r ≥ 0 , and let ψ ∈ C 2 ( [ 0 , ∞ ) ) \psi\in C^2([0,\infty)) ψ ∈ C 2 ([ 0 , ∞ )) with ψ ( 0 ) = 0 \psi(0)=0 ψ ( 0 ) = 0 and ψ ′ \psi' ψ ′ ,ψ ′ ′ \psi'' ψ ′′ bounded. Then:
r ↦ f r r\mapsto f_r r ↦ f r belongs to C 1 ( [ 0 , ∞ ) ; V ) C^1([0,\infty);V) C 1 ([ 0 , ∞ ) ; V ) , with derivative Δ f r = h r Δ f \Delta f_r=h_r\Delta f Δ f r = h r Δ f ;
G ∈ C 1 ( [ 0 , ∞ ) ; L 1 ( m ) ) G\in C^1([0,\infty);L^1(\mathfrak{m})) G ∈ C 1 ([ 0 , ∞ ) ; L 1 ( m )) with G ′ ( r ) = 2 Γ ( f r , Δ f r ) G'(r)=2\Gamma(f_r,\Delta f_r) G ′ ( r ) = 2Γ ( f r , Δ f r ) , and sup r ∈ [ 0 , T ] ∥ G ( r ) ∥ L ∞ ( m ) < ∞ \sup_{r\in[0,T]}\|G(r)\|_{L^\infty(\mathfrak{m})}<\infty sup r ∈ [ 0 , T ] ∥ G ( r ) ∥ L ∞ ( m ) < ∞ for every T > 0 T>0 T > 0 ;
ψ ∘ G ∈ C 1 ( [ 0 , ∞ ) ; L 1 ( m ) ) ∩ C ( [ 0 , ∞ ) ; L 2 ( m ) ) \psi\circ G\in C^1([0,\infty);L^1(\mathfrak{m}))\cap C([0,\infty);L^2(\mathfrak{m})) ψ ∘ G ∈ C 1 ([ 0 , ∞ ) ; L 1 ( m )) ∩ C ([ 0 , ∞ ) ; L 2 ( m )) with ( ψ ∘ G ) ′ ( r ) = ψ ′ ( G ( r ) ) G ′ ( r ) (\psi\circ G)'(r)=\psi'(G(r))G'(r) ( ψ ∘ G ) ′ ( r ) = ψ ′ ( G ( r )) G ′ ( r ) .
Proof. (i). Set v : = Δ f ∈ V v:=\Delta f\in V v := Δ f ∈ V . For σ , τ ≥ 0 \sigma,\tau\geq0 σ , τ ≥ 0 we have ∥ h σ + τ v − h σ v ∥ V ≤ ∥ h τ v − v ∥ V \|h_{\sigma+\tau}v-h_\sigma v\|_V\leq\|h_\tau v-v\|_V ∥ h σ + τ v − h σ v ∥ V ≤ ∥ h τ v − v ∥ V , by the L 2 L^2 L 2 -contractivity of h σ h_\sigma h σ and Ch ( h σ w ) ≤ Ch ( w ) \operatorname{Ch}(h_\sigma w)\leq\operatorname{Ch}(w) Ch ( h σ w ) ≤ Ch ( w ) , and h τ v → v h_\tau v\to v h τ v → v in V V V as τ ↓ 0 \tau\downarrow0 τ ↓ 0 , as shown in the proof of Lemma 4.2 . Hence σ ↦ h σ v \sigma\mapsto h_\sigma v σ ↦ h σ v is uniformly continuous from [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) to V V V . Since
f r − f r ′ = ∫ r ′ r h σ v d σ f_r-f_{r'}=\int_{r'}^r h_\sigma v\,\mathrm{d}\sigma f r − f r ′ = ∫ r ′ r h σ v d σ
in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) for 0 ≤ r ′ < r 0\leq r'<r 0 ≤ r ′ < r and the integrand is continuous with values in V V V , the identity holds in V V V , and (i) follows.
(ii). This follows from (i) and the continuity of Γ : V × V → L 1 ( m ) \Gamma:V\times V\to L^1(\mathfrak{m}) Γ : V × V → L 1 ( m ) ; the L ∞ L^\infty L ∞ bound is Lemma 4.3 .
(iii). Fix T > 0 T>0 T > 0 and M : = sup [ 0 , T ] ∥ G ∥ L ∞ ( m ) M:=\sup_{[0,T]}\|G\|_{L^\infty(\mathfrak{m})} M := sup [ 0 , T ] ∥ G ∥ L ∞ ( m ) . Since ∥ w ∥ L 2 ( m ) 2 ≤ ∥ w ∥ L ∞ ( m ) ∥ w ∥ L 1 ( m ) \|w\|_{L^2(\mathfrak{m})}^2\leq\|w\|_{L^\infty(\mathfrak{m})}\|w\|_{L^1(\mathfrak{m})} ∥ w ∥ L 2 ( m ) 2 ≤ ∥ w ∥ L ∞ ( m ) ∥ w ∥ L 1 ( m ) , G G G is continuous with values in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) on [ 0 , T ] [0,T] [ 0 , T ] , and so is ψ ∘ G \psi\circ G ψ ∘ G , ψ \psi ψ being Lipschitz with ψ ( 0 ) = 0 \psi(0)=0 ψ ( 0 ) = 0 . For r , r + τ ∈ [ 0 , T ] r,r+\tau\in[0,T] r , r + τ ∈ [ 0 , T ] , write Δ τ G : = G ( r + τ ) − G ( r ) \Delta_\tau G:=G(r+\tau)-G(r) Δ τ G := G ( r + τ ) − G ( r ) and Θ τ : = ∫ 0 1 ψ ′ ( G ( r ) + θ Δ τ G ) d θ \Theta_\tau:=\int_0^1\psi'(G(r)+\theta\Delta_\tau G)\,\mathrm{d}\theta Θ τ := ∫ 0 1 ψ ′ ( G ( r ) + θ Δ τ G ) d θ , so that
ψ ( G ( r + τ ) ) − ψ ( G ( r ) ) τ − ψ ′ ( G ( r ) ) G ′ ( r ) = Θ τ ( Δ τ G τ − G ′ ( r ) ) + ( Θ τ − ψ ′ ( G ( r ) ) ) G ′ ( r ) . \frac{\psi(G(r+\tau))-\psi(G(r))}{\tau}-\psi'(G(r))G'(r)=\Theta_\tau\left(\frac{\Delta_\tau G}{\tau}-G'(r)\right)+\left(\Theta_\tau-\psi'(G(r))\right)G'(r). τ ψ ( G ( r + τ )) − ψ ( G ( r )) − ψ ′ ( G ( r )) G ′ ( r ) = Θ τ ( τ Δ τ G − G ′ ( r ) ) + ( Θ τ − ψ ′ ( G ( r )) ) G ′ ( r ) .
The first term tends to 0 0 0 in L 1 ( m ) L^1(\mathfrak{m}) L 1 ( m ) because ∣ Θ τ ∣ ≤ ∥ ψ ′ ∥ ∞ |\Theta_\tau|\leq\|\psi'\|_\infty ∣ Θ τ ∣ ≤ ∥ ψ ′ ∥ ∞ . For the second,
∣ Θ τ − ψ ′ ( G ( r ) ) ∣ ≤ min { 2 ∥ ψ ′ ∥ ∞ , ∥ ψ ′ ′ ∥ ∞ ∣ Δ τ G ∣ } |\Theta_\tau-\psi'(G(r))|\leq\min\left\{2\|\psi'\|_\infty,\|\psi''\|_\infty|\Delta_\tau G|\right\} ∣ Θ τ − ψ ′ ( G ( r )) ∣ ≤ min { 2∥ ψ ′ ∥ ∞ , ∥ ψ ′′ ∥ ∞ ∣ Δ τ G ∣ }
tends to 0 0 0 in m \mathfrak{m} m -measure, and dominated convergence along subsequences gives convergence to 0 0 0 in L 1 ( m ) L^1(\mathfrak{m}) L 1 ( m ) . The continuity of r ↦ ψ ′ ( G ( r ) ) G ′ ( r ) r\mapsto\psi'(G(r))G'(r) r ↦ ψ ′ ( G ( r )) G ′ ( r ) in L 1 ( m ) L^1(\mathfrak{m}) L 1 ( m ) follows in the same way.
:::
The Gamma2 functional
Definition 4.5. For u , v ∈ Test ∞ ( X ) u,v \in\operatorname{Test}^{\infty}(X) u , v ∈ Test ∞ ( X ) and φ ∈ V \varphi\in\mathbb{V} φ ∈ V we set
Γ 2 [ u , v ] ( φ ) : = − 1 2 E ( Γ ( u , v ) , φ ) − 1 2 ∫ X ( Γ ( u , Δ v ) + Γ ( v , Δ u ) ) φ d m . \Gamma_{2}[u,v](\varphi):=-\frac{1}{2}\mathcal{E}(\Gamma(u,v),\varphi)-\frac{1}{2}\int_{X}\left(\Gamma(u,\Delta v)+\Gamma(v,\Delta u)\right)\varphi\,\mathrm{d}m. Γ 2 [ u , v ] ( φ ) := − 2 1 E ( Γ ( u , v ) , φ ) − 2 1 ∫ X ( Γ ( u , Δ v ) + Γ ( v , Δ u ) ) φ d m .
Γ 2 K [ u , v ] ( φ ) : = Γ 2 [ u , v ] ( φ ) − K ∫ X Γ ( u , v ) φ d m \Gamma_{2}^{K}[u,v](\varphi):=\Gamma_{2}[u,v](\varphi)-K\int_{X}\Gamma(u,v)\varphi\,\mathrm{d}m Γ 2 K [ u , v ] ( φ ) := Γ 2 [ u , v ] ( φ ) − K ∫ X Γ ( u , v ) φ d m , Γ 2 [ u ] : = Γ 2 [ u , u ] \Gamma_{2}[u]:=\Gamma_{2}[u,u] Γ 2 [ u ] := Γ 2 [ u , u ] and Γ 2 K [ u ] : = Γ 2 K [ u , u ] \Gamma_{2}^{K}[u]:=\Gamma_{2}^{K}[u,u] Γ 2 K [ u ] := Γ 2 K [ u , u ] . For u , v , w ∈ Test ∞ ( X ) u,v,w \in\operatorname{Test}^{\infty}(X) u , v , w ∈ Test ∞ ( X ) we set
H [ u ] ( v , w ) : = 1 2 ( Γ ( Γ ( u , v ) , w ) + Γ ( Γ ( u , w ) , v ) − Γ ( u , Γ ( v , w ) ) ) . H[u](v,w):=\frac{1}{2}\left(\Gamma(\Gamma(u,v),w)+\Gamma(\Gamma(u,w),v)-\Gamma(u,\Gamma(v,w))\right). H [ u ] ( v , w ) := 2 1 ( Γ ( Γ ( u , v ) , w ) + Γ ( Γ ( u , w ) , v ) − Γ ( u , Γ ( v , w )) ) .
Remark 4.6. By Lemma 2.5(v) and polarisation, Γ ( u , v ) ∈ V ∞ \Gamma(u,v) \in\mathbb{V}_{\infty} Γ ( u , v ) ∈ V ∞ for u , v ∈ Test ∞ ( X ) u,v \in\operatorname{Test}^{\infty}(X) u , v ∈ Test ∞ ( X ) , so E ( Γ ( u , v ) , ⋅ ) \mathcal{E}(\Gamma(u,v),\cdot) E ( Γ ( u , v ) , ⋅ ) is continuous on V \mathbb{V} V ; moreover Γ ( u , Δ v ) ∈ L 2 ( m ) \Gamma(u,\Delta v) \in L^{2}(m) Γ ( u , Δ v ) ∈ L 2 ( m ) , since ∣ D u ∣ ∈ L ∞ ( m ) |Du| \in L^{\infty}(m) ∣ D u ∣ ∈ L ∞ ( m ) and Δ v ∈ V \Delta v \in\mathbb{V} Δ v ∈ V . Hence Γ 2 [ u , v ] \Gamma_{2}[u,v] Γ 2 [ u , v ] and Γ 2 K [ u , v ] \Gamma_{2}^{K}[u,v] Γ 2 K [ u , v ] are continuous linear functionals on V \mathbb{V} V , symmetric and bilinear in ( u , v ) (u,v) ( u , v ) , and Γ 2 K [ u , v ] = 1 4 ( Γ 2 K [ u + v ] − Γ 2 K [ u − v ] ) \Gamma_{2}^{K}[u,v]=\frac{1}{4}\left(\Gamma_{2}^{K}[u+v]-\Gamma_{2}^{K}[u-v]\right) Γ 2 K [ u , v ] = 4 1 ( Γ 2 K [ u + v ] − Γ 2 K [ u − v ] ) . Similarly, H [ u ] ( v , w ) ∈ L 2 ( m ) H[u](v,w) \in L^{2}(m) H [ u ] ( v , w ) ∈ L 2 ( m ) is symmetric in ( v , w ) (v,w) ( v , w ) , and
Γ ( Γ ( u ) , w ) = 2 H [ u ] ( u , w ) m-a.e. (5) \Gamma(\Gamma(u),w)=2H[u](u,w) \qquad\text{m-a.e.} \tag*{(5)} Γ ( Γ ( u ) , w ) = 2 H [ u ] ( u , w ) m-a.e. (5)
Lemma 4.7. Assume (S). Then Γ 2 K [ u ] ( φ ) ≥ 0 \Gamma_{2}^{K}[u](\varphi)\geq0 Γ 2 K [ u ] ( φ ) ≥ 0 for every u ∈ Test ∞ ( X ) u \in\operatorname{Test}^{\infty}(X) u ∈ Test ∞ ( X ) and every φ ∈ V + \varphi\in\mathbb{V}_{+} φ ∈ V + .
Proof. For φ ∈ Test + ∞ ( X ) \varphi\in\operatorname{Test}^{\infty}_{+}(X) φ ∈ Test + ∞ ( X ) , Proposition 2.4(b) gives 1 2 ∫ X Γ ( u ) Δ φ d m = − 1 2 E ( Γ ( u ) , φ ) \frac{1}{2}\int_{X}\Gamma(u)\Delta\varphi\,\mathrm{d}m=-\frac{1}{2}\mathcal{E}(\Gamma(u),\varphi) 2 1 ∫ X Γ ( u ) Δ φ d m = − 2 1 E ( Γ ( u ) , φ ) , so (BE) reads Γ 2 K [ u ] ( φ ) ≥ 0 \Gamma_{2}^{K}[u](\varphi)\geq0 Γ 2 K [ u ] ( φ ) ≥ 0 . The general case follows from Lemma 2.5(iv) and the continuity of Γ 2 K [ u ] \Gamma_{2}^{K}[u] Γ 2 K [ u ] on V \mathbb{V} V . □
Lemma 4.8. Assume (S). Let u , v , w ∈ Test ∞ ( X ) u,v,w \in\operatorname{Test}^{\infty}(X) u , v , w ∈ Test ∞ ( X ) and φ ∈ V ∞ \varphi\in\mathbb{V}_{\infty} φ ∈ V ∞ . Then
Γ 2 [ u , v w ] ( φ ) = Γ 2 [ u , w ] ( v φ ) + Γ 2 [ u , v ] ( w φ ) + 2 ∫ X H [ u ] ( v , w ) φ d m (6) \Gamma_{2}[u,vw](\varphi)=\Gamma_{2}[u,w](v\varphi)+\Gamma_{2}[u,v](w\varphi)+2\int_{X}H[u](v,w)\varphi\,\mathrm{d}m \tag*{(6)} Γ 2 [ u , v w ] ( φ ) = Γ 2 [ u , w ] ( v φ ) + Γ 2 [ u , v ] ( w φ ) + 2 ∫ X H [ u ] ( v , w ) φ d m (6)
and
H [ w 2 ] ( w , w ) = 2 w H [ w ] ( w , w ) + 2 Γ ( w ) 2 m-a.e. (7) H[w^{2}](w,w)=2wH[w](w,w)+2\Gamma(w)^{2} \qquad\text{m-a.e.} \tag*{(7)} H [ w 2 ] ( w , w ) = 2 w H [ w ] ( w , w ) + 2Γ ( w ) 2 m-a.e. (7)
Proof. All products below belong to V ∞ \mathbb{V}_{\infty} V ∞ by the Leibniz rule and Lemma 2.5(v)–(vi). For a ∈ Test ∞ ( X ) a \in\operatorname{Test}^{\infty}(X) a ∈ Test ∞ ( X ) and b ∈ V ∞ b \in\mathbb{V}_{\infty} b ∈ V ∞ , the Leibniz rule and Proposition 2.4(b) give
E ( a b , φ ) = E ( b , a φ ) − 2 ∫ X φ Γ ( a , b ) d m − ∫ X b φ Δ a d m . \mathcal{E}(ab,\varphi)=\mathcal{E}(b,a\varphi)-2\int_{X}\varphi\Gamma(a,b)\,\mathrm{d}m-\int_{X}b\varphi\Delta a\,\mathrm{d}m. E ( ab , φ ) = E ( b , a φ ) − 2 ∫ X φ Γ ( a , b ) d m − ∫ X b φ Δ a d m .
Since Γ ( u , v w ) = v Γ ( u , w ) + w Γ ( u , v ) \Gamma(u,vw)=v\Gamma(u,w)+w\Gamma(u,v) Γ ( u , v w ) = v Γ ( u , w ) + w Γ ( u , v ) , applying this with ( a , b ) = ( v , Γ ( u , w ) ) (a,b)=(v,\Gamma(u,w)) ( a , b ) = ( v , Γ ( u , w )) and ( a , b ) = ( w , Γ ( u , v ) ) (a,b)=(w,\Gamma(u,v)) ( a , b ) = ( w , Γ ( u , v )) yields
E ( Γ ( u , v w ) , φ ) = E ( Γ ( u , w ) , v φ ) + E ( Γ ( u , v ) , w φ ) − 2 ∫ X φ ( Γ ( v , Γ ( u , w ) ) + Γ ( w , Γ ( u , v ) ) ) d m − ∫ X φ ( Γ ( u , w ) Δ v + Γ ( u , v ) Δ w ) d m . \begin{aligned}
\mathcal{E}(\Gamma(u,vw),\varphi)
={}&\mathcal{E}(\Gamma(u,w),v\varphi)+\mathcal{E}(\Gamma(u,v),w\varphi)\\
&-2\int_{X}\varphi\left(\Gamma(v,\Gamma(u,w))+\Gamma(w,\Gamma(u,v))\right)\,\mathrm{d}m\\
&-\int_{X}\varphi\left(\Gamma(u,w)\Delta v+\Gamma(u,v)\Delta w\right)\,\mathrm{d}m.
\end{aligned} E ( Γ ( u , v w ) , φ ) = E ( Γ ( u , w ) , v φ ) + E ( Γ ( u , v ) , w φ ) − 2 ∫ X φ ( Γ ( v , Γ ( u , w )) + Γ ( w , Γ ( u , v )) ) d m − ∫ X φ ( Γ ( u , w ) Δ v + Γ ( u , v ) Δ w ) d m .
By Proposition 2.4(b) and the Leibniz rule,
Γ ( u , Δ ( v w ) ) = v Γ ( u , Δ w ) + w Γ ( u , Δ v ) + Γ ( u , v ) Δ w + Γ ( u , w ) Δ v + 2 Γ ( u , Γ ( v , w ) ) , \Gamma(u,\Delta(vw))=v\Gamma(u,\Delta w)+w\Gamma(u,\Delta v)+\Gamma(u,v)\Delta w+\Gamma(u,w)\Delta v+2\Gamma(u,\Gamma(v,w)), Γ ( u , Δ ( v w )) = v Γ ( u , Δ w ) + w Γ ( u , Δ v ) + Γ ( u , v ) Δ w + Γ ( u , w ) Δ v + 2Γ ( u , Γ ( v , w )) ,
and Γ ( v w , Δ u ) = v Γ ( w , Δ u ) + w Γ ( v , Δ u ) \Gamma(vw,\Delta u)=v\Gamma(w,\Delta u)+w\Gamma(v,\Delta u) Γ ( v w , Δ u ) = v Γ ( w , Δ u ) + w Γ ( v , Δ u ) . Inserting these identities into the definition of Γ 2 [ u , v w ] ( φ ) \Gamma_{2}[u,vw](\varphi) Γ 2 [ u , v w ] ( φ ) , the terms containing Δ v \Delta v Δ v and Δ w \Delta w Δ w without a gradient cancel, and the remaining terms give (4.2). For (4.3), the Leibniz rule gives
Γ ( w 2 , w ) = 2 w Γ ( w ) , Γ ( w 2 , Γ ( w ) ) = 2 w Γ ( w , Γ ( w ) ) , \Gamma(w^{2},w)=2w\Gamma(w), \qquad\Gamma(w^{2},\Gamma(w))=2w\Gamma(w,\Gamma(w)), Γ ( w 2 , w ) = 2 w Γ ( w ) , Γ ( w 2 , Γ ( w )) = 2 w Γ ( w , Γ ( w )) ,
Γ ( 2 w Γ ( w ) , w ) = 2 Γ ( w ) 2 + 2 w Γ ( Γ ( w ) , w ) , \Gamma(2w\Gamma(w),w)=2\Gamma(w)^{2}+2w\Gamma(\Gamma(w),w), Γ ( 2 w Γ ( w ) , w ) = 2Γ ( w ) 2 + 2 w Γ ( Γ ( w ) , w ) ,
so H [ w 2 ] ( w , w ) = Γ ( Γ ( w 2 , w ) , w ) − 1 2 Γ ( Γ ( w 2 ) , Γ ( w ) ) = 2 Γ ( w ) 2 + w Γ ( Γ ( w ) , w ) H[w^{2}](w,w)=\Gamma(\Gamma(w^{2},w),w)-\frac{1}{2}\Gamma(\Gamma(w^{2}),\Gamma(w))=2\Gamma(w)^{2}+w\Gamma(\Gamma(w),w) H [ w 2 ] ( w , w ) = Γ ( Γ ( w 2 , w ) , w ) − 2 1 Γ ( Γ ( w 2 ) , Γ ( w )) = 2Γ ( w ) 2 + w Γ ( Γ ( w ) , w ) , and (4.1) concludes. □
In the smooth setting, Γ 2 K [ u ] ( φ ) = ∫ X ( Γ 2 ( u ) − K Γ ( u ) ) φ d m \Gamma_{2}^{K}[u](\varphi)=\int_{X}(\Gamma_{2}(u)-K\Gamma(u))\varphi\,\mathrm{d}m Γ 2 K [ u ] ( φ ) = ∫ X ( Γ 2 ( u ) − K Γ ( u )) φ d m for a function Γ 2 ( u ) \Gamma_{2}(u) Γ 2 ( u ) ; here Γ 2 K [ u ] \Gamma_{2}^{K}[u] Γ 2 K [ u ] is only a functional on V \mathbb{V} V , and the next lemma shows that, being nonnegative by (BE), it is given by a measure. This is where quasi-regularity enters the proof.
Lemma 4.9. Assume (S). For every u ∈ Test ∞ ( X ) u \in\operatorname{Test}^{\infty}(X) u ∈ Test ∞ ( X ) there is a unique positive Borel measure μ u \mu_u μ u on X X X , not charging E \mathcal{E} E -polar sets, such that Γ 2 K [ u ] ( φ ) = ∫ X φ ~ d μ u \Gamma_2^K[u](\varphi) = \int_X \widetilde{\varphi}\,d\mu_u Γ 2 K [ u ] ( φ ) = ∫ X φ d μ u for every φ ∈ V \varphi\in\mathbb{V} φ ∈ V . Moreover, μ u ( B ) < ∞ \mu_u(B) < \infty μ u ( B ) < ∞ for every admissible ball B B B , and Γ 2 K [ u ] ( φ ) = ∫ X φ d μ u \Gamma_2^K[u](\varphi) = \int_X \varphi\,d\mu_u Γ 2 K [ u ] ( φ ) = ∫ X φ d μ u for every continuous φ ∈ V \varphi\in\mathbb{V} φ ∈ V .
Proof. Set ℓ : = Γ 2 K [ u ] \ell:= \Gamma_2^K[u] ℓ := Γ 2 K [ u ] . By Remark 4.6, ℓ \ell ℓ is a continuous linear functional on V \mathbb{V} V , and by Lemma 4.7, ℓ ( φ ) ≥ 0 \ell(\varphi) \ge0 ℓ ( φ ) ≥ 0 for every φ ∈ V + \varphi\in\mathbb{V}_{+} φ ∈ V + . By Lemma 4.1, E \mathcal{E} E is quasi-regular, so Proposition 3.8(ii) gives a unique positive Borel measure μ u \mu_u μ u , not charging E \mathcal{E} E -polar sets, with ℓ ( φ ) = ∫ X φ ~ d μ u \ell(\varphi) = \int_X \widetilde{\varphi}\,d\mu_u ℓ ( φ ) = ∫ X φ d μ u for every φ ∈ V \varphi\in\mathbb{V} φ ∈ V . If φ ∈ V \varphi\in\mathbb{V} φ ∈ V is continuous, then φ ~ = φ \widetilde{\varphi} = \varphi φ = φ q.e. by Remark 3.9(i), hence μ u \mu_u μ u -a.e., and ℓ ( φ ) = ∫ X φ d μ u \ell(\varphi) = \int_X \varphi\,d\mu_u ℓ ( φ ) = ∫ X φ d μ u . If B = B r ( x ) B = B_r(x) B = B r ( x ) is admissible, the function β : = ( 1 − d ( ⋅ , B ) / r ) + \beta:= (1-d(\cdot,B)/r)^{+} β := ( 1 − d ( ⋅ , B ) / r ) + is continuous, belongs to V \mathbb{V} V by Lemma 2.3 and satisfies β ≥ χ B \beta\ge\chi_B β ≥ χ B , so μ u ( B ) ≤ ∫ X β d μ u = Γ 2 K [ u ] ( β ) < ∞ \mu_u(B) \le\int_X \beta\,d\mu_u = \Gamma_2^K[u](\beta) < \infty μ u ( B ) ≤ ∫ X β d μ u = Γ 2 K [ u ] ( β ) < ∞ .
We write μ u = ρ u m + μ u s \mu_u = \rho_u m+\mu_u^s μ u = ρ u m + μ u s for the Lebesgue decomposition of μ u \mu_u μ u with respect to m m m , with ρ u ≥ 0 \rho_u \ge0 ρ u ≥ 0 and μ u s ⊥ m \mu_u^s \perp m μ u s ⊥ m .
Lemma 4.10. Assume (S). Let Λ 1 , Λ 2 \Lambda_1,\Lambda_2 Λ 1 , Λ 2 be positive Borel measures on X X X , finite on admissible balls, such that ∫ X φ d Λ 1 ≥ ∫ X φ d Λ 2 \int_X \varphi\,d\Lambda_1 \ge\int_X \varphi\,d\Lambda_2 ∫ X φ d Λ 1 ≥ ∫ X φ d Λ 2 for every nonnegative φ ∈ LIP b ( X ) \varphi\in\operatorname{LIP}_b(X) φ ∈ LIP b ( X ) vanishing outside some admissible ball. Then Λ 1 ≥ Λ 2 \Lambda_1 \ge\Lambda_2 Λ 1 ≥ Λ 2 .
Proof. Fix an admissible ball B B B ; then \Lambda:= (\Lambda_1-\Lambda_2)\mathbin{\vrule height 1.4ex depth -0.3ex width .07ex\vrule height .07ex depth -0.02ex width .7ex}B is a finite signed Borel measure. For an open set U ⊆ B U \subseteq B U ⊆ B , the functions φ k : = min { 1 , k d ( ⋅ , X ∖ U ) } \varphi_k := \min\{1,k d(\cdot,X\setminus U)\} φ k := min { 1 , k d ( ⋅ , X ∖ U )} are Lipschitz, vanish outside U U U and satisfy 0 ≤ φ k ↑ χ U 0 \le\varphi_k \uparrow\chi_U 0 ≤ φ k ↑ χ U , so Λ ( U ) ≥ 0 \Lambda(U) \ge0 Λ ( U ) ≥ 0 by dominated convergence. Let B = P ∪ N B = P \cup N B = P ∪ N be a Hahn decomposition for Λ \Lambda Λ and δ > 0 \delta> 0 δ > 0 . The finite Borel measure ∣ Λ ∣ |\Lambda| ∣Λ∣ on the metric space X X X is outer regular, so there is an open set V ⊇ N V \supseteq N V ⊇ N with ∣ Λ ∣ ( V ∖ N ) < δ |\Lambda|(V\setminus N) < \delta ∣Λ∣ ( V ∖ N ) < δ ; then U : = V ∩ B U := V \cap B U := V ∩ B is open and 0 ≤ Λ ( U ) ≤ Λ ( N ) + δ 0 \le\Lambda(U) \le\Lambda(N)+\delta 0 ≤ Λ ( U ) ≤ Λ ( N ) + δ . Hence Λ ( N ) = 0 \Lambda(N)=0 Λ ( N ) = 0 , i.e. \Lambda_1\mathbin{\vrule height 1.4ex depth -0.3ex width .07ex\vrule height .07ex depth -0.02ex width .7ex}B \ge\Lambda_2\mathbin{\vrule height 1.4ex depth -0.3ex width .07ex\vrule height .07ex depth -0.02ex width .7ex}B . Since X X X is covered by countably many admissible balls (Lemma 2.1), Λ 1 ≥ Λ 2 \Lambda_1 \ge\Lambda_2 Λ 1 ≥ Λ 2 .
The L 1 L^1 L 1 Bakry–Émery estimate
We follow the self-improvement argument of Bakry [8 ] , in the form of [25 Section 3] . This gives the improved Bochner inequality Γ ( Γ ( f ) ) ≤ 4 ρ f Γ ( f ) \Gamma(\Gamma(f)) \le4\rho_f\Gamma(f) Γ ( Γ ( f )) ≤ 4 ρ f Γ ( f ) (Proposition 4.11). Then, to obtain the L 1 L^1 L 1 Bakry–Émery estimate, we differentiate
A ( s ) : = ∫ X h s φ ψ ( Γ ( h t − s f ) ) d m , s ∈ [ 0 , t ] , ψ ( r ) : = r + ε − ε , A(s):=\int_X h_s\varphi\psi(\Gamma(h_{t-s}f))\,\mathrm{d}m,\qquad s\in[0,t],\qquad\psi(r):=\sqrt{r+\varepsilon}-\sqrt{\varepsilon}, A ( s ) := ∫ X h s φ ψ ( Γ ( h t − s f )) d m , s ∈ [ 0 , t ] , ψ ( r ) := r + ε − ε ,
for nonnegative φ ∈ L 1 ( m ) ∩ L ∞ ( m ) \varphi\in L^1(m)\cap L^\infty(m) φ ∈ L 1 ( m ) ∩ L ∞ ( m ) ; by Proposition 4.11, A ′ ≥ K A A' \ge KA A ′ ≥ K A up to an error vanishing as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . Lemma 4.12 is the elementary differentiation rule used in this step.
Proposition 4.11 (Improved Bochner inequality). Assume (S) and let f ∈ Test ∞ ( X ) f\in\operatorname{Test}^{\infty}(X) f ∈ Test ∞ ( X ) . Then
Γ ( Γ ( f ) ) ≤ 4 ρ f Γ ( f ) m -a.e. (8) \Gamma(\Gamma(f)) \le4\rho_f\Gamma(f)\qquad m\text{-a.e.} \tag*{(8)} Γ ( Γ ( f )) ≤ 4 ρ f Γ ( f ) m -a.e. (8) Proof. Step 1: expansion. Fix w ∈ Test ∞ ( X ) w\in\operatorname{Test}^{\infty}(X) w ∈ Test ∞ ( X ) and λ , c ∈ R \lambda,c\in\mathbb{R} λ , c ∈ R . By Lemma 2.5(vi), u : = λ f + w 2 − 2 c w u:=\lambda f+w^2-2cw u := λ f + w 2 − 2 c w belongs to Test ∞ ( X ) \operatorname{Test}^{\infty}(X) Test ∞ ( X ) . We write H f : = H [ f ] ( w , w ) H_f:=H[f](w,w) H f := H [ f ] ( w , w ) and H w : = H [ w ] ( w , w ) H_w:=H[w](w,w) H w := H [ w ] ( w , w ) . We claim that, for every φ ∈ V ∞ \varphi\in\mathbb{V}_{\infty} φ ∈ V ∞ ,
Γ 2 K [ u ] ( φ ) = λ 2 Γ 2 K [ f ] ( φ ) + 4 λ Γ 2 K [ f , w ] ( ( w − c ) φ ) + 4 Γ 2 K [ w ] ( ( w − c ) 2 φ ) \Gamma_2^K[u](\varphi)=\lambda^2\Gamma_2^K[f](\varphi)+4\lambda\Gamma_2^K[f,w]((w-c)\varphi)+4\Gamma_2^K[w]((w-c)^2\varphi) Γ 2 K [ u ] ( φ ) = λ 2 Γ 2 K [ f ] ( φ ) + 4 λ Γ 2 K [ f , w ] (( w − c ) φ ) + 4 Γ 2 K [ w ] (( w − c ) 2 φ ) + ∫ X ( 4 λ H f + 8 ( w − c ) H w + 4 Γ ( w ) 2 ) φ d m . (9) \qquad+\int_X\left(4\lambda H_f+8(w-c)H_w+4\Gamma(w)^2\right)\varphi\,\mathrm{d}m. \tag*{(9)} + ∫ X ( 4 λ H f + 8 ( w − c ) H w + 4Γ ( w ) 2 ) φ d m . (9) Indeed, with g : = w 2 − 2 c w g:=w^2-2cw g := w 2 − 2 c w we have Γ 2 [ u ] = λ 2 Γ 2 [ f ] + 2 λ Γ 2 [ f , g ] + Γ 2 [ g ] \Gamma_2[u]=\lambda^2\Gamma_2[f]+2\lambda\Gamma_2[f,g]+\Gamma_2[g] Γ 2 [ u ] = λ 2 Γ 2 [ f ] + 2 λ Γ 2 [ f , g ] + Γ 2 [ g ] . By (4.2),
Γ 2 [ f , g ] ( φ ) = 2 Γ 2 [ f , w ] ( ( w − c ) φ ) + 2 ∫ X H f φ d m , \Gamma_2[f,g](\varphi)=2\Gamma_2[f,w]((w-c)\varphi)+2\int_X H_f\varphi\,\mathrm{d}m, Γ 2 [ f , g ] ( φ ) = 2 Γ 2 [ f , w ] (( w − c ) φ ) + 2 ∫ X H f φ d m , Γ 2 [ w , w 2 ] ( φ ) = 2 Γ 2 [ w ] ( w φ ) + 2 ∫ X H w φ d m , \Gamma_2[w,w^2](\varphi)=2\Gamma_2[w](w\varphi)+2\int_X H_w\varphi\,\mathrm{d}m, Γ 2 [ w , w 2 ] ( φ ) = 2 Γ 2 [ w ] ( w φ ) + 2 ∫ X H w φ d m , and, using also the symmetry of Γ 2 \Gamma_2 Γ 2 and (4.3),
Γ 2 [ w 2 ] ( φ ) = 2 Γ 2 [ w 2 , w ] ( w φ ) + 2 ∫ X H [ w 2 ] ( w , w ) φ d m \Gamma_2[w^2](\varphi)=2\Gamma_2[w^2,w](w\varphi)+2\int_X H[w^2](w,w)\varphi\,\mathrm{d}m Γ 2 [ w 2 ] ( φ ) = 2 Γ 2 [ w 2 , w ] ( w φ ) + 2 ∫ X H [ w 2 ] ( w , w ) φ d m = 4 Γ 2 [ w ] ( w 2 φ ) + ∫ X ( 8 w H w + 4 Γ ( w ) 2 ) φ d m . =4\Gamma_2[w](w^2\varphi)+\int_X\left(8wH_w+4\Gamma(w)^2\right)\varphi\,\mathrm{d}m. = 4 Γ 2 [ w ] ( w 2 φ ) + ∫ X ( 8 w H w + 4Γ ( w ) 2 ) φ d m .
Hence
Γ 2 [ g ] ( φ ) = Γ 2 [ w 2 ] ( φ ) − 4 c Γ 2 [ w 2 , w ] ( φ ) + 4 c 2 Γ 2 [ w ] ( φ ) = 4 Γ 2 [ w ] ( ( w − c ) 2 φ ) + ∫ X ( 8 ( w − c ) H w + 4 Γ ( w ) 2 ) φ d m . \begin{aligned}
\Gamma_{2}[g](\varphi)=\Gamma_{2}[w^{2}](\varphi)-4c\Gamma_{2}[w^{2},w](\varphi)+4c^{2}\Gamma_{2}[w](\varphi) \\
&=4\Gamma_{2}[w]\bigl((w-c)^{2}\varphi\bigr)+\int_X\bigl(8(w-c)H_w+4\Gamma(w)^{2}\bigr)\varphi\,\mathrm{d}m.
\end{aligned} Γ 2 [ g ] ( φ ) = Γ 2 [ w 2 ] ( φ ) − 4 c Γ 2 [ w 2 , w ] ( φ ) + 4 c 2 Γ 2 [ w ] ( φ ) = 4 Γ 2 [ w ] ( ( w − c ) 2 φ ) + ∫ X ( 8 ( w − c ) H w + 4Γ ( w ) 2 ) φ d m .
Finally, Γ ( u ) = λ 2 Γ ( f ) + 4 λ ( w − c ) Γ ( f , w ) + 4 ( w − c ) 2 Γ ( w ) \Gamma(u)=\lambda^{2}\Gamma(f)+4\lambda(w-c)\Gamma(f,w)+4(w-c)^{2}\Gamma(w) Γ ( u ) = λ 2 Γ ( f ) + 4 λ ( w − c ) Γ ( f , w ) + 4 ( w − c ) 2 Γ ( w ) by the Leibniz rule, and subtracting K ∫ X Γ ( u ) φ d m K\int_X\Gamma(u)\varphi\,\mathrm{d}m K ∫ X Γ ( u ) φ d m yields (9) .
Step 2: an inequality between measures. Set ν f , w : = 1 4 ( μ f + w − μ f − w ) \nu_{f,w}:=\frac{1}{4}(\mu_{f+w}-\mu_{f-w}) ν f , w := 4 1 ( μ f + w − μ f − w ) , so that Γ 2 K [ f , w ] ( φ ) = ∫ X φ d ν f , w \Gamma_{2}^{K}[f,w](\varphi)=\int_X\varphi\,\mathrm{d}\nu_{f,w} Γ 2 K [ f , w ] ( φ ) = ∫ X φ d ν f , w for every continuous φ ∈ V \varphi\in\mathsf{V} φ ∈ V , and
Λ λ , c : = λ 2 μ f + 4 λ ( w − c ) ν f , w + 4 ( w − c ) 2 μ w + ( 4 λ H f + 8 ( w − c ) H w + 4 Γ ( w ) 2 ) m . \Lambda_{\lambda,c}:=\lambda^{2}\mu_f+4\lambda(w-c)\nu_{f,w}+4(w-c)^{2}\mu_w+\bigl(4\lambda H_f+8(w-c)H_w+4\Gamma(w)^{2}\bigr)m. Λ λ , c := λ 2 μ f + 4 λ ( w − c ) ν f , w + 4 ( w − c ) 2 μ w + ( 4 λ H f + 8 ( w − c ) H w + 4Γ ( w ) 2 ) m .
Splitting each term into its positive and negative parts, Λ λ , c \Lambda_{\lambda,c} Λ λ , c is the difference of two positive Borel measures which are finite on admissible balls: μ f \mu_f μ f , μ w \mu_w μ w , μ f ± w \mu_{f\pm w} μ f ± w are finite on admissible balls by Lemma 4.9 , w w w is bounded, and H f H_f H f , H w H_w H w , Γ ( w ) 2 \Gamma(w)^2 Γ ( w ) 2 belong to L 1 ( m ) L^1(m) L 1 ( m ) . Let φ ∈ LIP b ( X ) \varphi\in\operatorname{LIP}_b(X) φ ∈ LIP b ( X ) vanish outside an admissible ball B r ( x ) B_r(x) B r ( x ) ; then supp ( φ ) ⊆ B r ( x ) \operatorname{supp}(\varphi)\subseteq B_r(x) supp ( φ ) ⊆ B r ( x ) , so φ ∈ LIP f m ( X ) \varphi\in\operatorname{LIP}_{fm}(X) φ ∈ LIP f m ( X ) , and φ \varphi φ , ( w − c ) φ (w-c)\varphi ( w − c ) φ , ( w − c ) 2 φ (w-c)^2\varphi ( w − c ) 2 φ are continuous elements of V \mathsf{V} V . By Lemma 4.9 , the right-hand side of (9) equals ∫ X φ d Λ λ , c \int_X\varphi\,\mathrm{d}\Lambda_{\lambda,c} ∫ X φ d Λ λ , c , and by Lemmas 4.7 and 4.10 , Λ λ , c ≥ 0 \Lambda_{\lambda,c}\geq0 Λ λ , c ≥ 0 for every λ , c ∈ R \lambda,c\in\mathbb{R} λ , c ∈ R .
Step 3: densities. Fix an admissible ball B B B , let \sigma:=(m+\mu_f+\mu_w+\mu_{f+w}+\mu_{f-w})\mathbin{\vrule height 1.4ex depth -0.3ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}B , a finite measure, and write M M M , F F F , W W W , N N N for the densities with respect to σ \sigma σ of m\mathbin{\vrule height 1.4ex depth -0.3ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}B , \mu_f\mathbin{\vrule height 1.4ex depth -0.3ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}B , \mu_w\mathbin{\vrule height 1.4ex depth -0.3ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}B , \nu_{f,w}\mathbin{\vrule height 1.4ex depth -0.3ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}B . Fix Borel representatives of H f H_f H f , H w H_w H w , Γ ( w ) \Gamma(w) Γ ( w ) ; changing them on an m m m -null set Z Z Z changes the products with M M M only on the σ \sigma σ -null set Z ∩ { M > 0 } Z\cap\{M>0\} Z ∩ { M > 0 } . By Step 2, for each ( λ , c ) ∈ Q 2 (\lambda,c)\in\mathbb{Q}^{2} ( λ , c ) ∈ Q 2 the density
D λ , c : = λ 2 F + 4 λ ( w − c ) N + 4 ( w − c ) 2 W + ( 4 λ H f + 8 ( w − c ) H w + 4 Γ ( w ) 2 ) M D_{\lambda,c}:=\lambda^{2}F+4\lambda(w-c)N+4(w-c)^{2}W+\bigl(4\lambda H_f+8(w-c)H_w+4\Gamma(w)^{2}\bigr)M D λ , c := λ 2 F + 4 λ ( w − c ) N + 4 ( w − c ) 2 W + ( 4 λ H f + 8 ( w − c ) H w + 4Γ ( w ) 2 ) M
is nonnegative outside a σ \sigma σ -null set E λ , c E_{\lambda,c} E λ , c . Outside E : = ⋃ ( λ , c ) ∈ Q 2 E λ , c E:=\bigcup_{(\lambda,c)\in\mathbb{Q}^{2}}E_{\lambda,c} E := ⋃ ( λ , c ) ∈ Q 2 E λ , c , the polynomial ( λ , c ) ↦ D λ , c ( x ) (\lambda,c)\mapsto D_{\lambda,c}(x) ( λ , c ) ↦ D λ , c ( x ) is nonnegative on Q 2 \mathbb{Q}^{2} Q 2 , hence on R 2 \mathbb{R}^{2} R 2 , and the choice c = w ( x ) c=w(x) c = w ( x ) gives
λ 2 F ( x ) + 4 λ H f ( x ) M ( x ) + 4 Γ ( w ) ( x ) 2 M ( x ) ≥ 0 \lambda^{2}F(x)+4\lambda H_f(x)M(x)+4\Gamma(w)(x)^{2}M(x)\geq0 λ 2 F ( x ) + 4 λ H f ( x ) M ( x ) + 4Γ ( w ) ( x ) 2 M ( x ) ≥ 0
for all λ ∈ R \lambda\in\mathbb{R} λ ∈ R and x ∈ B ∖ E x\in B\setminus E x ∈ B ∖ E . Since m ( { M = 0 } ∩ B ) = 0 m(\{M=0\}\cap B)=0 m ({ M = 0 } ∩ B ) = 0 , and since μ f = ρ f m + μ f s \mu_f=\rho_fm+\mu_f^{s} μ f = ρ f m + μ f s with μ f s \mu_f^{s} μ f s concentrated on an m m m -null set (see after Lemma 4.9 ), so that F = ρ f M F=\rho_fM F = ρ f M m m m -a.e. on B B B , dividing by M M M we get λ 2 ρ f + 4 λ H f + 4 Γ ( w ) 2 ≥ 0 \lambda^{2}\rho_f+4\lambda H_f+4\Gamma(w)^{2}\geq0 λ 2 ρ f + 4 λ H f + 4Γ ( w ) 2 ≥ 0 for all λ ∈ R \lambda\in\mathbb{R} λ ∈ R , m m m -a.e. on B B B . Since X X X is covered by countably many admissible balls,
∣ H [ f ] ( w , w ) ∣ ≤ ρ f Γ ( w ) m -a.e. in X . (10) |H[f](w,w)|\leq\sqrt{\rho_f}\,\Gamma(w)\qquad m\text{-a.e. in }X. \tag*{(10)} ∣ H [ f ] ( w , w ) ∣ ≤ ρ f Γ ( w ) m -a.e. in X . (10)
Step 4: polarisation. Let v , w ∈ Test ∞ ( X ) v,w\in\operatorname{Test}^{\infty}(X) v , w ∈ Test ∞ ( X ) . Applying (10) to s v ± s − 1 w sv\pm s^{-1}w s v ± s − 1 w for s ∈ Q ∖ { 0 } s\in\mathbb{Q}\setminus\{0\} s ∈ Q ∖ { 0 } , we get m m m -a.e.
∣ H [ f ] ( v , w ) ∣ = 1 4 ∣ H [ f ] ( s v + s − 1 w , s v + s − 1 w ) − H [ f ] ( s v − s − 1 w , s v − s − 1 w ) ∣ ≤ 1 4 ρ f ( Γ ( s v + s − 1 w ) + Γ ( s v − s − 1 w ) ) = 1 2 ρ f ( s 2 Γ ( v ) + s − 2 Γ ( w ) ) , \begin{aligned}
|H[f](v,w)|=\frac{1}{4}\left|H[f](sv+s^{-1}w,sv+s^{-1}w)-H[f](sv-s^{-1}w,sv-s^{-1}w)\right| \\
&\leq\frac{1}{4}\sqrt{\rho_f}\bigl(\Gamma(sv+s^{-1}w)+\Gamma(sv-s^{-1}w)\bigr)=\frac{1}{2}\sqrt{\rho_f}\bigl(s^{2}\Gamma(v)+s^{-2}\Gamma(w)\bigr),
\end{aligned} ∣ H [ f ] ( v , w ) ∣ = 4 1 H [ f ] ( s v + s − 1 w , s v + s − 1 w ) − H [ f ] ( s v − s − 1 w , s v − s − 1 w ) ≤ 4 1 ρ f ( Γ ( s v + s − 1 w ) + Γ ( s v − s − 1 w ) ) = 2 1 ρ f ( s 2 Γ ( v ) + s − 2 Γ ( w ) ) ,
and the infimum over s ∈ Q ∖ { 0 } s\in\mathbb{Q}\setminus\{0\} s ∈ Q ∖ { 0 } gives ∣ H [ f ] ( v , w ) ∣ ≤ ρ f ∣ D v ∣ ∣ D w ∣ |H[f](v,w)|\leq\sqrt{\rho_f}|Dv||Dw| ∣ H [ f ] ( v , w ) ∣ ≤ ρ f ∣ D v ∣∣ D w ∣ m m m -a.e. With v = f v=f v = f and (5) ,
∣ Γ ( Γ ( f ) , w ) ∣ ≤ 2 ρ f ∣ D f ∣ ∣ D w ∣ m -a.e., for every w ∈ Test ∞ ( X ) . (11) |\Gamma(\Gamma(f),w)|\leq2\sqrt{\rho_f}|Df||Dw|\qquad m\text{-a.e., for every }w\in\operatorname{Test}^{\infty}(X). \tag*{(11)} ∣Γ ( Γ ( f ) , w ) ∣ ≤ 2 ρ f ∣ D f ∣∣ D w ∣ m -a.e., for every w ∈ Test ∞ ( X ) . (11)
Step 5: approximation. By Lemma 2.5(v) and Lemma 4.2 , there are w k ∈ Test ∞ ( X ) w_k\in\operatorname{Test}^{\infty}(X) w k ∈ Test ∞ ( X ) with w k → Γ ( f ) w_k\to\Gamma(f) w k → Γ ( f ) in V \mathsf{V} V . Then Γ ( Γ ( f ) , w k ) → Γ ( Γ ( f ) ) \Gamma(\Gamma(f),w_k)\to\Gamma(\Gamma(f)) Γ ( Γ ( f ) , w k ) → Γ ( Γ ( f )) in L 1 ( m ) L^1(m) L 1 ( m ) and ∣ D w k ∣ → ∣ D Γ ( f ) ∣ |Dw_k|\to|D\Gamma(f)| ∣ D w k ∣ → ∣ D Γ ( f ) ∣ in L 2 ( m ) L^2(m) L 2 ( m ) , hence m m m -a.e. along a subsequence, and (11) gives Γ ( Γ ( f ) ) ≤ 2 ρ f ∣ D f ∣ Γ ( Γ ( f ) ) \Gamma(\Gamma(f))\leq2\sqrt{\rho_f}|Df|\sqrt{\Gamma(\Gamma(f))} Γ ( Γ ( f )) ≤ 2 ρ f ∣ D f ∣ Γ ( Γ ( f )) m m m -a.e., which is (8) .
Lemma 4.12. Let T > 0 T>0 T > 0 and let a , b : [ 0 , T ] → L 2 ( m ) a,b\colon[0,T]\to L^2(m) a , b : [ 0 , T ] → L 2 ( m ) satisfy
a ∈ C ( [ 0 , T ] ; L 2 ( m ) ) ∩ C 1 ( ( 0 , T ] ; L 2 ( m ) ) , b ∈ C ( [ 0 , T ] ; L 2 ( m ) ) ∩ C 1 ( [ 0 , T ] ; L 1 ( m ) ) , a\in C([0,T];L^2(m))\cap C^1((0,T];L^2(m)),\qquad b\in C([0,T];L^2(m))\cap C^1([0,T];L^1(m)), a ∈ C ([ 0 , T ] ; L 2 ( m )) ∩ C 1 (( 0 , T ] ; L 2 ( m )) , b ∈ C ([ 0 , T ] ; L 2 ( m )) ∩ C 1 ([ 0 , T ] ; L 1 ( m )) , and sup [ 0 , T ] ∥ a ∥ L ∞ ( m ) < ∞ \sup_{[0,T]}\|a\|_{L^\infty(m)}<\infty sup [ 0 , T ] ∥ a ∥ L ∞ ( m ) < ∞ . Then A ( s ) : = ∫ X a ( s ) b ( s ) d m A(s):=\int_Xa(s)b(s)\,\mathrm{d}m A ( s ) := ∫ X a ( s ) b ( s ) d m belongs to C ( [ 0 , T ] ) ∩ C 1 ( ( 0 , T ) ) C([0,T])\cap C^1((0,T)) C ([ 0 , T ]) ∩ C 1 (( 0 , T )) , and
A ′ ( s ) = ∫ X ( a ′ ( s ) b ( s ) + a ( s ) b ′ ( s ) ) d m for s ∈ ( 0 , T ) . A'(s)=\int_X\bigl(a'(s)b(s)+a(s)b'(s)\bigr)\,\mathrm{d}m\qquad\text{for }s\in(0,T). A ′ ( s ) = ∫ X ( a ′ ( s ) b ( s ) + a ( s ) b ′ ( s ) ) d m for s ∈ ( 0 , T ) .
Proof. Continuity follows from the continuity of the pairing on L 2 ( m ) × L 2 ( m ) L^2(\mathfrak{m}) \times L^2(\mathfrak{m}) L 2 ( m ) × L 2 ( m ) . For s ∈ ( 0 , T ) s \in(0,T) s ∈ ( 0 , T ) and small τ ≠ 0 \tau\ne0 τ = 0 ,
A ( s + τ ) − A ( s ) τ = ∫ X a ( s + τ ) − a ( s ) τ b ( s + τ ) d m + ∫ X a ( s ) b ( s + τ ) − b ( s ) τ d m ; \frac{A(s+\tau)-A(s)}{\tau}=\int_X\frac{a(s+\tau)-a(s)}{\tau}b(s+\tau)\,\mathrm{d}\mathfrak{m}+\int_Xa(s)\frac{b(s+\tau)-b(s)}{\tau}\,\mathrm{d}\mathfrak{m}; τ A ( s + τ ) − A ( s ) = ∫ X τ a ( s + τ ) − a ( s ) b ( s + τ ) d m + ∫ X a ( s ) τ b ( s + τ ) − b ( s ) d m ;
the first integral converges by convergence in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) of both factors, the second because a ( s ) ∈ L ∞ ( m ) a(s) \in L^\infty(\mathfrak{m}) a ( s ) ∈ L ∞ ( m ) and the difference quotient converges in L 1 ( m ) L^1(\mathfrak{m}) L 1 ( m ) . The continuity of A ′ A' A ′ follows in the same way. □
Proposition 4.13 (L 1 L^1 L 1 Bakry–Émery estimate). Assume (S). Then for every f ∈ V f \in\mathsf{V} f ∈ V and every t ≥ 0 t \ge0 t ≥ 0 ,
∣ D h t f ∣ ≤ e − K t h t ( ∣ D f ∣ ) m -a.e. in X . (12) |Dh_tf| \le e^{-Kt}h_t(|Df|) \quad\mathfrak{m}\text{-a.e. in }X. \tag*{(12)} ∣ D h t f ∣ ≤ e − K t h t ( ∣ D f ∣ ) m -a.e. in X . (12)
Proof. Step 1: setup. Let first f ∈ Test ∞ ( X ) f \in\operatorname{Test}^{\infty}(X) f ∈ Test ∞ ( X ) and t > 0 t>0 t > 0 , fix ε > 0 \varepsilon>0 ε > 0 and
ψ ( r ) : = r + ε − ε , ψ ′ ( r ) = 1 2 r + ε , ψ ′ ′ ( r ) = − 1 4 ( r + ε ) 3 / 2 , \psi(r):=\sqrt{r+\varepsilon}-\sqrt{\varepsilon},\qquad\psi'(r)=\frac{1}{2\sqrt{r+\varepsilon}},\qquad\psi''(r)=-\frac{1}{4(r+\varepsilon)^{3/2}}, ψ ( r ) := r + ε − ε , ψ ′ ( r ) = 2 r + ε 1 , ψ ′′ ( r ) = − 4 ( r + ε ) 3/2 1 ,
and fix φ ∈ L 1 ( m ) ∩ L ∞ ( m ) \varphi\in L^1(\mathfrak{m}) \cap L^\infty(\mathfrak{m}) φ ∈ L 1 ( m ) ∩ L ∞ ( m ) with φ ≥ 0 \varphi\ge0 φ ≥ 0 . Set a ( s ) : = h s φ a(s):=h_s\varphi a ( s ) := h s φ for s ∈ [ 0 , t ] s \in[0,t] s ∈ [ 0 , t ] . By Proposition 2.4(c), a ∈ C ( [ 0 , t ] ; L 2 ( m ) ) ∩ C 1 ( ( 0 , t ] ; L 2 ( m ) ) a \in C([0,t];L^2(\mathfrak{m})) \cap C^1((0,t];L^2(\mathfrak{m})) a ∈ C ([ 0 , t ] ; L 2 ( m )) ∩ C 1 (( 0 , t ] ; L 2 ( m )) with a ′ ( s ) = Δ a ( s ) a'(s)=\Delta a(s) a ′ ( s ) = Δ a ( s ) , 0 ≤ a ( s ) ≤ ∥ φ ∥ L ∞ ( m ) 0 \le a(s) \le\|\varphi\|_{L^\infty(\mathfrak{m})} 0 ≤ a ( s ) ≤ ∥ φ ∥ L ∞ ( m ) and ∫ X a ( s ) d m ≤ ∥ φ ∥ L 1 ( m ) \int_Xa(s)\,\mathrm{d}\mathfrak{m} \le\|\varphi\|_{L^1(\mathfrak{m})} ∫ X a ( s ) d m ≤ ∥ φ ∥ L 1 ( m ) . By Lemma 4.4(iii), the map s ↦ ψ ( Γ ( h t − s f ) ) s \mapsto\psi(\Gamma(h_{t-s}f)) s ↦ ψ ( Γ ( h t − s f )) belongs to C ( [ 0 , t ] ; L 2 ( m ) ) ∩ C 1 ( [ 0 , t ] ; L 1 ( m ) ) C([0,t];L^2(\mathfrak{m})) \cap C^1([0,t];L^1(\mathfrak{m})) C ([ 0 , t ] ; L 2 ( m )) ∩ C 1 ([ 0 , t ] ; L 1 ( m )) , with derivative
− 2 ψ ′ ( Γ ( h t − s f ) ) Γ ( h t − s f , Δ h t − s f ) . -2\psi'(\Gamma(h_{t-s}f))\Gamma(h_{t-s}f,\Delta h_{t-s}f). − 2 ψ ′ ( Γ ( h t − s f )) Γ ( h t − s f , Δ h t − s f ) .
Let
A ( s ) : = ∫ X a ( s ) ψ ( Γ ( h t − s f ) ) d m , s ∈ [ 0 , t ] . A(s):=\int_X a(s)\psi(\Gamma(h_{t-s}f))\,\mathrm{d}\mathfrak{m},\qquad s\in[0,t]. A ( s ) := ∫ X a ( s ) ψ ( Γ ( h t − s f )) d m , s ∈ [ 0 , t ] .
By Lemma 4.12, applied to a a a and s ↦ ψ ( Γ ( h t − s f ) ) s\mapsto\psi(\Gamma(h_{t-s}f)) s ↦ ψ ( Γ ( h t − s f )) , A ∈ C ( [ 0 , t ] ) ∩ C 1 ( ( 0 , t ) ) A\in C([0,t])\cap C^1((0,t)) A ∈ C ([ 0 , t ]) ∩ C 1 (( 0 , t )) and, writing g : = h t − s f ∈ Test ∞ ( X ) g:=h_{t-s}f\in\operatorname{Test}^{\infty}(X) g := h t − s f ∈ Test ∞ ( X ) (Lemma 4.3) and a : = a ( s ) a:=a(s) a := a ( s ) ,
A ′ ( s ) = ∫ X Δ a ψ ( Γ ( g ) ) d m − 2 ∫ X a ψ ′ ( Γ ( g ) ) Γ ( g , Δ g ) d m for s ∈ ( 0 , t ) . A'(s)=\int_X\Delta a\,\psi(\Gamma(g))\,\mathrm{d}\mathfrak{m}-2\int_Xa\psi'(\Gamma(g))\Gamma(g,\Delta g)\,\mathrm{d}\mathfrak{m}\qquad\text{for }s\in(0,t). A ′ ( s ) = ∫ X Δ a ψ ( Γ ( g )) d m − 2 ∫ X a ψ ′ ( Γ ( g )) Γ ( g , Δ g ) d m for s ∈ ( 0 , t ) .
Step 2: the derivative as a Γ 2 \Gamma_2 Γ 2 term. Let Ψ ∈ C 1 ( R ) \Psi\in C^1(\mathbb{R}) Ψ ∈ C 1 ( R ) be the extension of ψ ′ \psi' ψ ′ given by Ψ ( r ) : = ψ ′ ( 0 ) + ψ ′ ′ ( 0 ) r \Psi(r):=\psi'(0)+\psi''(0)r Ψ ( r ) := ψ ′ ( 0 ) + ψ ′′ ( 0 ) r for r < 0 r<0 r < 0 ; Ψ \Psi Ψ is Lipschitz and positive, and Ψ ≥ ψ ′ ( ∥ Γ ( g ) ∥ L ∞ ( m ) ) \Psi\ge\psi'(\|\Gamma(g)\|_{L^\infty(\mathfrak{m})}) Ψ ≥ ψ ′ ( ∥Γ ( g ) ∥ L ∞ ( m ) ) on ( − ∞ , ∥ Γ ( g ) ∥ L ∞ ( m ) ] (-\infty,\|\Gamma(g)\|_{L^\infty(\mathfrak{m})}] ( − ∞ , ∥Γ ( g ) ∥ L ∞ ( m ) ] . By the chain rule, ψ ( Γ ( g ) ) ∈ V \psi(\Gamma(g))\in\mathsf{V} ψ ( Γ ( g )) ∈ V and
χ : = Ψ ( Γ ( g ) ) a = Ψ ( 0 ) a + ( Ψ ( Γ ( g ) ) − Ψ ( 0 ) ) a ∈ V ∞ , ∇ χ = ψ ′ ( Γ ( g ) ) ∇ a + a ψ ′ ′ ( Γ ( g ) ) ∇ Γ ( g ) . \chi:=\Psi(\Gamma(g))a=\Psi(0)a+(\Psi(\Gamma(g))-\Psi(0))a\in\mathsf{V}_\infty,\qquad\nabla\chi=\psi'(\Gamma(g))\nabla a+a\psi''(\Gamma(g))\nabla\Gamma(g). χ := Ψ ( Γ ( g )) a = Ψ ( 0 ) a + ( Ψ ( Γ ( g )) − Ψ ( 0 )) a ∈ V ∞ , ∇ χ = ψ ′ ( Γ ( g )) ∇ a + a ψ ′′ ( Γ ( g )) ∇Γ ( g ) .
Hence, by Proposition 2.4(b),
∫ X Δ a ψ ( Γ ( g ) ) d m = − ∫ X ψ ′ ( Γ ( g ) ) ∇ a ⋅ ∇ Γ ( g ) d m = − E ( Γ ( g ) , χ ) + ∫ X a ψ ′ ′ ( Γ ( g ) ) Γ ( Γ ( g ) ) d m , \int_X\Delta a\,\psi(\Gamma(g))\,\mathrm{d}\mathfrak{m}=-\int_X\psi'(\Gamma(g))\nabla a\cdot\nabla\Gamma(g)\,\mathrm{d}\mathfrak{m}=-\mathcal{E}(\Gamma(g),\chi)+\int_Xa\psi''(\Gamma(g))\Gamma(\Gamma(g))\,\mathrm{d}\mathfrak{m}, ∫ X Δ a ψ ( Γ ( g )) d m = − ∫ X ψ ′ ( Γ ( g )) ∇ a ⋅ ∇Γ ( g ) d m = − E ( Γ ( g ) , χ ) + ∫ X a ψ ′′ ( Γ ( g )) Γ ( Γ ( g )) d m ,
and therefore
A ′ ( s ) = 2 Γ 2 [ g ] ( χ ) + ∫ X a ψ ′ ′ ( Γ ( g ) ) Γ ( Γ ( g ) ) d m . (13) A'(s)=2\Gamma_2[g](\chi)+\int_Xa\psi''(\Gamma(g))\Gamma(\Gamma(g))\,\mathrm{d}\mathfrak{m}. \tag*{(13)} A ′ ( s ) = 2 Γ 2 [ g ] ( χ ) + ∫ X a ψ ′′ ( Γ ( g )) Γ ( Γ ( g )) d m . (13)
Step 3: lower bound. By Lemma 2.5(ii), a a a has a Lipschitz representative a ^ \hat a a ^ , which is nonnegative everywhere because supp ( m ) = X \operatorname{supp}(\mathfrak{m})=X supp ( m ) = X . By Remark 3.9(ii)–(iii) applied to the decomposition of χ \chi χ above, χ ~ = Ψ ( Γ ( g ) ) a ^ ≥ 0 \widetilde{\chi}=\Psi(\Gamma(g))\hat a\ge0 χ = Ψ ( Γ ( g )) a ^ ≥ 0 q.e., hence μ g \mu_g μ g -a.e. By Lemma 4.9 and μ g ≥ ρ g m \mu_g\ge\rho_g\mathfrak{m} μ g ≥ ρ g m ,
Γ 2 [ g ] ( χ ) = ∫ X χ ~ d μ g + K ∫ X Γ ( g ) χ d m ≥ ∫ X ρ g χ d m + K ∫ X Γ ( g ) χ d m ; \Gamma_2[g](\chi)=\int_X\widetilde{\chi}\,\mathrm{d}\mu_g+K\int_X\Gamma(g)\chi\,\mathrm{d}\mathfrak{m}\ge\int_X\rho_g\chi\,\mathrm{d}\mathfrak{m}+K\int_X\Gamma(g)\chi\,\mathrm{d}\mathfrak{m}; Γ 2 [ g ] ( χ ) = ∫ X χ d μ g + K ∫ X Γ ( g ) χ d m ≥ ∫ X ρ g χ d m + K ∫ X Γ ( g ) χ d m ;
in particular a ρ g ∈ L 1 ( m ) a\rho_g\in L^1(\mathfrak{m}) a ρ g ∈ L 1 ( m ) , since Ψ ( Γ ( g ) ) ≥ ψ ′ ( ∥ Γ ( g ) ∥ L ∞ ( m ) ) > 0 \Psi(\Gamma(g))\ge\psi'(\|\Gamma(g)\|_{L^\infty(\mathfrak{m})})>0 Ψ ( Γ ( g )) ≥ ψ ′ ( ∥Γ ( g ) ∥ L ∞ ( m ) ) > 0 . Since Γ ( g ) ≥ 0 \Gamma(g)\ge0 Γ ( g ) ≥ 0 and Ψ = ψ ′ \Psi=\psi' Ψ = ψ ′ on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , we have χ = ψ ′ ( Γ ( g ) ) a \chi=\psi'(\Gamma(g))a χ = ψ ′ ( Γ ( g )) a , and inserting the last inequality into (13) gives
A ′ ( s ) ≥ ∫ X a ( 2 ρ g ψ ′ ( Γ ( g ) ) + 2 K Γ ( g ) ψ ′ ( Γ ( g ) ) + ψ ′ ′ ( Γ ( g ) ) Γ ( Γ ( g ) ) ) d m . A'(s)\ge\int_Xa\left(2\rho_g\psi'(\Gamma(g))+2K\Gamma(g)\psi'(\Gamma(g))+\psi''(\Gamma(g))\Gamma(\Gamma(g))\right)\,\mathrm{d}\mathfrak{m}. A ′ ( s ) ≥ ∫ X a ( 2 ρ g ψ ′ ( Γ ( g )) + 2 K Γ ( g ) ψ ′ ( Γ ( g )) + ψ ′′ ( Γ ( g )) Γ ( Γ ( g )) ) d m .
By Proposition 4.11 applied to g ∈ Test ∞ ( X ) g\in\operatorname{Test}^{\infty}(X) g ∈ Test ∞ ( X ) , Γ ( Γ ( g ) ) ≤ 4 ρ g Γ ( g ) \Gamma(\Gamma(g))\le4\rho_g\Gamma(g) Γ ( Γ ( g )) ≤ 4 ρ g Γ ( g ) m \mathfrak{m} m -a.e., and since ψ ′ ′ < 0 \psi''<0 ψ ′′ < 0 ,
ψ ′ ′ ( Γ ( g ) ) Γ ( Γ ( g ) ) ≥ 4 ρ g Γ ( g ) ψ ′ ′ ( Γ ( g ) ) m -a.e. \psi''(\Gamma(g))\Gamma(\Gamma(g))\ge4\rho_g\Gamma(g)\psi''(\Gamma(g))\qquad\mathfrak{m}\text{-a.e.} ψ ′′ ( Γ ( g )) Γ ( Γ ( g )) ≥ 4 ρ g Γ ( g ) ψ ′′ ( Γ ( g )) m -a.e.
Hence
A ′ ( s ) ≥ ∫ X a ( 2 ρ g ( ψ ′ ( Γ ( g ) ) + 2 Γ ( g ) ψ ′ ′ ( Γ ( g ) ) ) + 2 K Γ ( g ) ψ ′ ( Γ ( g ) ) ) d m . A'(s) \ge\int_X a\left(2\rho_g\left(\psi'(\Gamma(g))+2\Gamma(g)\psi''(\Gamma(g))\right)+2K\Gamma(g)\psi'(\Gamma(g))\right)\mathrm{d}m. A ′ ( s ) ≥ ∫ X a ( 2 ρ g ( ψ ′ ( Γ ( g )) + 2Γ ( g ) ψ ′′ ( Γ ( g )) ) + 2 K Γ ( g ) ψ ′ ( Γ ( g )) ) d m .
For r ≥ 0 r \ge0 r ≥ 0 we have
ψ ′ ( r ) + 2 r ψ ′ ′ ( r ) = ε 2 ( r + ε ) 3 / 2 ≥ 0 , 2 r ψ ′ ( r ) = r r + ε . \psi'(r)+2r\psi''(r)=\frac{\varepsilon}{2(r+\varepsilon)^{3/2}}\ge0,\qquad2r\psi'(r)=\frac{r}{\sqrt{r+\varepsilon}}. ψ ′ ( r ) + 2 r ψ ′′ ( r ) = 2 ( r + ε ) 3/2 ε ≥ 0 , 2 r ψ ′ ( r ) = r + ε r .
Since a ≥ 0 a\ge0 a ≥ 0 and ρ g ≥ 0 \rho_g\ge0 ρ g ≥ 0 , the term containing ρ g \rho_g ρ g is nonnegative, and
A ′ ( s ) ≥ K ∫ X a Γ ( g ) Γ ( g ) + ε d m . A'(s)\ge K\int_X a\frac{\Gamma(g)}{\sqrt{\Gamma(g)}+\varepsilon}\,\mathrm{d}m. A ′ ( s ) ≥ K ∫ X a Γ ( g ) + ε Γ ( g ) d m .
Since, for r ≥ 0 r\ge0 r ≥ 0 ,
0 ≤ r r + ε − ψ ( r ) = ε − ε r + ε ≤ ε , 0\le\frac{r}{\sqrt{r+\varepsilon}}-\psi(r)=\sqrt{\varepsilon}-\frac{\varepsilon}{\sqrt{r+\varepsilon}}\le\sqrt{\varepsilon}, 0 ≤ r + ε r − ψ ( r ) = ε − r + ε ε ≤ ε ,
and ∫ X a d m ≤ ∥ φ ∥ L 1 ( m ) \int_X a\,\mathrm{d}m\le\|\varphi\|_{L^1(\mathfrak m)} ∫ X a d m ≤ ∥ φ ∥ L 1 ( m ) , the right-hand side differs from K A ( s ) = K ∫ X a ψ ( Γ ( g ) ) d m KA(s)=K\int_Xa\psi(\Gamma(g))\mathrm{d}m K A ( s ) = K ∫ X a ψ ( Γ ( g )) d m by at most ∣ K ∣ ε ∥ φ ∥ L 1 ( m ) |K|\sqrt{\varepsilon}\|\varphi\|_{L^1(\mathfrak m)} ∣ K ∣ ε ∥ φ ∥ L 1 ( m ) . Hence
A ′ ( s ) ≥ K A ( s ) − ∣ K ∣ ε ∥ φ ∥ L 1 ( m ) for every s ∈ ( 0 , t ) . A'(s)\ge KA(s)-|K|\sqrt{\varepsilon}\|\varphi\|_{L^1(\mathfrak m)}
\qquad\text{for every }s\in(0,t). A ′ ( s ) ≥ K A ( s ) − ∣ K ∣ ε ∥ φ ∥ L 1 ( m ) for every s ∈ ( 0 , t ) .
Step 4: integration. Since A A A is continuous on [ 0 , t ] [0,t] [ 0 , t ] and of class C 1 C^1 C 1 on ( 0 , t ) (0,t) ( 0 , t ) , Gronwall’s lemma gives
A ( 0 ) ≤ e − K t A ( t ) + C K , t ε ∥ φ ∥ L 1 ( m ) , C K , t : = ∣ K ∣ ∫ 0 t e − K s d s . A(0)\le e^{-Kt}A(t)+C_{K,t}\sqrt{\varepsilon}\|\varphi\|_{L^1(\mathfrak m)},\qquad C_{K,t}:=|K|\int_0^t e^{-Ks}\,\mathrm{d}s. A ( 0 ) ≤ e − K t A ( t ) + C K , t ε ∥ φ ∥ L 1 ( m ) , C K , t := ∣ K ∣ ∫ 0 t e − K s d s .
where A ( 0 ) = ∫ X φ ψ ( Γ ( h t f ) ) d m A(0)=\int_X\varphi\psi(\Gamma(h_tf))\mathrm{d}m A ( 0 ) = ∫ X φ ψ ( Γ ( h t f )) d m and A ( t ) = ∫ X h t φ ψ ( Γ ( f ) ) d m A(t)=\int_Xh_t\varphi\psi(\Gamma(f))\mathrm{d}m A ( t ) = ∫ X h t φ ψ ( Γ ( f )) d m . Letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 by monotone convergence, since ψ ( r ) = r ( r + ε + ε ) − 1 \psi(r)=r(\sqrt{r+\varepsilon}+\sqrt{\varepsilon})^{-1} ψ ( r ) = r ( r + ε + ε ) − 1 increases to r \sqrt r r as ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , and using the self-adjointness of h t h_t h t ,
∫ X φ ∣ D h t f ∣ d m ≤ e − K t ∫ X h t φ ∣ D f ∣ d m = e − K t ∫ X φ h t ( ∣ D f ∣ ) d m . \int_X\varphi|Dh_tf|\,\mathrm{d}m\le e^{-Kt}\int_Xh_t\varphi|Df|\,\mathrm{d}m=e^{-Kt}\int_X\varphi h_t(|Df|)\,\mathrm{d}m. ∫ X φ ∣ D h t f ∣ d m ≤ e − K t ∫ X h t φ ∣ D f ∣ d m = e − K t ∫ X φ h t ( ∣ D f ∣ ) d m .
Choosing φ \varphi φ as the indicator of { ∣ D h t f ∣ > e − K t h t ( ∣ D f ∣ ) } ∩ E n \{|Dh_tf|>e^{-Kt}h_t(|Df|)\}\cap E_n { ∣ D h t f ∣ > e − K t h t ( ∣ D f ∣ )} ∩ E n , where ( E n ) n ∈ N (E_n)_{n\in\mathbb N} ( E n ) n ∈ N are Borel sets of finite measure with E n ↑ X E_n\uparrow X E n ↑ X (Lemma 2.1 ), and letting n → ∞ n\to\infty n → ∞ gives (12) for f ∈ Test ∞ ( X ) f\in\operatorname{Test}^{\infty}(X) f ∈ Test ∞ ( X ) .
Step 5: density. For f ∈ V f\in\mathbb V f ∈ V , take f k ∈ Test ∞ ( X ) f_k\in\operatorname{Test}^{\infty}(X) f k ∈ Test ∞ ( X ) with f k → f f_k\to f f k → f in V \mathbb V V (Lemma 2.5(iv)). Then ∣ D h t f k ∣ → ∣ D h t f ∣ |Dh_tf_k|\to|Dh_tf| ∣ D h t f k ∣ → ∣ D h t f ∣ and h t ( ∣ D f k ∣ ) → h t ( ∣ D f ∣ ) h_t(|Df_k|)\to h_t(|Df|) h t ( ∣ D f k ∣ ) → h t ( ∣ D f ∣ ) in L 2 ( m ) L^2(\mathfrak m) L 2 ( m ) , and (12) passes to the limit m \mathfrak m m -a.e. along a subsequence. The case t = 0 t=0 t = 0 is trivial. □
The Harnack inequality
Lemma 4.14 (Power-type Harnack inequality). Assume (S). Let f ∈ LIP f m ( X ) f\in\operatorname{LIP}_{\mathrm{fm}}(X) f ∈ LIP fm ( X ) with 0 ≤ f ≤ 1 / 2 0\le f\le1/2 0 ≤ f ≤ 1/2 and let t > 0 t>0 t > 0 . Then h t f h_tf h t f has a Lipschitz representative, still denoted by h t f h_tf h t f , and
( h t f ( x ) ) 2 ≤ h t f ( y ) exp ( d ( x , y ) 2 2 I K ( t ) ) for all x , y ∈ X . (14) (h_tf(x))^2\le h_tf(y)\exp\left(\frac{\mathrm d(x,y)^2}{2I_K(t)}\right)\qquad\text{for all }x,y\in X. \tag*{(14)} ( h t f ( x ) ) 2 ≤ h t f ( y ) exp ( 2 I K ( t ) d ( x , y ) 2 ) for all x , y ∈ X . (14)
where I K ( t ) I_K(t) I K ( t ) is given by (2) .
Proof. For r ∈ ( 0 , t ] r\in(0,t] r ∈ ( 0 , t ] set f r : = h r f f_r:=h_rf f r := h r f . Since f ∈ LIP f m ( X ) f\in\operatorname{LIP}_{\mathrm{fm}}(X) f ∈ LIP fm ( X ) , we have f ∈ L 1 ( m ) ∩ L ∞ ( m ) ∩ V f\in L^1(\mathfrak m)\cap L^\infty(\mathfrak m)\cap\mathbb V f ∈ L 1 ( m ) ∩ L ∞ ( m ) ∩ V by Lemma 2.3 . By positivity and the weak maximum principle, 0 ≤ f r ≤ 1 / 2 0\le f_r\le1/2 0 ≤ f r ≤ 1/2 . By Lemma 2.5(ii), f r f_r f r admits a Lipschitz representative; moreover f r ∈ D ( Δ ) ∩ L 1 ( m ) ∩ L ∞ ( m ) f_r\in D(\Delta)\cap L^1(\mathfrak m)\cap L^\infty(\mathfrak m) f r ∈ D ( Δ ) ∩ L 1 ( m ) ∩ L ∞ ( m ) . We always use these representatives; since supp ( m ) = X \operatorname{supp}(\mathfrak m)=X supp ( m ) = X , two continuous functions that agree m \mathfrak m m -a.e. agree everywhere.
Fix δ ∈ ( 0 , 1 / 4 ) \delta\in(0,1/4) δ ∈ ( 0 , 1/4 ) and consider the smooth convex function ϕ δ : [ 0 , 1 / 2 ] → [ 0 , ∞ ) \phi_\delta:[0,1/2]\to[0,\infty) ϕ δ : [ 0 , 1/2 ] → [ 0 , ∞ ) ,
ϕ δ ( r ) : = ( r + δ ) log ( r + δ ) − r log δ − δ log δ . \phi_\delta(r):=(r+\delta)\log(r+\delta)-r\log\delta-\delta\log\delta. ϕ δ ( r ) := ( r + δ ) log ( r + δ ) − r log δ − δ log δ .
Since ϕ δ ( 0 ) = 0 \phi_\delta(0)=0 ϕ δ ( 0 ) = 0 and ϕ δ ′ ′ ( r ) = ( r + δ ) − 1 ∈ [ 1 , δ − 1 ] \phi_\delta''(r)=(r+\delta)^{-1}\in[1,\delta^{-1}] ϕ δ ′′ ( r ) = ( r + δ ) − 1 ∈ [ 1 , δ − 1 ] on [ 0 , 1 / 2 ] [0,1/2] [ 0 , 1/2 ] , by Proposition 2.4 (b), applied to a C 2 C^2 C 2 extension of ϕ δ \phi_\delta ϕ δ to R \mathbb R R with bounded second derivative, we have ϕ δ ( f t − s ) ∈ D ( Δ ) ∩ L 2 ( m ) \phi_\delta(f_{t-s})\in D(\Delta)\cap L^2(\mathfrak m) ϕ δ ( f t − s ) ∈ D ( Δ ) ∩ L 2 ( m ) for every s ∈ [ 0 , t ) s\in[0,t) s ∈ [ 0 , t ) , with
Δ ( ϕ δ ( f t − s ) ) = ϕ δ ′ ( f t − s ) Δ f t − s + ∣ D f t − s ∣ 2 f t − s + δ . (15) \Delta(\phi_\delta(f_{t-s}))=\phi_\delta'(f_{t-s})\Delta f_{t-s}+\frac{|Df_{t-s}|^2}{f_{t-s}+\delta}. \tag*{(15)} Δ ( ϕ δ ( f t − s )) = ϕ δ ′ ( f t − s ) Δ f t − s + f t − s + δ ∣ D f t − s ∣ 2 . (15)
Fix g ∈ L 2 ( m ) ∩ L ∞ ( m ) g \in L^{2}(\mathfrak{m}) \cap L^{\infty}(\mathfrak{m}) g ∈ L 2 ( m ) ∩ L ∞ ( m ) with g ≥ 0 g \ge0 g ≥ 0 , and set Φ δ ( s ) : = ∫ X h s g ϕ δ ( f t − s ) d m \Phi_{\delta}(s) := \int_X h_sg\phi_{\delta}(f_{t-s})\mathrm{d}\mathfrak{m} Φ δ ( s ) := ∫ X h s g ϕ δ ( f t − s ) d m for s ∈ [ 0 , t ] s \in[0,t] s ∈ [ 0 , t ] . The curve s ↦ h s g s \mapsto h_sg s ↦ h s g belongs to C ( [ 0 , t ] ; L 2 ( m ) ) ∩ C 1 ( ( 0 , t ] ; L 2 ( m ) ) C([0,t];L^{2}(\mathfrak{m})) \cap C^{1}((0,t];L^{2}(\mathfrak{m})) C ([ 0 , t ] ; L 2 ( m )) ∩ C 1 (( 0 , t ] ; L 2 ( m )) with derivative Δ h s g \Delta h_sg Δ h s g and h s g ∈ D ( Δ ) h_sg \in D(\Delta) h s g ∈ D ( Δ ) for s > 0 s>0 s > 0 . Since ϕ δ \phi_{\delta} ϕ δ is of class C 2 C^{2} C 2 on [ 0 , 1 / 2 ] [0,1/2] [ 0 , 1/2 ] with ϕ δ ( 0 ) = 0 \phi_{\delta}(0)=0 ϕ δ ( 0 ) = 0 , the curve s ↦ ϕ δ ( f t − s ) s \mapsto\phi_{\delta}(f_{t-s}) s ↦ ϕ δ ( f t − s ) belongs to C ( [ 0 , t ] ; L 2 ( m ) ) ∩ C 1 ( [ 0 , t ) ; L 2 ( m ) ) C([0,t];L^{2}(\mathfrak{m})) \cap C^{1}([0,t);L^{2}(\mathfrak{m})) C ([ 0 , t ] ; L 2 ( m )) ∩ C 1 ([ 0 , t ) ; L 2 ( m )) with derivative − ϕ δ ′ ( f t − s ) Δ f t − s -\phi_{\delta}'(f_{t-s})\Delta f_{t-s} − ϕ δ ′ ( f t − s ) Δ f t − s . Hence Φ δ ∈ C ( [ 0 , t ] ) ∩ C 1 ( ( 0 , t ) ) \Phi_{\delta} \in C([0,t]) \cap C^{1}((0,t)) Φ δ ∈ C ([ 0 , t ]) ∩ C 1 (( 0 , t )) , and differentiating on ( 0 , t ) (0,t) ( 0 , t ) , using (15) and the symmetry of Δ \Delta Δ on D ( Δ ) D(\Delta) D ( Δ ) , we get
Φ δ ′ ( s ) = ∫ X h s g ∣ D f t − s ∣ 2 f t − s + δ d m = ∫ X g h s ( ∣ D f t − s ∣ 2 f t − s + δ ) d m . (16) \Phi_{\delta}'(s)=\int_X h_sg\frac{|Df_{t-s}|^2}{f_{t-s}+\delta}\mathrm{d}\mathfrak{m}=\int_X gh_s\left(\frac{|Df_{t-s}|^2}{f_{t-s}+\delta}\right)\mathrm{d}\mathfrak{m}. \tag*{(16)} Φ δ ′ ( s ) = ∫ X h s g f t − s + δ ∣ D f t − s ∣ 2 d m = ∫ X g h s ( f t − s + δ ∣ D f t − s ∣ 2 ) d m . (16)
For λ ∈ R \lambda\in\mathbb{R} λ ∈ R , expanding ( ( f t − s + δ ) − 1 / 2 ∣ D f t − s ∣ + λ ( f t − s + δ ) 1 / 2 ) 2 ≥ 0 \left((f_{t-s}+\delta)^{-1/2}|Df_{t-s}|+\lambda(f_{t-s}+\delta)^{1/2}\right)^2 \ge0 ( ( f t − s + δ ) − 1/2 ∣ D f t − s ∣ + λ ( f t − s + δ ) 1/2 ) 2 ≥ 0 and applying the positivity-preserving operator h s h_s h s , with h s f t − s = f t h_sf_{t-s}=f_t h s f t − s = f t and h s δ ≤ δ h_s\delta\le\delta h s δ ≤ δ , we obtain
h s ( ∣ D f t − s ∣ 2 f t − s + δ ) + 2 λ h s ( ∣ D f t − s ∣ ) + λ 2 ( f t + δ ) ≥ 0 m -a.e. h_s\left(\frac{|Df_{t-s}|^2}{f_{t-s}+\delta}\right)+2\lambda h_s(|Df_{t-s}|)+\lambda^2(f_t+\delta)\ge0\quad\mathfrak{m}\text{-a.e.} h s ( f t − s + δ ∣ D f t − s ∣ 2 ) + 2 λ h s ( ∣ D f t − s ∣ ) + λ 2 ( f t + δ ) ≥ 0 m -a.e.
For each λ \lambda λ this holds outside an m \mathfrak{m} m -null set N λ N_\lambda N λ . The set N : = ⋃ λ ∈ Q N λ N:=\bigcup_{\lambda\in\mathbb{Q}}N_\lambda N := ⋃ λ ∈ Q N λ is m \mathfrak{m} m -null, and by continuity in λ \lambda λ the inequality holds at every point of X ∖ N X\setminus N X ∖ N for every λ ∈ R \lambda\in\mathbb{R} λ ∈ R . At such points, the choice λ = − h s ( ∣ D f t − s ∣ ) / ( f t + δ ) \lambda=-h_s(|Df_{t-s}|)/(f_t+\delta) λ = − h s ( ∣ D f t − s ∣ ) / ( f t + δ ) gives
( h s ( ∣ D f t − s ∣ ) ) 2 ≤ ( f t + δ ) h s ( ∣ D f t − s ∣ 2 f t − s + δ ) m -a.e. (17) \left(h_s(|Df_{t-s}|)\right)^2\le(f_t+\delta)h_s\left(\frac{|Df_{t-s}|^2}{f_{t-s}+\delta}\right)\quad\mathfrak{m}\text{-a.e.} \tag*{(17)} ( h s ( ∣ D f t − s ∣ ) ) 2 ≤ ( f t + δ ) h s ( f t − s + δ ∣ D f t − s ∣ 2 ) m -a.e. (17)
Then, applying Proposition 4.13 at time s s s to f t − s ∈ V f_{t-s}\in\mathbb{V} f t − s ∈ V and using (17) , we get
∣ D f t ∣ 2 ≤ e − 2 K s ( h s ( ∣ D f t − s ∣ ) ) 2 ≤ e − 2 K s ( f t + δ ) h s ( ∣ D f t − s ∣ 2 f t − s + δ ) m -a.e. (18) |Df_t|^2\le e^{-2Ks}\left(h_s(|Df_{t-s}|)\right)^2\le e^{-2Ks}(f_t+\delta)h_s\left(\frac{|Df_{t-s}|^2}{f_{t-s}+\delta}\right)\quad\mathfrak{m}\text{-a.e.} \tag*{(18)} ∣ D f t ∣ 2 ≤ e − 2 K s ( h s ( ∣ D f t − s ∣ ) ) 2 ≤ e − 2 K s ( f t + δ ) h s ( f t − s + δ ∣ D f t − s ∣ 2 ) m -a.e. (18)
Dividing (18) by e − 2 K s ( f t + δ ) e^{-2Ks}(f_t+\delta) e − 2 K s ( f t + δ ) and inserting it into (16) , we obtain
Φ δ ′ ( s ) ≥ e 2 K s ∫ X g ∣ D f t ∣ 2 f t + δ d m for every s ∈ ( 0 , t ) . \Phi_{\delta}'(s)\ge e^{2Ks}\int_Xg\frac{|Df_t|^2}{f_t+\delta}\mathrm{d}\mathfrak{m}\quad\text{for every }s\in(0,t). Φ δ ′ ( s ) ≥ e 2 K s ∫ X g f t + δ ∣ D f t ∣ 2 d m for every s ∈ ( 0 , t ) .
Since Φ δ ∈ C ( [ 0 , t ] ) \Phi_{\delta}\in C([0,t]) Φ δ ∈ C ([ 0 , t ]) , integrating over s ∈ ( 0 , t ) s\in(0,t) s ∈ ( 0 , t ) and writing Φ δ ( t ) = ∫ X g h t ( ϕ δ ( f ) ) d m \Phi_{\delta}(t)=\int_Xg\,h_t(\phi_{\delta}(f))\mathrm{d}\mathfrak{m} Φ δ ( t ) = ∫ X g h t ( ϕ δ ( f )) d m by the self-adjointness of h t h_t h t , we get
I K ( t ) ∫ X g ∣ D f t ∣ 2 f t + δ d m ≤ Φ δ ( t ) − Φ δ ( 0 ) = ∫ X g ( h t ( ϕ δ ( f ) ) − ϕ δ ( f t ) ) d m . I_K(t)\int_Xg\frac{|Df_t|^2}{f_t+\delta}\mathrm{d}\mathfrak{m}\le\Phi_{\delta}(t)-\Phi_{\delta}(0)=\int_Xg\left(h_t(\phi_{\delta}(f))-\phi_{\delta}(f_t)\right)\mathrm{d}\mathfrak{m}. I K ( t ) ∫ X g f t + δ ∣ D f t ∣ 2 d m ≤ Φ δ ( t ) − Φ δ ( 0 ) = ∫ X g ( h t ( ϕ δ ( f )) − ϕ δ ( f t ) ) d m .
Both ∣ D f t ∣ 2 / ( f t + δ ) ≤ δ − 1 ∣ D f t ∣ 2 |Df_t|^2/(f_t+\delta)\le\delta^{-1}|Df_t|^2 ∣ D f t ∣ 2 / ( f t + δ ) ≤ δ − 1 ∣ D f t ∣ 2 and h t ( ϕ δ ( f ) ) − ϕ δ ( f t ) h_t(\phi_{\delta}(f))-\phi_{\delta}(f_t) h t ( ϕ δ ( f )) − ϕ δ ( f t ) belong to L 1 ( m ) L^1(\mathfrak{m}) L 1 ( m ) , the latter because ϕ δ \phi_{\delta} ϕ δ is Lipschitz on [ 0 , 1 / 2 ] [0,1/2] [ 0 , 1/2 ] with ϕ δ ( 0 ) = 0 \phi_{\delta}(0)=0 ϕ δ ( 0 ) = 0 and f , f t ∈ L 1 ( m ) f,f_t\in L^1(\mathfrak{m}) f , f t ∈ L 1 ( m ) . Since g ∈ L 2 ( m ) ∩ L ∞ ( m ) g\in L^2(\mathfrak{m})\cap L^\infty(\mathfrak{m}) g ∈ L 2 ( m ) ∩ L ∞ ( m ) with g ≥ 0 g\ge0 g ≥ 0 is arbitrary, we obtain
I K ( t ) ∣ D f t ∣ 2 f t + δ ≤ h t ( ϕ δ ( f ) ) − ϕ δ ( f t ) m -a.e. (19) I_K(t)\frac{|Df_t|^2}{f_t+\delta}\le h_t(\phi_{\delta}(f))-\phi_{\delta}(f_t)\quad\mathfrak{m}\text{-a.e.} \tag*{(19)} I K ( t ) f t + δ ∣ D f t ∣ 2 ≤ h t ( ϕ δ ( f )) − ϕ δ ( f t ) m -a.e. (19)
We now bound the right-hand side of (19) from above. The linear part of ϕ δ \phi_{\delta} ϕ δ gives no contribution: writing
ϕ δ ( r ) = θ δ ( r ) − r log δ , θ δ ( r ) : = ( r + δ ) log ( r + δ ) − δ log δ , \phi_{\delta}(r)=\theta_{\delta}(r)-r\log\delta,\qquad\theta_{\delta}(r):=(r+\delta)\log(r+\delta)-\delta\log\delta, ϕ δ ( r ) = θ δ ( r ) − r log δ , θ δ ( r ) := ( r + δ ) log ( r + δ ) − δ log δ ,
the linearity of h t h_t h t and h t f = f t h_tf=f_t h t f = f t give
h t ( ϕ δ ( f ) ) − ϕ δ ( f t ) = h t ( θ δ ( f ) ) − θ δ ( f t ) . h_t(\phi_{\delta}(f))-\phi_{\delta}(f_t)=h_t(\theta_{\delta}(f))-\theta_{\delta}(f_t). h t ( ϕ δ ( f )) − ϕ δ ( f t ) = h t ( θ δ ( f )) − θ δ ( f t ) .
Moreover 0 < f + δ ≤ 3 / 4 < 1 0<f+\delta\le3/4<1 0 < f + δ ≤ 3/4 < 1 , so ( f + δ ) log ( f + δ ) ≤ 0 (f+\delta)\log(f+\delta)\le0 ( f + δ ) log ( f + δ ) ≤ 0 and hence θ δ ( f ) ≤ θ δ ( f ) + ≤ − δ log δ \theta_{\delta}(f)\le\theta_{\delta}(f)^+\le-\delta\log\delta θ δ ( f ) ≤ θ δ ( f ) + ≤ − δ log δ m \mathfrak{m} m -a.e. Since θ δ \theta_{\delta} θ δ is Lipschitz on [ 0 , 1 / 2 ] [0,1/2] [ 0 , 1/2 ] with θ δ ( 0 ) = 0 \theta_{\delta}(0)=0 θ δ ( 0 ) = 0 , we have θ δ ( f ) + ∈ L 2 ( m ) \theta_{\delta}(f)^+\in L^2(\mathfrak{m}) θ δ ( f ) + ∈ L 2 ( m ) , and positivity together with the weak maximum principle gives
h t ( θ δ ( f ) ) ≤ h t ( θ δ ( f ) + ) ≤ − δ log δ . h_t(\theta_{\delta}(f))\le h_t(\theta_{\delta}(f)^+)\le-\delta\log\delta. h t ( θ δ ( f )) ≤ h t ( θ δ ( f ) + ) ≤ − δ log δ .
Therefore
h t ( ϕ δ ( f ) ) − ϕ δ ( f t ) ≤ − δ log δ − θ δ ( f t ) = ( f t + δ ) log 1 f t + δ , h_t(\phi_{\delta}(f))-\phi_{\delta}(f_t)\le-\delta\log\delta-\theta_{\delta}(f_t)=(f_t+\delta)\log\frac{1}{f_t+\delta}, h t ( ϕ δ ( f )) − ϕ δ ( f t ) ≤ − δ log δ − θ δ ( f t ) = ( f t + δ ) log f t + δ 1 ,
and (19) becomes
I K ( t ) ∣ D f t ∣ 2 ≤ ( f t + δ ) 2 log 1 f t + δ m-a.e. (20) I_K(t)|D f_t|^2 \le(f_t+\delta)^2\log\frac{1}{f_t+\delta}\quad\text{m-a.e.} \tag*{(20)} I K ( t ) ∣ D f t ∣ 2 ≤ ( f t + δ ) 2 log f t + δ 1 m-a.e. (20)
Consider η δ ( r ) : = − log δ − − log ( r + δ ) \eta_\delta(r):=\sqrt{-\log\delta}-\sqrt{-\log(r+\delta)} η δ ( r ) := − log δ − − log ( r + δ ) for r ∈ [ 0 , 1 / 2 ] r\in[0,1/2] r ∈ [ 0 , 1/2 ] . Since r + δ ∈ [ δ , 3 / 4 ] r+\delta\in[\delta,3/4] r + δ ∈ [ δ , 3/4 ] , we have − log ( r + δ ) ≥ log ( 4 / 3 ) > 0 -\log(r+\delta)\ge\log(4/3)>0 − log ( r + δ ) ≥ log ( 4/3 ) > 0 , so η δ \eta_\delta η δ is smooth on [ 0 , 1 / 2 ] [0,1/2] [ 0 , 1/2 ] with η δ ( 0 ) = 0 \eta_\delta(0)=0 η δ ( 0 ) = 0 and
0 < η δ ′ ( r ) = 1 2 ( r + δ ) − log ( r + δ ) ≤ 1 2 δ log ( 4 / 3 ) . 0<\eta_\delta'(r)=\frac{1}{2(r+\delta)\sqrt{-\log(r+\delta)}}\le\frac{1}{2\delta\sqrt{\log(4/3)}}. 0 < η δ ′ ( r ) = 2 ( r + δ ) − log ( r + δ ) 1 ≤ 2 δ log ( 4/3 ) 1 .
Extending η δ \eta_\delta η δ to a Lipschitz function on R \mathbb{R} R , the chain rule gives w δ : = η δ ( f t ) ∈ V w_\delta:=\eta_\delta(f_t)\in\mathbb{V} w δ := η δ ( f t ) ∈ V with ∣ D w δ ∣ = η δ ′ ( f t ) ∣ D f t ∣ |Dw_\delta|=\eta_\delta'(f_t)|Df_t| ∣ D w δ ∣ = η δ ′ ( f t ) ∣ D f t ∣ m-a.e., and by (20)
∣ D w δ ∣ 2 = ∣ D f t ∣ 2 4 ( f t + δ ) 2 ( − log ( f t + δ ) ) ≤ 1 4 I K ( t ) m-a.e. |Dw_\delta|^2=\frac{|Df_t|^2}{4(f_t+\delta)^2(-\log(f_t+\delta))}\le\frac{1}{4I_K(t)}\quad\text{m-a.e.} ∣ D w δ ∣ 2 = 4 ( f t + δ ) 2 ( − log ( f t + δ )) ∣ D f t ∣ 2 ≤ 4 I K ( t ) 1 m-a.e.
By (SL), w δ w_\delta w δ has a ( 2 I K ( t ) ) − 1 (2\sqrt{I_K(t)})^{-1} ( 2 I K ( t ) ) − 1 -Lipschitz representative, which coincides with the continuous function η δ ( f t ) \eta_\delta(f_t) η δ ( f t ) . Hence, for all x , y ∈ X x,y\in X x , y ∈ X ,
− log ( f t ( y ) + δ ) ≤ − log ( f t ( x ) + δ ) + d ( x , y ) 2 I K ( t ) . \sqrt{-\log(f_t(y)+\delta)}\le\sqrt{-\log(f_t(x)+\delta)}+\frac{d(x,y)}{2\sqrt{I_K(t)}}. − log ( f t ( y ) + δ ) ≤ − log ( f t ( x ) + δ ) + 2 I K ( t ) d ( x , y ) .
Squaring with ( a + b ) 2 ≤ 2 a 2 + 2 b 2 (a+b)^2\le2a^2+2b^2 ( a + b ) 2 ≤ 2 a 2 + 2 b 2 and exponentiating gives
( f t ( x ) + δ ) 2 ≤ ( f t ( y ) + δ ) exp ( d ( x , y ) 2 2 I K ( t ) ) , (f_t(x)+\delta)^2\le(f_t(y)+\delta)\exp\left(\frac{d(x,y)^2}{2I_K(t)}\right), ( f t ( x ) + δ ) 2 ≤ ( f t ( y ) + δ ) exp ( 2 I K ( t ) d ( x , y ) 2 ) ,
and letting δ ↓ 0 \delta\downarrow0 δ ↓ 0 yields (14) . □
Proof of Theorem 1.4
Proof of Theorem 1.4. The function f f f is ( 2 r 0 ) − 1 (2r_0)^{-1} ( 2 r 0 ) − 1 -Lipschitz, 0 ≤ f ≤ 1 / 2 0\le f\le1/2 0 ≤ f ≤ 1/2 and supp ( f ) ⊆ B ‾ r 0 ( x 0 ) \operatorname{supp}(f)\subseteq\overline{B}_{r_0}(x_0) supp ( f ) ⊆ B r 0 ( x 0 ) , so f ∈ LIP f m ( X ) f\in\operatorname{LIP}_{fm}(X) f ∈ LIP f m ( X ) and Lemma 4.14 applies.
Non-degeneracy. The heat flow is injective on L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) . Indeed, let g ∈ L 2 ( m ) g\in L^2(\mathfrak{m}) g ∈ L 2 ( m ) and s > 0 s>0 s > 0 with h s g = 0 h_sg=0 h s g = 0 . By the semigroup property and the self-adjointness of h s / 2 h_{s/2} h s /2 ,
∥ h s / 2 g ∥ L 2 ( m ) 2 = ∫ X g h s / 2 h s / 2 g d m = ∫ X g h s g d m = 0 , \|h_{s/2}g\|^2_{L^2(\mathfrak{m})}=\int_X gh_{s/2}h_{s/2}g\,\mathrm{d}\mathfrak{m}=\int_X gh_sg\,\mathrm{d}\mathfrak{m}=0, ∥ h s /2 g ∥ L 2 ( m ) 2 = ∫ X g h s /2 h s /2 g d m = ∫ X g h s g d m = 0 ,
so h s / 2 g = 0 h_{s/2}g=0 h s /2 g = 0 ; by induction h s 2 − k g = 0 h_{s2^{-k}}g=0 h s 2 − k g = 0 for every k ∈ N k\in\mathbb{N} k ∈ N . Since τ ↦ h τ g \tau\mapsto h_\tau g τ ↦ h τ g is continuous from [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) to L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) , letting k → ∞ k\to\infty k → ∞ gives g = 0 g=0 g = 0 . Now f ≠ 0 f\ne0 f = 0 in L 2 ( m ) L^2(\mathfrak{m}) L 2 ( m ) because supp ( m ) = X \operatorname{supp}(\mathfrak{m})=X supp ( m ) = X , hence h t f ≠ 0 h_tf\ne0 h t f = 0 . Since h t f ≥ 0 h_tf\ge0 h t f ≥ 0 is continuous, there is x ˉ ∈ X \bar{x}\in X x ˉ ∈ X with h t f ( x ˉ ) > 0 h_tf(\bar{x})>0 h t f ( x ˉ ) > 0 . Applying (14) with ( x , y ) = ( x ˉ , x 0 ) (x,y)=(\bar{x},x_0) ( x , y ) = ( x ˉ , x 0 ) gives h t f ( x 0 ) ≥ ( h t f ( x ˉ ) ) 2 exp ( − d ( x ˉ , x 0 ) 2 / ( 2 I K ( t ) ) ) > 0 h_tf(x_0)\ge(h_tf(\bar{x}))^2\exp(-d(\bar{x},x_0)^2/(2I_K(t)))>0 h t f ( x 0 ) ≥ ( h t f ( x ˉ ) ) 2 exp ( − d ( x ˉ , x 0 ) 2 / ( 2 I K ( t ))) > 0 .
Integration. Applying (14) with ( x , y ) = ( x 0 , y ) (x,y)=(x_0,y) ( x , y ) = ( x 0 , y ) , for every y ∈ B r ( x 0 ) y\in B_r(x_0) y ∈ B r ( x 0 ) we get
h t f ( y ) ≥ ( h t f ( x 0 ) ) 2 exp ( − r 2 2 I K ( t ) ) . h_tf(y)\ge(h_tf(x_0))^2\exp\left(-\frac{r^2}{2I_K(t)}\right). h t f ( y ) ≥ ( h t f ( x 0 ) ) 2 exp ( − 2 I K ( t ) r 2 ) .
Integrating over B r ( x 0 ) B_r(x_0) B r ( x 0 ) and using h t f ≥ 0 h_tf\ge0 h t f ≥ 0 and the L 1 ( m ) L^1(\mathfrak{m}) L 1 ( m ) -contractivity of h t h_t h t ,
m ( B r ( x 0 ) ) ( h t f ( x 0 ) ) 2 exp ( − r 2 2 I K ( t ) ) ≤ ∫ B r ( x 0 ) h t f d m ≤ ∥ f ∥ L 1 ( m ) , \mathfrak{m}(B_r(x_0))(h_tf(x_0))^2\exp\left(-\frac{r^2}{2I_K(t)}\right)\le\int_{B_r(x_0)}h_tf\,\mathrm{d}\mathfrak{m}\le\|f\|_{L^1(\mathfrak{m})}, m ( B r ( x 0 )) ( h t f ( x 0 ) ) 2 exp ( − 2 I K ( t ) r 2 ) ≤ ∫ B r ( x 0 ) h t f d m ≤ ∥ f ∥ L 1 ( m ) ,
which is (3) . Since every bounded set is contained in a ball centred at x 0 x_0 x 0 , m \mathfrak{m} m is finite on bounded sets; then ( X , d , m ) (X,d,\mathfrak{m}) ( X , d , m ) satisfies the assumptions of Definition 1.1, with (G) given by (3) for x ˉ = x 0 \bar{x}=x_0 x ˉ = x 0 , and it is an RCD ( K , ∞ ) \operatorname{RCD}(K,\infty) RCD ( K , ∞ ) space. □
Acknowledgements
The author thanks Nicola Gigli for carefully checking the arguments of this note. This work was supported by the Italian Ministry of University and Research (MUR) through the programme “Fondo Italiano per la Scienza” (FIS 3), project “Modern challEnges in Geometric Analysis – MEGA” (PI: N. Gigli), CUP G53C25000920001. Parts of the arguments in this note were developed, checked and typeset with the assistance of the large language models Claude Opus 5.5 (Anthropic) and Gemini 4 Argon (Google DeepMind). The author takes full responsibility for the content. The author is a member of the GNAMPA group of INdAM.
Data availability
Data sharing is not applicable to this article, as no datasets were generated or analysed during the current study.
Conflict of interest
The author has no competing interests to declare that are relevant to the content of this article.
References
(Marco Picerni) SISSA, Via Bonomea 265, 34136 Trieste, Italy
Email address: mpicerni@sissa.it