Departments of Mathematics Montana State University Fall 2015 Prof. Kevin Wildrick An introduction to non-smooth analysis and geometry Lecture 15: The Proof of Cheeger’s Theorem 1. The Lip-lip condition and finite dimensionality In this section, we will show Theorem 1.1. Suppose that (X, d, µ) is a complete doubling metric space that satisfies the K-Lip-lip condition. Then there is a constant N0 such that for almost every x ∈ X, every (N0 + 1)-tuple of Lipschitz functions is dependent to first order at x. As we have seen in the past, if (X, d, µ) has a measurable differentiable structure, then the conclusion of Theorem 1.1 is valid. We will see that in fact the two are essentially equivalent. Let f1 , . . . fN : X → R be a collection of Lipschitz functions so that the set of points ind where they are not dependent has positive µ measure. We must bound the number N from above. We consider the collection of all rational linear combinations of f1 , . . . , fn : F := { N X λi fi : X → R : λ1 , . . . , λN ∈ Q}. i=1 Note that F is a countable collection. Definition 1.2. Let > 0. A set Y ⊆ X is -good for F if for every f ∈ F, there is a scale rf > 0 such that if 0 < r < rf and x ∈ Y , the n Lip f (x) − ≤ lip f (x) − ≤ var f (x, r) ≤ Lip f (x) + . K We say that Y is good for F if it is -good for all > 0. Notice that the first inequality in the above definition is valid for µ-almost every x ∈ X by the K-Lip-lip condition. Lemma 1.3. Let Y0 be any measurable subset of X of finite (and positive) measure. For every δ > 0, there is a subset Y ⊆ Y0 so that µ(Y0 \Y ) < δ and Y is good for F. Proof. For f ∈ F, recall that lip f (x) = lim inf var f (x, r) ≤ lim sup var f (x, r) = Lip f (x). r→0 r→0 By a version of Egoroff’s theorem for lim inf and lim sup, for any δ > 0, there is a set Y0 (f, δ) ⊆ Y0 so that the convergence of the lim inf and lim sup above is uniform on Y0 (f, δ) and µ(Y0 \Y0 (f, δ)) < δ. In other words, for > 0, there is r such that if r < r , then for all x ∈ Y0 (f, δ) lip f (x) − ≤ var f (x, r) ≤ Lip f (x) + . This is just another way of saying that Y0 (f, δ) is good for f . Define i+1 [ 2−(i+1) Yi+1 = Yi f j , δ . i + 1 j=1 1 Then Y0 ⊇ Y1 ⊇ Y2 . . . , and so we set Y = ∞ \ Yi i=1 and see that µ (Y0 \Y 0 ) ≤ δ. Then Y is good for each function in F. Let Y be a subset of ind satisfying the conclusions of Lemma 1.3. By Lusin’s theorem we may also assume that for each f ∈ F, the function lip f is continuous on Y . Lemma 1.4. Let x ∈ Y be a point of density of Y . Let (X∞ , d∞ , x∞ ) be a GromovHausdorff tangent of X at x such that each f ∈ F has a tangent function f∞ : X∞ → R. Then, for every f ∈ F, every p ∈ X∞ and every r > 0, Lip f (x) ≤ var f∞ (p, r) ≤ Lip f (x). K In particular, the Lipschitz constant of f∞ is no greater than Lip f (x). Proof. The novelty of this lemma, compared to what we did in the last lecture, is that we estimate var f∞ at every point of X∞ , not just at the basepoint. Let {φi : (X∞ , d∞ , x∞ ) → (X, d/ri , x)} be a Hausdorff approximation. We begin with an exercise: since x is a density point of Y and µ is doubling, we may find another Hausdorff approximation {φ0i : (X∞ , d∞ , x∞ ) → (Y, d/ri , x)}. To complete this exercise, you may want to use the lemma involved in the proof of Stepanov’s theorem. We have seen that f ◦ φ0i − f (x) f∞ = lim i→∞ ri uniformly on compact subsets of X∞ . Fix p, q ∈ X∞ , and set pi = φ0i (p) and qi = φ0i (q). Then for each > 0, for all sufficiently large i ∈ N, we use the fact that Y is good for F to see that |f∞ (p) − f∞ (q)| |f (pi ) − f (qi )| ≤ + d∞ (p, q) d(pi , qi ) ≤ (1 + ) var f (pi , (1 + )d(pi , qi )) + ≤ Lip f (pi ) + 2. Notice that as the sequence of scales ri is tending to 0, it holds that d(pi , x) → 0. Hence, the continuity of Lip f on Y implies that for sufficiently large i, Lip f (pi ) ≤ Lip f (x) + . An easy argument now shows that var f∞ (p, r) ≤ Lip f (x) as well as the final statement of the lemma. For the other inequality, for > 0, we may choose ai ∈ B(pi , (r − )ri )) so that var f (pi , (r − )ri )) ≤ f (pi ) − f (ai ) + (r − )ri Since pi tends to x, and x is a density point of Y , an argument as we have seen in the proof of Stepanov’s theorem implies that we may assume that ai is a point of Y , and hence ai = φ0i (vi ) for some vi ∈ B∞ (p, r). Thus var f (pi , (r − )ri )) ≤ |fi (p) − fi (vi )| + . (r − ) Since fi converges uniformly to f∞ on compact sets, we see that for sufficiently large i ∈ N, var f (pi , (r − )ri )) ≤ |f∞ (p) − f∞ (vi )| r + 2 ≤ var f∞ (p, r) + 2. r− r− Since Lip f is continuous on Y , Lip f (x) = lim Lip f (pi ). i→∞ Since Y is good for f , for sufficiently large i, Lip f (pi ) ≤ K(var f (pi , (r − )ri )) + ). Thus, combining and sending → 0, we see that Lip f (x) ≤ K var f∞ (p, r). Let F∞ be a collection of tangent functions as in the statement of Lemma 1.4 (we know such a collection exists by Gromov’s compactness theorem, as in the last lecture). Note that F∞ is the Q-span of {(f1 )∞ , . . . , (fN )∞ ). Since each element of F∞ satisfies the uniform quasilinearity condition (i.e., the conclusion of Lemma 1.4), so does each element of the R-span of {(f1 )∞ , . . . , (fN )∞ }. Denote this R-span by F ∞ . Now, recall that (X∞ , d∞ ) is a doubling metric space. Consider a maximal 1/(4K)separated set T in B∞ (x∞ , 1). Then card T is finite, and depends only on K and the doubling constant. Consider that the dimension of L∞ (T ) is card T , and that there is a natural restriction mapping from F ∞ to L∞ (T ) given by f∞ 7→ (f∞ )|T . Suppose that (f∞ )|T is zero, so that f∞ is in the kernel of this mapping. Let z ∈ B := B∞ (x∞ , 1), and let z 0 ∈ T with d∞ (z, z 0 ) ≤ 1/(4K). Then |f∞ (z)| = |f∞ (z) − f∞ (z 0 )| ≤ Lip f (x) var f∞ (x∞ , 1) ≤ . 4K 4 By the triangle inequality, var f∞ (x∞ , 1) ≤ 2||f∞ |B ||L∞ , and so the above argument shows that ||f∞ |B ||L∞ = 0. This implies that f∞ |B = 0, and so var f∞ (x∞ , 1) = 0. By the quasilinearity condition, this implies that the variation on any ball is 0, and so f∞ is constantly 0. This means that F ∞ is the injective image of L∞ (T ), and so its dimension is at most card T . Suppose that N > card T . Then there are b1 , . . . , bN ∈ R such that n X bi (fi )∞ = 0. i=1 We may write bi = limk→∞ ai,k where each ai,k is rational. Then n X i=1 ai,k (fi )∞ converges uniformly on compact subsets of X∞ to the zero function as k → ∞. This implies that for any p ∈ X∞ and r > 0, ! n X lim var ai,k (fi )∞ (p, r) = 0. k→∞ Since each of the functions Pn i=1 n X i=1 ai,k (fi )∞ is uniformly quasilinear, and ! n X ai,k (fi )∞ = ai,k fi , i=1 i=1 ∞ this implies that lim Lip k→∞ n X ! ai,k fi (x) = 0. i=1 Finally, Lip n X i=1 ! bi f i (x) ≤ lim sup Lip k→∞ n X i=1 ! ai,k fi (x) + Lip n X ! (bi − ai,k )fi (x) = 0. i=1 This is exactly the statement that the fi are dependent to first order at x, as desired.