Departments of Mathematics Montana State University Fall 2015 Prof. Kevin Wildrick An introduction to non-smooth analysis and geometry Lecture 7: The local Lipschitz constant and Stepanov’s Theorem 1. The Local Lipschitz constant Rademacher’s theorem is clearly non-optimal in the sense that the Lipschitz condition is global, while differentiation is a local issue. It is indeed possible to improve Rademacher’s theorem to a more appropriate local statement. The key quantity is a local version of the Lipschitz condition. Definition 1.1. Let f : X → Y be a mapping between metric spaces, and let x ∈ X. We define the (upper) local Lipschitz constant at x by supy∈B(x,r) dY (f (x), f (y)) . r r→0 Similarly, define the lower local Lipschitz constant at x by supy∈B(x,r) dY (f (x), f (y)) . lip f (x) := lim inf r→0 r Theorem 1.2 (Stepanov’s Theorem). Let f : Rn → R be a mapping. Then f is differentiable at almost every point of the set Lip f (x) := lim sup Sf := {x ∈ Rn : Lip f (x)} < ∞. 2. Extension of Lipschitz functions In order to prove Stepanov’s Theorem, we will first need another pillar of analysis: the Lipschitz extension theorem. The proof of this result is easy, but there is a large body of much more difficult results that vastly generalize it. Theorem 2.1 (McShane-Whitney Lipschitz Extension). Let A be a non-empty subset of an arbitrary metric space, and let f : A → R be an L-Lipschitz function. Then there is an L-Lipschitz function F : X → R such that F |A = f . Proof. Consider, for each a ∈ A, the globally defined function fa : X → R given by fa (x) = f (a) + Ld(x, a). Then, by the triangle inequality, |fa (x) − fa (y)| = L|d(x, a) − d(y, a)| ≤ Ld(x, y), and so fa is again L-Lipschitz. We will show that F (x) := inf fa (x) a∈A is also L-Lipschitz, and restricts to f on A . First note that if a ∈ A, then F (a) ≤ fa (a) = f (a) < ∞, 0 and moreover, if a ∈ A is an arbitrary point of A, then fa0 (a) = f (a0 ) + Ld(a, a0 ) ≥ f (a) by the Lipschitz condition. This shows that F |A = f. 1 Now, for x, y ∈ X and a ∈ A, we see that fa (y) ≥ fa (x) − Ld(x, y) ≥ F (x) − Ld(x, y). Hence F (y) ≥ F (x) − Ld(x, y), which implies (using the finiteness of F ) F (x) − F (y) ≤ Ld(x, y). Exchanging the roles of x and y yields the desired result. 3. Density points and derivatives Lemma 3.1. Let (X, d, µ) be a doubling metric measure space, let A ⊆ X be a measurable set, and let x0 be a density point of A. Then, for each > 0, there exists a radius r > 0 such that for each x ∈ B(x0 , r), there is a point y ∈ A with d(x, y) ≤ d(x, x0 ). Proof. If not, then there is > 0 and a sequence of points {xj }j∈N Rn tending to x0 such that A ∩ B(xj , d(xj , x0 )) = ∅. This implies that µ(A ∩ B(x0 , (1 + )d(xj , x0 )) µ(B(xj , d(xj , x0 )) ≤1− µ(B(x0 , (1 + )d(xj , x0 )) µ(B(xj , (1 + )d(xj , x0 )) s 1 ≤1− <1 C 1+ where C and s are the constants associated to the homogeneity condition for µ (See Lecture 3 Proposition 2.3). Letting xj tend to x0 now violates the density condition. 4. The proof of Stepanov’s Theorem Proof of Stepanov’s Theorem. For j ∈ N, consider the sets 1 |f (x) − f (y)| n Cj := x ∈ R : |x − y| < =⇒ ≤j . j |x − y| We first claim that [ Cj . Sf = j∈N If x ∈ / S j∈N Cj , then there is a sequence of points {yj }j∈N such that |x − yj | < 1/j and |f (x) − f (yj )| > j. |x − yj | Let rj = 2|x − yj |, then Lip f (x) ≥ lim supy∈B(x,rj ) |f (x) − f (y)| j→∞ rj j , j→∞ 2 ≥ lim showing that x ∈ / Sf . On the other hand, if x ∈ / Sf , then we may find a sequence of scales rj & 0 and a sequence of points yj ∈ B(x, rj ) such that |f (x) − f (yj )| = ∞. j→∞ rj lim By passing to a subsequence, we may assume that for each j ∈ N, it holds that rj < 1/j and |f (x) − f (yj )| ≥ j. rj Now, |x − yj | ≤ rj implies that |f (x) − f (yj )| ≥j |x − yj | as well. This shows that x ∈ / Cj for each j ∈ N. Although this isn’t necessary for the proof, note that each Cj is closed. If {xk }k∈N ∈ Cj tend to a point x, and |x − y| < 1/j, then |xk − y| < 1/j as well for sufficiently large k, and hence |f (x) − f (y)| ≤ lim |f (x) − f (xk )| + |f (xk ) − f (y)| ≤ lim j|x − xk | + j|xk − y| = j|x − y|. k→∞ j→∞ This implies that x ∈ Cj . This also gives us a hint as to how to proceed; if x and x0 are elements of Cj with the property that |x − x0 | < 1/j, then by definition |f (x) − f (x0 )| ≤ j|x − x0 |. S Hence, if we write Cj = k∈N Cj,k where each Cj,k has diameter at most 1/j, then f |Cj,k is j-Lipschitz. To find such sets, one can extract a countable subcover of the cover {B(x, 1/2j)}x∈Cj and intersect with Cj . This construction also implies that each Cj,k is closed. Now, using the McShane-Whitney extension theorem, we may extend each f |Cj,k to a globally defined j-Lipschitz function Fj,k , which, by Rademacher’s theorem, is differentiable almost everywhere, and in particular, at almost every point of Cj,k . It remains to show that f is also differentiable at almost all of the points of Cj,k ; we will show that f is differentiable at each density pont of Cj,k . To this end, let x0 be a density point of C := Cj,k at which F := Fj,k is differentiable, and let > 0. By Lemma 3.1, there is a radius r > 0 such that if x ∈ B(x0 , r), then there is a point y ∈ C with |x − y| < |x − x0 |. For r < 1/j, the definition of C implies that |f (x) − f (y)| ≤ j|x − y| ≤ j|x − x0 |. Hence, |f (x) − f (x0 ) − ∇F (x0 )(x − x0 )| |F (y) − F (x0 ) − ∇F (x0 )(y − x0 )| ≤ j+ (1 + j) |x − x0 | |y − x0 | + |∇F (x0 )|j. If x → x0 , then y → x0 as well. Hence, the differentiability of F at x0 implies that |f (x) − f (x0 ) − ∇F (x0 )(x − x0 )| lim ≤ j + |∇F (x0 )|j. x→x0 |x − x0 | Letting tend to 0 shows that f is differentiable at x0 with ∇f (x0 ) = ∇F (x0 ).