Lebesgue showed that a striking result that a monotone function is differentiable almost everywhere on an interval. We begin the discussion from a key concept, Vitali Cover.

Definition Vitali Cover

  1. F:F: a collection of closed, bounded non-degenerate(of positive length) intervals.
  2. E:E: a set.
  3. Then FF is a Vitali cover of EE if ∀x∈E\forall x \in E, ∀ϵ>0\forall \epsilon > 0, ∃\exists interval I∈FI \in F, s.t., x∈Ix \in I and μ(I)<ϵ\mu(I) < \epsilon.

Theorem Vitali Covering Lemma

  1. EE a set of finite outer measure.

  2. FF a Vitali cover of EE.

  3. ∀ϵ>0\forall \epsilon > 0, ∃\exists a finite disjoint collection of intervals {Ik}k=1n\{ I_k \} _{k = 1}^n of FF for which

    μ∗(E∖∪k=1nIk)≤ϵ \mu^*(E \setminus \cup_{k = 1}^n I_k) \le \epsilon

Proof: Since EE is of finite outer measure, there is a open set OO of finite measure that contains EE for which μ(O)≤μ∗(E)+ϵ\mu(O) \le \mu^*(E) + \epsilon. For each x∈Ex \in E, we have x∈Ox \in O. As OO is open, ∃ϵ>0\exists \epsilon > 0, s.t., Nϵ(x)⊂ON_\epsilon(x) \subset O. Because FF is a Vitali cover of EE, there exists an interval II that contains xx with length μ(I)<ϵ/2\mu(I) < \epsilon / 2. This means I⊂Nϵ(x)⊂OI \subset N_\epsilon(x) \subset O. Therefore, we may assume that FF contains only intervals within OO.

Then we are going to select a disjoint countable (either finite or infinite) sub-collection of intervals from FF. Suppose nn is a natural number and the collection {Ik}k=1n\{ I_k \} _{k = 1 }^n of FF has been chosen. If {Ik}k=1n\{ I_k \} _{k = 1}^n covers EE, then we are done. Otherwise, we pick a new interval Ik+1I_{k + 1} from FF according to two criteria:

  1. Ik+1I_{k + 1} should be disjoint with the collection {Ik}k=1n\{ I_k \}_{k = 1}^n, i.e., Ik+1∩(∪k=1nIk)=∅I_{k + 1} \cap (\cup_{k = 1}^n I_k) = \emptyset.
  2. The length of Ik+1I_{k + 1} should be as large as possible.

Such Ik+1I_{k + 1} always exists. Let x∈E∖∪k=1nIkx \in E \setminus \cup_{k = 1}^n I_k. As ∪k=1nIk\cup_{k = 1}^n I_k is closed, O∖∪k=1nIkO \setminus \cup_{k = 1}^n I_k is open. Therefore, there is an interval in FF that is inside that open set and contains xx. We can now define the nonempty collection of such intervals in FF as FnF_n:

Fn≐{I∈F:I∩(∪k=1nIk)=∅} F_n \doteq \{ I \in F : I \cap (\cup_{k = 1}^n I_k ) = \emptyset \}

As all interval in FnF_n are contained in OO with finite measure,

αn≐sup⁡I∈Fnμ(I) \alpha_n \doteq \sup_{I \in F_n} \mu(I)

exists and is a real number. Although we may not be able to find an interval II that achieves the maximum length, i.e., μ(I)=αn\mu(I) = \alpha_n, it suffices to find one that is "a little bit" shorter. The criteria is not unique and for our proof we choose an Ik+1I_{k + 1} of measure μ(Ik+1)>αn/2\mu(I_{k + 1}) > \alpha_n / 2.

Now {Ik}k=1∞\{ I_k \}_{k = 1}^\infty is a disjoint collection whose union is within OO. By the countable additivity and monotonicity of measure,

μ(∪k=1∞Ik)=∑k=1∞μ(Ik)≤μ(O)≤μ(E)+ϵ \mu(\cup_{k = 1}^\infty I_k) = \sum_{k = 1}^\infty \mu( I_k) \le \mu(O) \le \mu(E) + \epsilon

We infer that μ(Ik)→0\mu(I_k) \rightarrow 0. Moreover, {Ik}k=1∞\{ I_k \}_{k = 1}^\infty has the following property: for each interval I∈FI \in F that contains an element x∈Ex \in E, either it belongs to or overlaps with {Ik}k=1∞\{ I_k \}_{k = 1}^\infty. Otherwise, II is disjoint with {Ik}k=1∞\{ I_k \}_{k = 1}^\infty. But the selection process guarantees that μ(I)≤2μ(Ik)\mu(I) \le 2 \mu(I_k) for all kk, which implies μ(I)=0\mu(I) = 0, a contradiction (any I∈FI \in F must have positive length).

We claim that for any natural number nn, it holds that

E∖∪k=1nIk⊂∪k=n+1∞5∗Ik E \setminus \cup_{k = 1}^n I_k \subset \cup_{k = n + 1}^\infty 5 * I_k

As before, for x∈E∖∪k=1nIkx \in E \setminus \cup_{k = 1}^n I_k, ∃ I∈Fn\exists \ I \in F_n for which x∈Ix \in I. Then II must belongs to or intersect with ∪k=n+1∞Ik\cup_{k = n + 1}^\infty I_k. Let mm be the first natural number that Im∩I≠∅I_{m} \cap I \neq \emptyset. By the selection process of ImI_m, μ(Im)≥μ(I)/2\mu(I_m) \ge \mu(I) / 2. This means I⊂5∗ImI \subset 5 * I_m.

Let ϵ>0\epsilon > 0, ∃N>0\exists N > 0, s.t., ∑k=N∞μ(Ik)<ϵ/5\sum_{k = N}^\infty \mu(I_k) < \epsilon / 5. This choice of NN, together with the fact that E∖∪k=1nIk⊂(∪k=n+1∞5∗Ik)E \setminus \cup_{k = 1}^n I_k \subset (\cup_{k = n + 1}^\infty 5 * I_k), establishes the desired result.

■\blacksquare

Lemma

Let f:[a,b]→Rf : [a, b] \rightarrow R be a non-decreasing function, then for each α>0\alpha > 0,

αμ∗({x∈[a,b]:Dˉf(x)≥α})≤f(b)−f(a) \alpha \mu^* (\{ x \in [a, b] : \bar D f(x) \ge \alpha \}) \le f(b) - f(a)

and

μ∗({x∈(a,b):Dˉf(x)=∞})=0 \mu^*(\{ x \in (a, b) : \bar D f(x) = \infty \}) = 0

Proof:
Define E={x∈[a,b]:Dˉf(x)≥α}E = \{ x \in [a, b] : \bar D f(x) \ge \alpha \}. For each 0<λ<α0 < \lambda < \alpha, let Fλ={[c,d]⊂[a,b],f(d)−f(c)≥λ(d−c)}F_{\lambda} = \{ [c, d] \subset [a, b], f(d) - f(c) \ge \lambda (d - c) \}. Since Dˉf≥α\bar D f \ge \alpha on EE, FF is a Vitali cover of EE. By Vitali Covering Theorem, there is a finite disjoint sub-collection of intervals {[ck,dk]}k=1n\{ [c_k, d_k] \}_{k = 1}^n of FF for which

μ∗(E∖∪k=1nIk)≤ϵ \mu^*(E \setminus \cup_{k = 1}^n I_k) \le \epsilon

Therefore,

μ∗(E)≤μ(∪k=1nIk)+μ∗(E∖∪k=1nIk)≤μ(∪k=1nIk)+ϵ=∑k=1nμ(Ik)+ϵ=∑k=1n(dk−ck)+ϵ≤∑k=1n1λ[f(dk)−f(ck)]+ϵ≤1λ[f(d)−f(c)]+ϵ \begin{aligned} \mu^*(E) &\le \mu(\cup_{k = 1}^n I_k) + \mu^*(E \setminus \cup_{k = 1}^n I_k) \\ &\le \mu(\cup_{k = 1}^n I_k) + \epsilon \\ &= \sum_{k = 1}^n \mu(I_k) + \epsilon \\ &= \sum_{k = 1}^n (d_k - c_k) + \epsilon \\ &\le \sum_{k = 1}^n \frac{1}{\lambda} [f(d_k) - f(c_k)] + \epsilon \\ &\le \frac{1}{\lambda} [f(d) - f(c)] + \epsilon \end{aligned}

Taking the limit of λ→α\lambda \rightarrow \alpha, and ϵ→0\epsilon \rightarrow 0, we proves

αμ∗(E)≤f(d)−f(c) \alpha \mu^* (E) \le f(d) - f(c)

■\blacksquare.

Theorem Lebesgue

If a function ff is monotone on a interval (a,b)(a, b), then it is differentiable almost everywhere.

Proof:
Without lost of generality, assume ff is increasing. The set of non-differentiable points can be written as

E=∪α,β∈Q{x∈(a,b):Dˉf(x)≥α>β>D‾f(x)} E = \cup_{\alpha, \beta \in Q} \{ x \in (a, b) : \bar D f(x) \ge \alpha > \beta > \underline D f(x) \}

where QQ is the set of rational numbers. It suffices to consider just Eα,β={x∈(a,b):Dˉf(x)≥α>β>D‾f(x)}E_{\alpha, \beta} = \{ x \in (a, b) : \bar D f(x) \ge \alpha > \beta > \underline D f(x) \}. We prove that Eα,βE_{\alpha, \beta} has zero measure and therefore by countable additivity EE has measure zero.

For fix Eα,βE_{\alpha, \beta}, choose an open set OO for which Eα,β⊂O⊂(a,b)E_{\alpha, \beta} \subset O \subset (a, b) and μ∗(O∖Eα,β)≤ϵ\mu^*(O \setminus E_{\alpha, \beta}) \le \epsilon. Let Fβ={[c,d]⊂O,f(d)−f(c)≤β(d−c)}F_{\beta} = \{ [c, d] \subset O, f(d) - f(c) \le \beta (d - c) \} be a Vitali cover of Eα,βE_{\alpha, \beta}. By Vitali Covering Theorem, there is a finite disjoint sub-collection of intervals {[ck,dk]}k=1n\{ [c_k, d_k] \}_{k = 1}^n for which

μ∗(Eα,β∖∪k=1nIk)≤ϵ \mu^*(E_{\alpha, \beta} \setminus \cup_{k = 1}^n I_k) \le \epsilon

Therefore,

μ∗(Eα,β)≥μ(O)−ϵ≥μ(∪k=1nIk)−ϵ=∑k=1nμ(Ik)−ϵ=∑k=1n(dk−ck)−ϵ≥∑k=1n1β[f(dk)−f(ck)]−ϵ \begin{aligned} \mu^*(E_{\alpha, \beta}) &\ge \mu(O) - \epsilon \\ &\ge \mu(\cup_{k = 1}^n I_k) - \epsilon \\ &= \sum_{k = 1}^n \mu(I_k) - \epsilon \\ &= \sum_{k = 1}^n (d_k - c_k) - \epsilon \\ &\ge \sum_{k = 1}^n \frac{1}{\beta} [f(d_k) - f(c_k)] - \epsilon \\ \end{aligned}

On the other hand, for each interval [ck,dk][c_k, d_k], by the previous lemma,

f(dk)−f(ck)≥αμ({x∈[ck,dk]∩Eα,β}) f(d_k) - f(c_k) \ge \alpha \mu(\{x \in [c_k, d_k] \cap E_{\alpha, \beta} \})

Together with sub-additivity of outer measure,

μ∗(Eα,β)≤μ∗(∪k=1nIk∩Eα,β)+μ∗(Eα,β∖∪k=1nIk)≤μ(∪k=1nIk∩Eα,β)+ϵ≤∑k=1n1α[f(dk)−f(ck)]+ϵ \begin{aligned} \mu^*(E_{\alpha, \beta}) &\le \mu^*(\cup_{k = 1}^n I_k \cap E_{\alpha, \beta}) + \mu^*(E_{\alpha, \beta} \setminus \cup_{k = 1}^n I_k) \\ &\le \mu(\cup_{k = 1}^n I_k \cap E_{\alpha, \beta}) + \epsilon \\ &\le \sum_{k = 1}^n \frac{1}{\alpha} [f(d_k) - f(c_k)] + \epsilon \\ \end{aligned}

Combined, we obtained

∑k=1n1β[f(dk)−f(ck)]−ϵ≤μ∗(Eα,β)≤∑k=1n1α[f(dk)−f(ck)]+ϵ \sum_{k = 1}^n \frac{1}{\beta} [f(d_k) - f(c_k)] - \epsilon \le \mu^*(E_{\alpha, \beta}) \le \sum_{k = 1}^n \frac{1}{\alpha} [f(d_k) - f(c_k)] + \epsilon

By arbitrariness of ϵ\epsilon, it holds that

βμ∗(Eα,β)≥αμ∗(Eα,β) \beta \mu^*(E_{\alpha, \beta}) \ge \alpha \mu^*(E_{\alpha, \beta})

It concludes that μ∗(Eα,β)=0\mu^*(E_{\alpha, \beta}) = 0.

■\blacksquare.

Reference

  1. H.L.Royden, P.M. Fitzpatrick, Real Analysis