Notation. For real p<q write J=[p,q] for the closed interval with these endpoints, ∣J∣=q−p and, when J⊆I, ΔJ=g(q)−g(p); since g is nondecreasing, ΔJ≥0. By claim 4 of Existence of Lebesgue Measure on the Real Line we have λ(J)=∣J∣, and J is the closed ball with centre (p+q)/2 and radius (q−p)/2, so families of such intervals are families of closed balls in the sense of The Vitali Covering Theorem in Rn, an interval of length ℓ having radius ℓ/2.
Step S: increments add up inside an interval. Let J⊆I be a closed interval and let K1,…,KM⊆J be pairwise disjoint closed intervals. Then ∑j=1MΔKj≤ΔJ. Indeed, relabelling the finitely many intervals we may assume their left endpoints increase, writing Kj=[pj,qj] with p1≤⋯≤pM; disjointness then forces qj<pj+1 for j<M, and g nondecreasing gives g(qj)≤g(pj+1). Writing J=[p,q] we have p≤p1 and qM≤q, so
j=1∑M(g(qj)−g(pj))≤j=1∑M(g(qj)−g(pj))+j=1∑M−1(g(pj+1)−g(qj))=g(qM)−g(p1)≤g(q)−g(p),
all the discarded terms being nonnegative.
Step 0: reduction to a compact subinterval. For every x∈I there are rationals a<b with x∈(a,b) and [a,b]⊆I: since I is open there is a real ρ>0 with {t:∣t−x∣<ρ}⊆I, and The Rational Numbers are Dense in the Real Numbers supplies rationals a and b with x−ρ<a<x<b<x+ρ. The set of pairs of rationals is countable by The Integers and the Rational Numbers are Countable and Products and Powers of Countable Sets, so I is the union of a countable family of open intervals (a,b) with a,b rational and [a,b]⊆I. Since a countable union of null sets is null by the null-set claim, it suffices to prove that (a,b)∖D is null for each such pair.
Fix such a<b, put J0=[a,b]⊆I and T=ΔJ0=g(b)−g(a)≥0. All intervals below are contained in (a,b), and every subset of (a,b) has outer measure at most b−a<∞ by monotonicity and agreement on Borel sets.
Step 1: the points of unbounded difference quotients. For M∈N let GM be the set of x∈(a,b) such that for every real η>0 there is a closed interval J with x∈J⊆(a,b), 0<∣J∣<η and ΔJ>M∣J∣. We claim λ∗(GM)≤T/M.
The pairs carrying these intervals finely cover GM in the sense of the fine covering condition: given x∈GM and η>0, apply the definition with 2η in place of η to obtain such a J containing x of radius ∣J∣/2<η. Let ε>0. By the Vitali covering theorem there are finitely many pairwise disjoint intervals J1,…,JN of this family with λ∗(GM∖⋃iJi)≤ε. Using subadditivity and monotonicity of λ∗, then M∣Ji∣<ΔJi, and finally Step S with J=J0,
λ∗(GM)≤ε+i=1∑N∣Ji∣≤ε+M1i=1∑NΔJi≤ε+MT.
This holds for every real ε>0; were λ∗(GM)>T/M, taking ε=(λ∗(GM)−T/M)/2 would contradict it. Hence λ∗(GM)≤T/M, as claimed. Consequently P=⋂M∈NGM satisfies λ∗(P)≤T/M for every M∈N, by monotonicity; if λ∗(P) were positive then The Archimedean Property of the Real Numbers would supply M∈N with T/M<λ∗(P), a contradiction. So λ∗(P)=0 and P is null by the null-set claim.
Step 2: the points of oscillating difference quotients. Let u,v be rationals with 0≤u<v and let Eu,v be the set of x∈(a,b) such that for every real η>0 there are closed intervals J,K with x∈J⊆(a,b), x∈K⊆(a,b), 0<∣J∣<η, 0<∣K∣<η,
ΔJ<u∣J∣andΔK>v∣K∣.
We claim λ∗(Eu,v)=0. Put s=λ∗(Eu,v), a real number with 0≤s≤b−a, and suppose 0<s. Let ε be a real number with 0<ε<s/2.
By outer regularity of the outer measure there is an open O′ with Eu,v⊆O′ and λ(O′)≤s+ε; put O=O′∩(a,b), again open, still containing Eu,v, with λ(O)≤s+ε.
First application. Let F1 consist of the pairs carrying the closed intervals J with J⊆O and ΔJ<u∣J∣. This family finely covers Eu,v: for x∈Eu,v and η>0, choose a real ρ>0 with {t:∣t−x∣<ρ}⊆O and apply the definition of Eu,v with min{2η,ρ} in place of η; the resulting J contains x, has ∣J∣<ρ and therefore J⊆O, and has radius <η. By the Vitali covering theorem there are pairwise disjoint J1,…,JN∈F1 with λ∗(Eu,v∖⋃iJi)≤ε. Being pairwise disjoint Borel subsets of O, they satisfy ∑i∣Ji∣≤λ(O)≤s+ε, whence
i=1∑NΔJi<ui=1∑N∣Ji∣≤u(s+ε).(1)
Second application. Let A=Eu,v∩⋃i=1N(pi,qi), where Ji=[pi,qi]. The finite set F of the endpoints pi,qi is null, being a finite union of singletons, each of measure 0 by claim 4 of Existence of Lebesgue Measure on the Real Line applied to an interval with equal endpoints, and a countable union of null sets being null by the null-set claim; moreover Eu,v⊆(Eu,v∖⋃iJi)∪A∪F, so subadditivity gives s≤ε+λ∗(A), that is λ∗(A)≥s−ε.
Let F2 consist of the pairs carrying the closed intervals K with K⊆(pi,qi) for some i and ΔK>v∣K∣. This family finely covers A: for x∈A pick i with x∈(pi,qi) and a real ρ>0 with {t:∣t−x∣<ρ}⊆(pi,qi), then apply the definition of Eu,v with min{2η,ρ} in place of η. By the Vitali covering theorem there are pairwise disjoint K1,…,KM∈F2 with λ∗(A∖⋃jKj)≤ε, so subadditivity gives
j=1∑M∣Kj∣≥λ∗(A)−ε≥s−2ε>0.
Each Kj lies in exactly one Ji, the Ji being pairwise disjoint; grouping the Kj accordingly and applying Step S inside each Ji gives ∑jΔKj≤∑iΔJi. Combining this with ΔKj>v∣Kj∣ and with (1),
v(s−2ε)≤vj=1∑M∣Kj∣<j=1∑MΔKj≤i=1∑NΔJi<u(s+ε).
This holds for every real ε with 0<ε<s/2. If vs>us then, v and u being fixed, the Archimedean property supplies such an ε with (2v+u)ε<vs−us, contradicting the display; hence vs≤us, so (v−u)s≤0. As u<v and s>0 this is impossible, and therefore s=0.
Step 3: assembling. Let x∈(a,b) belong to none of the sets ⋂MGM and Eu,v (over rationals 0≤u<v). We show x∈D.
For h=0 with x+h∈(a,b) let Jh be the closed interval with endpoints x and x+h and put q(h)=(g(x+h)−g(x))/h. Then ∣Jh∣=∣h∣ and ΔJh=∣g(x+h)−g(x)∣, because g is nondecreasing; as g(x+h)−g(x) and h have the same sign, q(h)=ΔJh/∣Jh∣≥0.
Since x∈/⋂MGM there is M∈N with x∈/GM, so there is a real η0>0 such that every closed interval J with x∈J⊆(a,b) and 0<∣J∣<η0 has ΔJ≤M∣J∣. Applying this to Jh gives 0≤q(h)≤M whenever 0<∣h∣<η0 and x+h∈(a,b). For real η with 0<η≤η0 let
Q(η)={q(h):0<∣h∣<η, x+h∈(a,b)},
a nonempty subset of [0,M]; let m(η) be its greatest lower bound and mˉ(η) its least upper bound, which exist by the least upper bound property and by Existence of the Infimum of a Nonempty Subset of R Bounded Below. As η decreases Q(η) shrinks, so m does not decrease and mˉ does not increase; both stay in [0,M]. Let α be the least upper bound of {m(η):0<η≤η0} and β the greatest lower bound of {mˉ(η):0<η≤η0}; then 0≤α≤β≤M, since m(η)≤mˉ(η) for every η and m(η)≤mˉ(η′) for all η,η′ by comparing at min{η,η′}.
Suppose α<β. By The Rational Numbers are Dense in the Real Numbers choose rationals u,v with α<u<v<β; then 0≤u. For every η with 0<η≤η0 we have m(η)≤α<u, so u is not a lower bound of Q(η) and there is h with 0<∣h∣<η and q(h)<u, that is ΔJh<u∣Jh∣; and mˉ(η)≥β>v, so there is h′ with 0<∣h′∣<η and ΔJh′>v∣Jh′∣. Hence x∈Eu,v, contrary to the choice of x. Therefore α=β; call this common value c.
Finally, let ε>0. There is η1 with m(η1)>c−ε and η2 with mˉ(η2)<c+ε, since c is the least upper bound of the m(η) and the greatest lower bound of the mˉ(η). Put δ=min{η1,η2,x−a,b−x}, a positive real because a<x<b. If 0<∣h∣<δ and x+h∈I, then ∣h∣<min{x−a,b−x} forces x+h∈(a,b), so q(h) belongs to Q(η1) and to Q(η2) and therefore satisfies c−ε<m(η1)≤q(h)≤mˉ(η2)<c+ε. Thus for every real ε>0 there is a real δ>0 such that the difference quotient of g at x differs from c by less than ε for all h with 0<∣h∣<δ and x+h∈I; that is, g has a derivative g′(x)=c at the interior point x of I, so x∈D.
Therefore (a,b)∖D is contained in the union of ⋂MGM and of the sets Eu,v over the countably many pairs of rationals with 0≤u<v, all of which are null; by the null-set claim it is null. By Step 0, I∖D is null.