346 CHAPTER 14. THE LEBESGUE INTEGRAL

Now these balls Bk are disjoint and contained in a bounded set so letting 1 > θ > θ 1, if nis large enough, and since the sum on the right is the tail of a convergent series,

∑k=n

m(Bk)≤ (θ −θ 1) m̄(F)

Thus there exists n1 such that m̄(

F \∪n1j=1B j

)< θ m̄(F) . If m̄

(F \∪n1

j=1B j

)= 0, stop.

Otherwise, do for F \∪n1j=1B j exactly the same thing that was done for F. Since ∪n1

j=1B j isclosed, you can arrange to have the approximating open set be contained in the open set(∪n1

j=1B j

)C. It follows there exist disjoint closed balls from C called Bn1+1, · · · ,Bn2 such

thatm̄((

F \∪n1j=1B j

)\∪n2

j=n1+1B j

)< θ m̄

(F \∪n1

j=1B j

)< θ

2m̄(F)

continuing this way and noting that limn→∞ θn = 0 while m(F)< ∞, this shows the desired

result. Either the process stops because m̄(

F \∪nkj=1B j

)= 0 or else you obtain an infinite

sequence {Bk} and m̄(

F \∪∞j=1B j

)≤ m̄

(F \∪nk

j=1B j

)≤ θ

km̄(F) for each k, showing that

m̄(

F \∪∞j=1B j

)= 0.

14.13 Differentiation of Increasing FunctionsAs a spectacular application of the covering theorem, is the famous theorem that an increas-ing function has a derivative a.e. Here the a.e. refers to Lebesgue measure, the Stieltjesmeasure from the increasing function F (x) = x.

I will write yn ↑ x to mean limn→∞ yn = x and yn < x,yn ≤ yn+1. I will write yn ↓ x tomean limn→∞ yn = x and yn > x,yn ≥ yn+1.

Definition 14.13.1 The Dini derivates are as follows. In these formulas, f is areal valued function defined on R. yn ↓ x refers to a decreasing sequence as just described.

D+ f (x) ≡ sup{

lim supn→∞

f (yn)− f (x)yn− x

: yn ↓ x},

D+ f (x) ≡ inf{

lim infn→∞

f (yn)− f (x)yn− x

: yn ↓ x},

D− f (x) ≡ sup{

lim supn→∞

f (yn)− f (x)yn− x

: yn ↑ x},

D− f (x) ≡ inf{

lim infn→∞

f (yn)− f (x)yn− x

: yn ↑ x}

The notation means that the sup and inf refer to all sequences of the sort described in {} .

Lemma 14.13.2 The function f : R→ R has a derivative if and only if all the Diniderivates are equal for any sequence just described.

Proof: If D+ f (x) = D+ f (x) , then for any yn ↓ x, it must be the case that

lim supn→∞

f (yn)− f (x)yn− x

= lim infn→∞

f (yn)− f (x)yn− x

= limn→∞

f (yn)− f (x)yn− x

= D+ f (x) = D+ f (x)

346 CHAPTER 14. THE LEBESGUE INTEGRALNow these balls B, are disjoint and contained in a bounded set so letting 1 > 0 > 61, ifnis large enough, and since the sum on the right is the tail of a convergent series,coym (By) < (8-01) m(F)k=nThus there exists n; such that m (F \ Ui Bi) < Om(F).Ifm (F \ Uj! Bi) = 0, stop.Otherwise, do for F \U'!_,B; exactly the same thing that was done for F. Since Ui! ,B; isclosed, you can arrange to have the approximating open set be contained in the open set(UiLB i) It follows there exist disjoint closed balls from @ called By,41,--+ ,Bn, suchthatm ((F\UiL Bi) \ Uni Bi) < Om (F\UjLB)) < 0°m(F)continuing this way and noting that lim,_,.. 0” = 0 while m(F) < ©, this shows the desiredresult. Either the process stops because m (F \ Ure |B i) = 0 or else you obtain an infinitesequence {B,} and m (F \ UE B)) <m (F \ Uik Bj) < 0*im(F) for each k, showing thatm(F\UB;) =0. &14.13 Differentiation of Increasing FunctionsAs a spectacular application of the covering theorem, is the famous theorem that an increas-ing function has a derivative a.e. Here the a.e. refers to Lebesgue measure, the Stieltjesmeasure from the increasing function F (x) = x.I will write y, +x to mean limp... y, =x and yy <X,¥n < yn41. I will write y, | x tomean limy— soo Yn =X and yy > X,Y > Yn41-Definition 14.13.1 The Dini derivates are as follows. In these formulas, f is areal valued function defined on R. yn | x refers to a decreasing sequence as just described.D'f(x) = sup {tim sup FE), yh,D.f(x) = in {Kim nt FOAL) 5, Lah,Df(x) = sup {tim sup LPF), ph,D_f(x) = int {tim ing SFE 5, }The notation means that the sup and inf refer to all sequences of the sort described in {}.Lemma 14.13.2 The function f :R—R has a derivative if and only if all the Diniderivates are equal for any sequence just described.Proof: If D* f (x) = Df (x), then for any y, | x, it must be the case thatlim sup Peis tim ing SOFtim £0) -FO) _ p+ py =D, F(x)n—yoo Yn —X