Chapter V · Limits of functions of a continuous variable. Continuous and discontinuous functions

The Heine–Borel theorem and uniform continuity

§§106–107 40 min

From infinitely many local facts to one global fact; covering a closed interval by finitely many intervals, and why the oscillation of a continuous function can be made small uniformly.

Why this matters

Continuity is a local property: near each point, separately, the function varies little. Many theorems need a global statement: one step size that works everywhere at once. The bridge from local to global is the Heine–Borel theorem of §106: if every point of a closed interval has an interval of its own around it, then finitely many of those intervals already cover the whole. In §107 Hardy uses it to prove that a continuous function on [a,b][a, b] has small oscillation on every sufficiently short piece. That fact, today called uniform continuity, is exactly what the definite integral of Chapter VII needs.

The idea

Suppose each point yy of [0,1][0, 1] comes with an open interval (y−δ(y),y+δ(y))(y - \delta(y), y + \delta(y)). Infinitely many intervals certainly cover [0,1][0, 1], since each point is in its own. Can we throw all but finitely many away? If the radii δ(y)\delta(y) were all at least some fixed r>0r > 0, yes: the intervals round 0,r,2r,…0, r, 2r, \dots would do. The danger is that δ(y)\delta(y) shrinks towards 00 somewhere, forcing ever more intervals.

Hardy’s second proof measures this danger with a function. For each xx let ρ(x)\rho(x) be the largest λ\lambda such that the window (x−λ,x+λ)(x - \lambda, x + \lambda) fits inside one interval of the cover. He shows that ρ\rho is continuous; on a closed interval it then has a positive least value mm (last lesson). Any window shorter than 2m2m, placed anywhere, fits inside one interval of the cover, and finitely many such windows reach across [0,1][0, 1].

Explore The covering game

Try this. With the first cover, drag the purple window along the bottom and change its length ll. Find the largest ll for which every position fits. Then choose the wavy cover and do the same. Finally choose δ(y)=y/2\delta(y) = y/2 on (0,1](0, 1].

Cover of [0,1][0, 1] by the intervals (y−δ(y), y+δ(y))(y - \delta(y),\ y + \delta(y)), δ(y)=0.05+0.2y\delta(y) = 0.05 + 0.2y
xρ(x)
  • ρ(x)
  • δ(x)
  • minimum m of ρ
  • window of length l (drag)
  • interval of the cover containing it; below: the finite sub-cover
The least value of ρ is m≈0.0499m \approx 0.0499, so every window shorter than 2m≈0.09982m \approx 0.0998 fits inside one interval of the cover. This window fits, inside the interval round y=0.34y = 0.34. Overlapping windows of length l, starting at 0,12l,l,…0, \tfrac12 l, l, \dots, need only 24 intervals of the cover: a finite sub-cover (the first 16 are drawn).
What to notice

A window fits everywhere exactly when l<2ml < 2m, where mm is the minimum of the blue curve ρ\rho; then the green rows below show a finite sub-cover, one interval per window. For the wavy cover mm is small, so many intervals are needed, but still finitely many. For δ(y)=y/2\delta(y) = y/2, ρ(x)=x/2\rho(x) = x/2 sinks to 00 at the left end: no positive ll works near 00, and no finite sub-cover exists. The interval (0,1](0, 1] is not closed, and the point 00 has no interval of its own.

Precise statements

Hardy states the theorem for an arbitrary set of intervals II inside [a,b][a, b], finite or infinite, overlapping or not.

Theorem The Heine–Borel theorem Hardy: §106

Let II be a set of intervals, each contained in [a,b][a, b], such that

  1. every point of [a,b][a, b] other than aa and bb lies inside (not at an end of) at least one interval of II, and
  2. aa is the left-hand end of at least one interval of II, and bb the right-hand end of at least one.

Then a finite number of the intervals of II already have properties 1 and 2.

The heart of the proof is a lemma about a single number ll.

Lemma A uniform window length Hardy: §106

Under the hypotheses of the theorem there is a positive number ll such that every sub-interval of [a,b][a, b] of length ll is contained in at least one interval of II.

The lemma gives the theorem at once: cover [a,b][a, b] by finitely many overlapping sub-intervals of length ll (starting at aa, each shifted by 12l\tfrac12 l, the last ending at bb), and for each choose one interval of II containing it.

Now fix ε>0\varepsilon > 0 and a function ϕ\phi continuous on [a,b][a, b]. By continuity, each point ξ\xi has an interval round it in which the oscillation of ϕ\phi is less than ε\varepsilon (Hardy, whose tolerance is δ\delta rather than ε\varepsilon, calls these the δ\delta-intervals of ξ\xi). Heine–Borel selects finitely many of them.

Theorem Finitely many pieces of small oscillation Hardy: Theorem I, §107

If ϕ\phi is continuous on [a,b][a, b] and ε>0\varepsilon > 0, then [a,b][a, b] can be divided into a finite number of sub-intervals [a,x1],[x1,x2],…,[xn,b][a, x_1], [x_1, x_2], \dots, [x_n, b] in each of which the oscillation of ϕ\phi is less than ε\varepsilon.

Theorem One length that works everywhere Hardy: Theorem II, §107

If ϕ\phi is continuous on [a,b][a, b] and ε>0\varepsilon > 0, there is a δ>0\delta > 0 such that, if [a,b][a, b] is divided in any manner into sub-intervals of length less than δ\delta, the oscillation of ϕ\phi in each of them is less than ε\varepsilon.

Theorem II says: ∣ϕ(x)−ϕ(x′)∣<ε|\phi(x) - \phi(x')| < \varepsilon whenever x,x′x, x' are in [a,b][a, b] and ∣x−x′∣<δ|x - x'| < \delta, with one δ\delta serving every point. In ordinary continuity δ\delta may depend on the point.

Explore Uniformly small oscillation

Try this. With the continuous function, increase the number of pieces nn until every piece is green, for ε=0.25\varepsilon = 0.25; then halve ε\varepsilon. Repeat with xsin⁡(1/x)x\sin(1/x), which is continuous on [−1,1][-1, 1]. Then choose sin⁡(1/x)\sin(1/x), and also x−[x]x - [x].

ϕ(x)=sin⁡2x+0.3x\phi(x) = \sin 2x + 0.3x on [0,4][0, 4]
xy
  • y = φ(x)
  • piece with oscillation < ε
  • piece with oscillation ≥ ε
With n=4n = 4 pieces of length 11, the largest oscillation is 1.571.57. 4 pieces still oscillate by ε or more. Increase n.
What to notice

For both continuous functions a large enough nn makes every piece green, for any ε\varepsilon: that is Theorem II (and xsin⁡(1/x)x\sin(1/x) needs more pieces near 00, where it wiggles, but only finitely many). For sin⁡(1/x)\sin(1/x) the pieces touching 00 always have oscillation 22, and for x−[x]x - [x] the pieces ending at an integer always have oscillation 11: at a discontinuity no partition helps.

Proofs

Hardy gives two proofs of the lemma. Here is the first, which uses Weierstrass’s theorem on points of accumulation (§19).

Proof A uniform window length (Hardy’s first proof)
Strategy

Argue by contradiction. If no ll works, there are bad windows of every length 2−n2^{-n}. Their left ends have a point of accumulation cc. But cc lies inside an interval of II, which must then contain the small bad windows near cc.

Suppose no such ll exists

Then for each nn there is a sub-interval [cn,cn+2−n][c_n, c_n + 2^{-n}] of [a,b][a, b] that is contained in no interval of II.

Take a point of accumulation cc of the cnc_n

The numbers cnc_n lie in [a,b][a, b], so by Weierstrass’s theorem they have a point of accumulation cc in [a,b][a, b] (if only finitely many distinct values occur, one of them occurs for infinitely many nn; take that one as cc).

cc lies well inside some interval of II

If a<c<ba < c < b, by hypothesis 1 there is an interval [x′,x′′][x', x''] of II with x′<c<x′′x' < c < x''. If c=ac = a, hypothesis 2 gives an interval [a,x′′][a, x''] of II with c<x′′c < x'' (and the cnc_n are ≥a\ge a); similarly if c=bc = b. Let h>0h > 0 be the distance from cc to the nearer end that matters.

A bad window falls inside it

Choose nn with 2−n<12h2^{-n} < \tfrac12 h and ∣cn−c∣<12h|c_n - c| < \tfrac12 h (possible for infinitely many nn, since cc is a point of accumulation). Then [cn,cn+2−n][c_n, c_n + 2^{-n}] lies inside that interval of II, contrary to its choice. So ll exists.

The second proof is the one the covering game draws: ρ(x)\rho(x) is continuous because shifting xx by ∣x−x′∣|x - x'| shrinks the best window by at most ∣x−x′∣|x - x'|, so ∣ρ(x)−ρ(x′)∣≤∣x−x′∣|\rho(x) - \rho(x')| \le |x - x'|; its minimum mm on a closed sub-interval is positive; and any l<2ml < 2m works there. (Hardy says “less than 2m2m”, not “2m2m”, because the largest window round a point need not itself fit.) A third proof is by bisection: if [a,b][a, b] could not be covered by finitely many intervals of II, neither could one of its halves, nor one of that half’s halves, and so on; the nested halves close on a point cc, which lies inside one interval of II, and that one interval covers all the small halves near cc.

Explore Heine–Borel by bisection

Try this. Press “Halve” repeatedly. Each orange piece is not yet inside a single interval of the cover δ(x)=0.05+0.2x\delta(x) = 0.05 + 0.2x and is halved again; green pieces are done.

Bisection and the cover δ(x)=0.05+0.2x\delta(x)=0.05+0.2x
00.51
  • piece inside one interval of the cover
  • piece still to be halved
Halving 0 times: 0 of 1 pieces lie inside a single interval of the cover. Halve the remaining pieces.
What to notice

After a few rounds every piece is green: finitely many pieces, each inside one interval, so finitely many intervals cover [0,1][0, 1]. The process must stop, because if it did not, the orange pieces would nest down to a point cc, and the interval belonging to cc would swallow them.

Proof Theorem I: finitely many pieces of small oscillation
Strategy

Collect, for every point, the intervals round it on which the oscillation is less than ε\varepsilon. They satisfy the hypotheses of Heine–Borel, so finitely many suffice; their end points cut [a,b][a, b] into the required pieces.

Each point has ε\varepsilon-intervals

If a<ξ<ba < \xi < b, continuity at ξ\xi gives h>0h > 0 with ∣ϕ(x)−ϕ(ξ)∣<13ε|\phi(x) - \phi(\xi)| < \tfrac13\varepsilon on [ξ−h,ξ+h][\xi - h, \xi + h]. All values there lie within 13ε\tfrac13\varepsilon of ϕ(ξ)\phi(\xi), so the oscillation is at most 23ε<ε\tfrac23\varepsilon < \varepsilon. At aa and bb use one-sided intervals [a,a+h][a, a + h] and [b−h,b][b - h, b].

Apply Heine–Borel

The set of all these intervals satisfies hypotheses 1 and 2, so finitely many of them, say J1,…,JkJ_1, \dots, J_k, also cover [a,b][a, b] in that sense.

Cut at all their end points

The end points of J1,…,JkJ_1, \dots, J_k that lie in [a,b][a, b] divide it into finitely many pieces. Each piece lies inside one JiJ_i (it lies between consecutive end points, so no JiJ_i starts or stops within it, and some JiJ_i covers its interior), and oscillation can only decrease on a sub-interval (§104, (2)). So each piece has oscillation less than ε\varepsilon.

Proof Theorem II: one length that works everywhere
Strategy

Apply Theorem I with 12ε\tfrac12\varepsilon (strictly, with some ε1<12ε\varepsilon_1 < \tfrac12\varepsilon) and let δ\delta be the length of the shortest piece. A short interval can straddle at most two neighbouring pieces, and the oscillation over two neighbouring pieces is at most the sum of theirs.

Pieces from Theorem I

Choose ε1<12ε\varepsilon_1 < \tfrac12\varepsilon and divide [a,b][a, b] into finitely many pieces j1,…,jrj_1, \dots, j_r, each of oscillation less than ε1\varepsilon_1. Let δ\delta be the length of the shortest; δ>0\delta > 0 because there are finitely many.

A short interval meets at most two pieces

An interval of length less than δ\delta cannot contain a whole piece in its interior, so it lies within the union of at most two successive pieces ji∪ji+1j_i \cup j_{i+1}.

Add the oscillations

By property (3) of §104, the oscillation on ji∪ji+1j_i \cup j_{i+1} is at most that on jij_i plus that on ji+1j_{i+1}, which is less than 2ε1<ε2\varepsilon_1 < \varepsilon; on the short interval it is no larger.

Worked examples and exercises

Exercise ●○○ Hardy: §106, examples (i)–(ii)

(i) Divide [0,1][0, 1] into the nn closed intervals [0,1n],[1n,2n],…[0, \tfrac1n], [\tfrac1n, \tfrac2n], \dots. Do they satisfy the hypotheses of Heine–Borel? (ii) Give each ξ\xi in [0,1][0, 1] the interval [ξ−ε,ξ+ε][\xi - \varepsilon, \xi + \varepsilon] cut down to lie in [0,1][0, 1], with 0<ε<10 < \varepsilon < 1. Find a finite subset with properties 1 and 2.

Hint
In (i), where is the point 1n\tfrac1n: inside an interval, or at an end?

Solution

(i) No: the points 1n,2n,…,n−1n\tfrac1n, \tfrac2n, \dots, \tfrac{n-1}n lie at the ends of intervals and inside none, so hypothesis 1 fails. (The conclusion is trivially true, as the set is already finite; the example shows what “inside” means.)

(ii) Hardy’s choice: the intervals belonging to ε,2ε,3ε,…\varepsilon, 2\varepsilon, 3\varepsilon, \dots, namely [0,2ε],[ε,3ε],[2ε,4ε],…[0, 2\varepsilon], [\varepsilon, 3\varepsilon], [2\varepsilon, 4\varepsilon], \dots, ending with one whose right end is 11 (for the last, use the interval of the point 1−ε1 - \varepsilon, or of 11 itself, which is [1−ε,1][1 - \varepsilon, 1]). Consecutive intervals overlap by ε\varepsilon, so every interior point is inside one of them; 00 is the left end of the first and 11 the right end of the last.

Exercise An explicit δ\delta ●○○ Hardy: §107

For ϕ(x)=x2\phi(x) = x^2 on [0,2][0, 2] and a given ε>0\varepsilon > 0, find a δ\delta as in Theorem II.

Hint
∣x2−x′2∣=∣x−x′∣ ∣x+x′∣≤4∣x−x′∣|x^2 - x'^2| = |x - x'|\,|x + x'| \le 4|x - x'| on [0,2][0, 2].

Solution

On any sub-interval of [0,2][0, 2] of length less than δ\delta, two points x,x′x, x' satisfy ∣x2−x′2∣≤4∣x−x′∣<4δ|x^2 - x'^2| \le 4|x - x'| < 4\delta. So the oscillation is at most 4δ4\delta, and δ=14ε\delta = \tfrac14\varepsilon works (the oscillation is then <ε< \varepsilon, since the sup and inf over a closed sub-interval are values). The bound 44 comes from the right-hand end; on [0,1][0, 1] alone, δ=12ε\delta = \tfrac12\varepsilon would do. One δ\delta for the whole interval is possible because the slope is bounded there.

Exercise Theorem II needs a closed interval ●●○ Hardy: §107

ϕ(x)=sin⁡(1/x)\phi(x) = \sin(1/x) is continuous at every point of (0,1](0, 1]. Show that the conclusion of Theorem II fails on (0,1](0, 1] for ε=1\varepsilon = 1. Which step of the proof breaks?

Hint
Any interval (0,δ)(0, \delta) contains a point where sin⁡(1/x)=1\sin(1/x) = 1 and one where it is −1-1, closer together than δ\delta.

Solution

Given any δ>0\delta > 0, choose an integer kk with 1/(2kπ)<δ1/(2k\pi) < \delta. The points x=1/(2kπ+12π)x = 1/(2k\pi + \tfrac12\pi) and x′=1/(2kπ+32π)x' = 1/(2k\pi + \tfrac32\pi) both lie in (0,δ)(0, \delta), so they are less than δ\delta apart, and sin⁡(1/x)=1\sin(1/x) = 1, sin⁡(1/x′)=−1\sin(1/x') = -1. So a division of (0,1](0, 1] that has the interval [x′,x][x', x] (of length less than δ\delta) as one of its pieces has a piece with oscillation 2>12 > 1. As δ\delta was arbitrary, no δ\delta works.

The proof breaks at the very start: Heine–Borel needs every point of a closed interval to have an ε\varepsilon-interval, including the end point 00, but sin⁡(1/x)\sin(1/x) is not even defined at 00, and no value there would make it continuous. The ε\varepsilon-intervals of points near 00 shrink to nothing.

Exercise ●●● Hardy: §106, example (iii)

Give each rational p/qp/q in [0,1][0, 1] (lowest terms, q≥1q \ge 1) the interval (p/q−ε/q3, p/q+ε/q3)(p/q - \varepsilon/q^3,\ p/q + \varepsilon/q^3), cut down to [0,1][0, 1]. Show that if ε≤14\varepsilon \le \tfrac14 some points of [0,1][0, 1] lie in none of these intervals. You may use the fact that finitely many open intervals covering [0,1][0, 1] have total length greater than 11.

Hint
Suppose every point is inside some interval. Use Heine–Borel to pass to finitely many intervals, and bound their total length.
Hint
For q≥2q \ge 2 there are at most q−1q - 1 fractions, each with an interval of length 2ε/q32\varepsilon/q^3; and ∑q≥2(q−1)/q3<∑q≥21/q2<∑q≥21/(q(q−1))=1\sum_{q\ge2}(q-1)/q^3 < \sum_{q \ge 2} 1/q^2 < \sum_{q\ge2} 1/(q(q-1)) = 1.

Solution

Suppose every point of [0,1][0, 1] lies inside (or, for 00 and 11, at the appropriate end of) one of the intervals. By Heine–Borel finitely many of them already do, and these have total length greater than 11.

But the total length of any finite collection is at most the sum over all denominators. For q=1q = 1 there are two intervals (at 00 and 11), each cut to length ε\varepsilon: total 2ε2\varepsilon. For q≥2q \ge 2 there are at most q−1q - 1 intervals, of total length at most 2ε(q−1)/q32\varepsilon(q - 1)/q^3. So any finite collection has total length less than

2ε(1+∑q≥2q−1q3)<2ε(1+1)=4ε≤1,2\varepsilon\Bigl(1 + \sum_{q \ge 2}\frac{q - 1}{q^3}\Bigr) < 2\varepsilon(1 + 1) = 4\varepsilon \le 1,

using q−1q3<1q2<1q(q−1)\frac{q-1}{q^3} < \frac{1}{q^2} < \frac{1}{q(q - 1)} and ∑q=2N1q(q−1)=1−1N<1\sum_{q=2}^{N}\frac{1}{q(q-1)} = 1 - \frac1N < 1. This contradicts “greater than 11”. Hence some points are covered by no interval. (Those points are irrational numbers that are badly approximable by fractions.) Hardy stresses that the shortcut “the total length is less than 11, so they cannot cover” is exactly what needs Heine–Borel when there are infinitely many intervals.

Check yourself

Check yourself Choose one

Which family of open intervals covers its set but has no finite sub-family that does?

Why

Finitely many of the intervals (1n,1)(\tfrac1n, 1) have a largest nn, say NN, and miss (0,1N](0, \tfrac1N]. Heine–Borel does not apply because (0,1)(0, 1) is not closed. In the other options the set is closed and every radius is at least 0.010.01, so finitely many intervals suffice, as the theorem promises.

Check yourself True or false?

For ϕ(x)=x2\phi(x) = x^2 on the whole line, there is one δ>0\delta > 0 such that ∣x2−x′2∣<1|x^2 - x'^2| < 1 whenever ∣x−x′∣<δ|x - x'| < \delta.

Why

False. Take x′=x+12δx' = x + \tfrac12\delta: then x′2−x2=xδ+14δ2>1x'^2 - x^2 = x\delta + \tfrac14\delta^2 > 1 once x>1/δx > 1/\delta. The line is not a closed bounded interval, and the slope of x2x^2 is unbounded on it. On any [a,b][a, b], Theorem II does give such a δ\delta.

Check yourself Choose one

What is the difference between continuity on [a,b][a, b] and the conclusion of Theorem II?

Why

Continuity: for each ξ\xi and ε\varepsilon there is δ(ξ,ε)\delta(\xi, \varepsilon). Theorem II: for each ε\varepsilon there is δ(ε)\delta(\varepsilon) working at all points at once. On a closed interval the first implies the second; that is the content of the theorem, and it needs Heine–Borel.

Hardy’s language

It is impossible, without the use of this or some similar theorem, to prove that a function continuous throughout an interval necessarily possesses an integral over that interval.

— G. H. Hardy, §107

Summary

Summary
  • Heine–Borel: if every point of [a,b][a, b] lies inside some interval of a family (end points at an end), finitely many of the intervals do the same job. The interval must be closed.
  • Key lemma: one window length l>0l > 0 fits inside some interval of the family wherever it is placed (Hardy proves it by Weierstrass’s theorem, or by minimising the continuous function ρ\rho).
  • For ϕ\phi continuous on [a,b][a, b]: finitely many pieces of oscillation <ε< \varepsilon (Theorem I), and even one length δ\delta such that every piece shorter than δ\delta has oscillation <ε< \varepsilon (Theorem II, uniform continuity).
  • Both fail off closed intervals (sin⁡(1/x)\sin(1/x) on (0,1](0, 1], x2x^2 on the whole line) and at discontinuities.