Chapter I · Real variables

Weierstrass's theorem

§19 35 min

Every infinite set inside a bounded interval has a point of accumulation, proved from Dedekind's theorem and, alternatively, by repeated bisection.

Why this matters

If infinitely many points are squeezed into a bounded interval, they must crowd together somewhere. That sounds obvious, but it is false for the rationals: the decimals 1,1.4,1.41,1.414,…1, 1.4, 1.41, 1.414, \dots crowd only towards 2\sqrt2, which is not rational. So the statement is really a property of the real numbers, a consequence of their completeness. Hardy proves it in §19 directly from Dedekind’s theorem, and it is the tool that later produces limits of bounded sequences (§71), maxima of continuous functions and much else.

The idea

Hardy’s proof is a clever use of a section. Given the set SS in an interval, sort the points PP of the line by a single question: are there infinitely many points of SS to the right of PP? Points far to the left say yes; points to the right of the whole interval say no; and once the answer is no it stays no as you move right. That is a section of the reals, so Dedekind’s theorem puts a number ξ\xi at the join. Just left of ξ\xi there are infinitely many points of SS further right; just right of ξ\xi only finitely many. So infinitely many points of SS sit within any distance of ξ\xi.

Explore Hardy's section, point by point

Try this. With S={1/n}S = \{1/n\}, move PP across the line and read how many points of SS lie to its right. Find the join. Then try (−1)n(1+1/n)(-1)^n(1 + 1/n), 1/m+1/n1/m + 1/n and sin⁡n\sin n, and finally the two sets that break the hypotheses: the finite set and the positive integers.

S={1/n}S = \{1/n\}: put P in L if infinitely many points of S lie to the right of P, otherwise in R
LR−0.200.20.40.60.81.01.2Pξ = 0
  • class L
  • class R
  • your point P
  • the join ξ
To the right of P=0.5P = 0.5 there is only 1 point of S, so P is in R. Every point left of 0 is in L and every point right of it in R; ξ = 0 itself is in L. For any δ, ξ − δ is in L (infinitely many points beyond it) and ξ + δ is in R (finitely many beyond it), so (ξ−δ, ξ+δ)(\xi-\delta,\ \xi+\delta) holds infinitely many points: ξ is a point of accumulation.
What to notice

The answer flips from “infinitely many” to “finitely many” at exactly one place, the join ξ\xi, which is a point of accumulation. It is always the right-most one: for (−1)n(1+1/n)(-1)^n(1 + 1/n) it finds 11, not −1-1. The join may lie in either class: 00 is in LL for {1/n}\{1/n\} (all the 1/n1/n are to its right), but 11 is in RR for {sin⁡n}\{\sin n\}. For a finite set LL is empty, and for the unbounded integers RR is empty: no section, no conclusion.

Precise statements

Theorem Weierstrass's theorem Hardy: Theorem, §19

If a set SS contains infinitely many points and lies entirely in an interval α≤x≤β\alpha \le x \le \beta, then at least one point of the interval is a point of accumulation of SS.

Corollary The greatest point of accumulation

The point ξ\xi produced by Hardy’s proof is the greatest point of accumulation of SS.

Both hypotheses are needed. The positive integers are infinite but not in any bounded interval, and have no point of accumulation; a finite set lies in an interval but has none either.

Proofs

Proof Weierstrass's theorem (Hardy's proof)
Strategy

Define a section of the reals by “infinitely many points of SS lie to the right of PP”. Check Dedekind’s three conditions using the hypotheses, take the number ξ\xi at the join, and show every interval round ξ\xi contains infinitely many points of SS.

Define the classes

Put a real number PP in LL if infinitely many points of SS are greater than PP, and in RR otherwise. Condition (i) holds by construction.

Both classes are non-empty

α∈L\alpha \in L: all points of SS are ≥α\ge \alpha, and at most one of them equals α\alpha, so infinitely many are greater. β∈R\beta \in R: no point of SS exceeds β\beta. Here both hypotheses are used: SS infinite gives L≠∅L \ne \emptyset, and SS bounded gives R≠∅R \ne \emptyset.

Every member of LL is below every member of RR

Let P∈LP \in L and Q∈RQ \in R. If Q≤PQ \le P, every point of SS greater than PP would be greater than QQ, so infinitely many points would exceed QQ, and QQ would be in LL. So P<QP < Q.

Apply Dedekind's theorem

There is a number ξ\xi with every number below ξ\xi in LL and every number above ξ\xi in RR. Since α∈L\alpha \in L and β∈R\beta \in R, α≤ξ≤β\alpha \le \xi \le \beta.

Every interval round ξ\xi holds infinitely many points of SS

Let δ>0\delta > 0. Then ξ−δ∈L\xi - \delta \in L: infinitely many points of SS exceed ξ−δ\xi - \delta. And ξ+12δ∈R\xi + \tfrac12\delta \in R: only finitely many exceed ξ+12δ\xi + \tfrac12\delta. Removing those finitely many leaves infinitely many points of SS in (ξ−δ,ξ+12δ](\xi - \delta, \xi + \tfrac12\delta], inside (ξ−δ,ξ+δ)(\xi - \delta, \xi + \delta).

Conclude

In particular every such interval contains a point of SS other than ξ\xi, so ξ\xi is a point of accumulation of SS, and it lies in the interval [α,β][\alpha, \beta].

There is a second proof, which Hardy gives in §71, by repeated bisection. Halve the interval. At least one half contains infinitely many points of SS; keep it (the left one, say, if both do). Halve again, and so on. The chosen intervals are nested, each contains infinitely many points of SS, and their lengths (β−α)/2k(\beta - \alpha)/2^k shrink to 00, so they close in on a single point, which is a point of accumulation. Completeness enters at the words “close in on a single point”.

Explore The bisection proof

Try this. Press “Halve” repeatedly and watch which half is kept. Note the rule in the read-out when both halves contain infinitely many points.

Bisection for a point of accumulation of {1/m+1/n:m,n≥1}\{1/m+1/n : m, n \ge 1\}
−113
  • earlier intervals
  • current interval
  • points of the set
Step 0: [−1, 3][−1,\ 3], width 4; it contains infinitely many points of the set (1830 of the first 1830 drawn). Every interval in the chain has this property, so the point they close on is a point of accumulation.
What to notice

Every interval in the chain contains infinitely many points of the set, and the intervals shrink to a point, here 00, which is therefore a point of accumulation. Keeping the left half whenever possible finds the least point of accumulation; Hardy's section finds the greatest.

Worked examples and exercises

Exercise Hardy's ξ\xi is the greatest point of accumulation ●●○ Hardy: §19

Show that no number greater than the ξ\xi of the proof is a point of accumulation of SS.

Hint
Take η>ξ\eta > \xi and a number η′\eta' strictly between them. Which class is η′\eta' in?

Solution

Let η>ξ\eta > \xi and ξ<η′<η\xi < \eta' < \eta. Then η′∈R\eta' \in R, so only finitely many points of SS exceed η′\eta'. The interval (η′,2η−η′)(\eta', 2\eta - \eta') round η\eta (radius η−η′\eta - \eta') lies to the right of η′\eta', so it contains only finitely many points of SS. By the theorem of §18, an interval round a point of accumulation contains infinitely many points; so η\eta is not a point of accumulation. (In Chapter IV this ξ\xi reappears as the upper limit, lim sup⁡\limsup, of a bounded sequence.)

Exercise The theorem fails for the rationals ●●○ Hardy: §19

Let SS be the set of decimal truncations of 2\sqrt2: 1,1.4,1.41,1.414,…1, 1.4, 1.41, 1.414, \dots Show that SS is an infinite set of rationals in [1,2][1, 2] whose only point of accumulation is 2\sqrt2. Why does Hardy’s proof fail if we try to run it inside the rationals?

Hint
The kk-th truncation tkt_k satisfies 2−10−k<tk≤2\sqrt2 - 10^{-k} < t_k \le \sqrt2.

Solution

The truncations never decrease, and infinitely many of them are different (a digit 00 repeats a truncation, as 1.4142135623731.414213562373 is followed by 1.41421356237301.4142135623730, but the expansion of the irrational 2\sqrt2 never ends in zeros, so new values keep coming); they all lie in [1,2][1, 2], and 2−10−k<tk<2\sqrt2 - 10^{-k} < t_k < \sqrt2. Every interval round 2\sqrt2 contains all tkt_k with 10−k10^{-k} less than its radius, so 2\sqrt2 is a point of accumulation. Any η≠2\eta \ne \sqrt2 is not: if η>2\eta > \sqrt2, the interval of radius η−2\eta - \sqrt2 round η\eta contains no tkt_k; if η<2\eta < \sqrt2, only the finitely many tk≤12(η+2)t_k \le \tfrac12(\eta + \sqrt2) can be near η\eta. So within the rationals, SS has no point of accumulation. Hardy’s classes, restricted to rational PP, form a section of the rationals with a gap at 2\sqrt2; it is Dedekind’s theorem for the reals that fills the gap.

Exercise The bisection proof in detail ●●○ Hardy: §19 and §71

Write out the bisection proof: construct the nested intervals and prove that the point they close on is a point of accumulation.

Hint
Let I0=[α,β]I_0 = [\alpha, \beta] and Ik+1I_{k+1} a half of IkI_k containing infinitely many points of SS. Why does such a half exist?
Hint
The left end-points aka_k increase and are bounded by β\beta. Use Dedekind’s theorem (or the least upper bound) to get a single point.

Solution

If both halves of IkI_k contained only finitely many points of SS, so would IkI_k; so one half contains infinitely many: call it Ik+1=[ak+1,bk+1]I_{k+1} = [a_{k+1}, b_{k+1}]. Then a0≤a1≤…a_0 \le a_1 \le \dots, b0≥b1≥…b_0 \ge b_1 \ge \dots, each aj<bka_j < b_k, and bk−ak=(β−α)/2kb_k - a_k = (\beta - \alpha)/2^k. Put xx in LL if x≤akx \le a_k for some kk, in RR otherwise; this is a section (it is non-empty on both sides since a0∈La_0 \in L and b0+1∈Rb_0 + 1 \in R, and LL is closed downwards), so Dedekind’s theorem gives ξ\xi with ak≤ξ≤bka_k \le \xi \le b_k for every kk. Given δ>0\delta > 0, choose kk with (β−α)/2k<δ(\beta - \alpha)/2^k < \delta; then Ik⊆(ξ−δ,ξ+δ)I_k \subseteq (\xi - \delta, \xi + \delta), so this interval contains infinitely many points of SS, and ξ\xi is a point of accumulation.

Exercise Applying the theorem ●○○ Hardy: §19

Show that the set {sin⁡n:n=1,2,3,… }\{\sin n : n = 1, 2, 3, \dots\} has a point of accumulation, without knowing anything about where the numbers sin⁡n\sin n lie except that they are in [−1,1][-1, 1] and are all different.

Hint
Check the two hypotheses. Why are the numbers sin⁡n\sin n all different?

Solution

All sin⁡n\sin n lie in [−1,1][-1, 1]. They are all different: sin⁡m=sin⁡n\sin m = \sin n with m≠nm \ne n would require m−nm - n or m+nm + n to be a multiple of π\pi (from sin⁡m−sin⁡n=2cos⁡m+n2sin⁡m−n2\sin m - \sin n = 2\cos\tfrac{m+n}{2}\sin\tfrac{m-n}{2}), which would make π\pi rational. So the set is infinite and bounded, and Weierstrass’s theorem gives a point of accumulation. (In fact every point of [−1,1][-1, 1] is one, but that is much harder.)

Check yourself

Check yourself Choose one

In Hardy’s proof applied to S={1/n}S = \{1/n\} in [0,1][0, 1], what is ξ\xi, and which class contains it?

Why

Every 1/n1/n is greater than 00, so infinitely many points of SS lie to the right of 00: 0∈L0 \in L. Every P>0P > 0 has only finitely many 1/n1/n to its right, so is in RR. The join is 00, the greatest member of LL. Whether a point of accumulation belongs to SS is irrelevant.

Check yourself Choose one

Which set does not satisfy the hypotheses of Weierstrass’s theorem?

Why

{n2}\{n^2\} is infinite but not contained in any bounded interval, and indeed it has no point of accumulation. The other three are infinite sets in bounded intervals.

Check yourself True or false?

Weierstrass’s theorem says that a bounded infinite set has exactly one point of accumulation.

Why

False: it says at least one. {(−1)n(1+1/n)}\{(-1)^n(1 + 1/n)\} has two, {1/m+1/n}\{1/m + 1/n\} infinitely many, and the rationals in [0,1][0, 1] have every point of [0,1][0, 1].

Hardy’s language

The general theory of sets of points is of the utmost interest and importance in the higher branches of analysis; but it is for the most part too difficult to be included in a book such as this.

— G. H. Hardy, §19

Summary

Summary
  • Weierstrass: an infinite set contained in a bounded interval has at least one point of accumulation, lying in that interval.
  • Hardy’s proof: P∈LP \in L if infinitely many points of SS exceed PP; this is a section of the reals, and the number at the join is a point of accumulation (in fact the greatest).
  • Alternative proof: repeated bisection, keeping a half with infinitely many points.
  • Both hypotheses (infinite, bounded) and the completeness of the reals are essential; inside the rationals the theorem fails.