Chapter I · Real variables

Sections of the reals and Dedekind's theorem

§17 35 min

Cutting the real numbers produces nothing new; every section of the reals has a real number at the join.

Why this matters

Sections of the rationals turned out to be a richer system than the rationals: they gave us 2\sqrt2, π\pi and every other real number. So it is natural to ask whether sections of the reals give something richer still, numbers beyond the real numbers. §17 answers no. Every section of the real numbers is “at” a real number. This property, completeness, is what the rationals lack and the reals possess, and it is the property every existence theorem of analysis (limits of monotone sequences, the intermediate value theorem, maxima of continuous functions) ultimately rests on.

The idea

A section of the reals is defined just like a section of the rationals: two properties, every real number has exactly one, both occur, and everything with the first is less than everything with the second. Among the rationals there were three possibilities at the join: the lower class has a greatest member, or the upper class has a least member, or neither (a gap). Among the reals Hardy shows that the third possibility, the gap, never happens.

Explore Look for a gap among the reals

Try this. Start with “x ≤ √2 | x > √2” among the rationals and zoom in on the join. Switch to the reals. Then try “x < √2 | x ≥ √2”, the cubic, and 1/21/2, each among the rationals and among the reals. Last, try “x < √2 | x > √2” among the reals.

L:x≤2L: x \le \sqrt2, R:x>2R: x > \sqrt2, among the rationals
LR1.01.11.21.31.41.51.61.71.81.9no rational at √2
width of view: 1
  • lower class L
  • upper class R
  • the join
Divide
Among the rationals there is a gap: 2\sqrt2 is not rational, so the rational members of L have no greatest element and those of R no least. Zoom in as far as you like; the hollow mark never fills. (For the rationals the pairs “x≤2x \le \sqrt2” and “x<2x < \sqrt2” give the same section: they differ only at an irrational point.)
What to notice

Among the rationals an irrational join is a hole: neither class has an end member, and the two √2 pairs give the same section. Among the reals the join is always a real number, which sits in exactly one class: L has a greatest member or R a least one. The last pair is not a section of the reals at all, because √2 has neither property, although it is a perfectly good section of the rationals.

Why should the reals behave better? Because the only candidates for a gap are positions on the line, and every position is already described by a section of the rationals, which is a real number. Hardy’s proof makes that precise: look at the rational members of the two classes. They form a section of the rationals, hence a real number α\alpha, and α\alpha must be the end member of whichever class it falls in.

Precise statements

Definition Section of the real numbers Hardy: §17

A section of the real numbers is a division of all real numbers into two classes LL and RR such that (i) every real number belongs to exactly one class; (ii) both classes are non-empty; (iii) every member of LL is less than every member of RR.

Theorem Dedekind's theorem Hardy: §17

If the real numbers are divided into two classes LL and RR satisfying (i), (ii) and (iii), then there is a real number α\alpha such that every number less than α\alpha belongs to LL and every number greater than α\alpha belongs to RR. The number α\alpha itself may belong to either class: it is either the greatest member of LL or the least member of RR.

Corollary Sections of an interval Hardy: §17

The same holds for a section of the numbers of an interval β≤x≤γ\beta \le x \le \gamma, and then β≤α≤γ\beta \le \alpha \le \gamma.

Proofs

Proof Dedekind's theorem
Strategy

First show that LL and RR cannot both have end members. Then restrict the classes to the rationals: the rational members form a section of the rationals, which by definition is a real number α\alpha. Show that α\alpha is an end member of LL or of RR.

Not both an end in LL and an end in RR

If ll were greatest in LL and rr least in RR, then l<rl < r and 12(l+r)\tfrac12(l + r) would be a real number greater than every member of LL and less than every member of RR, so in neither class. That contradicts (i).

The rational members form a section of the rationals

Let L1L_1 and R1R_1 be the rational members of LL and RR. Every rational is in exactly one of them, and every member of L1L_1 is below every member of R1R_1. Both are non-empty: if l∈Ll \in L, any rational r<lr < l cannot be in RR (by (iii) it would exceed ll), so r∈L1r \in L_1; similarly any rational above a member of RR is in R1R_1.

Case 1: L1L_1 has a greatest member aa

Then aa is also the greatest member of LL. Otherwise some β∈L\beta \in L has β>a\beta > a, and a rational rr with a<r<βa < r < \beta exists by density. Being less than the member β\beta of LL, rr is in LL (it cannot be in RR by (iii)), so r∈L1r \in L_1 and r>ar > a, a contradiction. So α=a\alpha = a works.

Case 2: L1L_1 has no greatest member

Then (L1,R1)(L_1, R_1) is a section of the rationals whose lower class has no greatest member, that is, a real number α\alpha (§8). The number α\alpha belongs to LL or to RR.

α\alpha is an end member of its class

Suppose α∈L\alpha \in L but some β∈L\beta \in L exceeds it. Choose rationals r,sr, s with α<r<s<β\alpha < r < s < \beta. Both are below β\beta, so they lie in LL, hence in L1L_1. But every member of L1L_1 is less than α\alpha (the lower class of the section α\alpha consists of rationals below α\alpha), and r>αr > \alpha: contradiction. So α\alpha is the greatest member of LL. In the same way, if α∈R\alpha \in R it is the least member of RR.

Conclude

In every case there is a real α\alpha that is the greatest member of LL or the least member of RR. Say α\alpha is greatest in LL. A number x>αx > \alpha cannot be in LL, so it is in RR; a number x<αx < \alpha cannot be in RR, since it would then be smaller than the member α\alpha of LL, against (iii); so it is in LL. The case where α\alpha is least in RR is the same with the classes exchanged.

The proof uses the definition of real numbers in exactly one place, Case 2: the section of the rationals cut out by LL and RR is a real number by definition. That is where completeness comes from. For the rationals alone, Case 2 would leave us with a gap and nothing to fill it.

Sections of an interval

In applications one often cuts only the numbers of an interval [β,γ][\beta, \gamma]. Such a section becomes a section of all reals by adding every number below β\beta to LL and every number above γ\gamma to RR; the theorem then gives a number α\alpha, and β≤α≤γ\beta \le \alpha \le \gamma because β∈L\beta \in L (or is the least member of RR) and γ∈R\gamma \in R (or is the greatest member of LL). This is the form used for Weierstrass’s theorem in the last lesson of the chapter.

Worked examples and exercises

Exercise Which class has the end? ●○○ Hardy: §17

For each section of the reals, find α\alpha and say whether it is the greatest member of LL or the least member of RR: (a) L={x:x≤3}L = \{x : x \le 3\}; (b) L={x:x2<2 or x<0}L = \{x : x^2 < 2 \text{ or } x < 0\}; (c) L={x:x<0 or x3≤3x+8}L = \{x : x < 0 \text{ or } x^3 \le 3x + 8\}.

Hint
In each case decide where the “border” number itself goes.

Solution

(a) α=3\alpha = 3, the greatest member of LL. (b) α=2\alpha = \sqrt2: since (2)2=2(\sqrt2)^2 = 2 is not less than 22, 2∈R\sqrt2 \in R, and it is the least member of RR. (c) α=z\alpha = z, the real root of z3=3z+8z^3 = 3z + 8, which satisfies z3≤3z+8z^3 \le 3z + 8, so it is the greatest member of LL. (For 0≤x0 \le x the inequality x3≤3x+8x^3 \le 3x + 8 holds exactly up to zz, by the monotonicity shown in §15.)

Exercise Properties that cut the rationals but not the reals ●○○ Hardy: §17

Show that the pair "x2<2x^2 < 2", "x2>2x^2 > 2" defines a section of the positive rationals but not of the positive reals.

Hint
Which positive real number has neither property?

Solution

No positive rational has x2=2x^2 = 2, so every positive rational has exactly one of the properties; the other conditions were checked in §4. Among the positive reals, 2\sqrt2 has (2)2=2(\sqrt2)^2 = 2 and so has neither property: condition (i) fails. To get a section of the reals, 2\sqrt2 must be put into one of the classes, and it then becomes its end member, as Dedekind’s theorem predicts.

Exercise Least upper bounds ●●○ Hardy: §17

Let SS be a non-empty set of real numbers with an upper bound (a number MM with x≤Mx \le M for all x∈Sx \in S). Use Dedekind’s theorem to show that SS has a least upper bound.

Hint
Put xx in RR if xx is an upper bound of SS, and in LL otherwise.
Hint
Check (i)–(iii), apply the theorem, and ask whether α\alpha can be in LL.

Solution

Let RR be the set of upper bounds of SS and LL the rest. (i) holds by construction. (ii) M∈RM \in R; and if s∈Ss \in S then s−1s - 1 is not an upper bound, so L≠∅L \ne \emptyset. (iii) If x∈Lx \in L and y∈Ry \in R, then some s∈Ss \in S has s>xs > x, while y≥sy \ge s; so x<yx < y. Dedekind’s theorem gives α\alpha. If α\alpha were in LL, some s∈Ss \in S would exceed α\alpha; then 12(α+s)\tfrac12(\alpha + s) is greater than α\alpha and still not an upper bound, so it is in LL, contradicting that α\alpha is greatest in LL. So α∈R\alpha \in R is the least member of RR: the least upper bound. This is the form of completeness used in most modern books.

Exercise Completeness fails for the rationals ●●○ Hardy: §17

Show that the set of rationals xx with x2<2x^2 < 2 has rational upper bounds but no least rational upper bound.

Hint
An upper bound u>0u > 0 must have u2≥2u^2 \ge 2, hence u2>2u^2 > 2. Use the map x↦(2x+2)/(x+2)x \mapsto (2x + 2)/(x + 2) from the earlier lesson.

Solution

22 is a rational upper bound. Let uu be any rational upper bound; then u>1>0u > 1 > 0, and u2<2u^2 < 2 is impossible (then uu would be in the set, and the larger member (2u+2)/(u+2)(2u+2)/(u+2) would exceed it), while u2=2u^2 = 2 is impossible for rational uu. So u2>2u^2 > 2. Then u′=(2u+2)/(u+2)u' = (2u + 2)/(u + 2) is rational, smaller than uu, and still has u′2>2u'^2 > 2, so it is also an upper bound (anything with square below 22 is below u′u'). Hence no rational upper bound is least.

Check yourself

Check yourself Choose one

For a section of the real numbers, which situation is impossible?

Why

That is exactly the “gap” case, which Dedekind’s theorem rules out for the reals (it is possible for the rationals). Both of the first two cases occur, and the join may well be irrational, as with 2\sqrt2; it is then a real number in one of the classes.

Check yourself True or false?

Sections of the real numbers define a new, larger kind of number, just as sections of the rationals did.

Why

False. That is the whole point of the theorem: every section of the reals corresponds to a real number at its join. Hardy expresses this by saying that the continuum is “closed”, or has a completeness the rationals lack.

Check yourself Choose one

Where in the proof of Dedekind’s theorem is the definition of a real number (as a section of the rationals) used?

Why

When the rational members leave a “gap”, that gap is a real number by definition. For the rationals the same proof breaks down at exactly this step, and the theorem is false for them.

Hardy’s language

the continuum has a kind of completeness which the aggregate of the rational numbers lacked, a completeness which is expressed in technical language by saying that the continuum is closed.

— G. H. Hardy, §17

Summary

Summary
  • A section of the reals splits all real numbers into non-empty classes LL below RR.
  • Dedekind’s theorem: there is a real number α\alpha at the join, either the greatest member of LL or the least member of RR, never neither and never both.
  • Proof: the rational members of LL and RR form a section of the rationals, which is a real number by definition, and it is the end member of its class.
  • Equivalent form: every non-empty set of reals with an upper bound has a least upper bound. Both fail for the rationals.