Dedekind cuts

Paul Taylor

Many of the accounts of Dedekind cuts that I have seen, including the original one, wave their hands in the most blatant way, especially with regard to multiplication. Nevertheless, several deep and powerful ideas have been incorporated into this theory in the course of 150 years. This posting is a summary of them, and was written in the hope that somebody might tell me who first discovered each of them. However, whilst I am always glad to receive mathematical, historical and philosophical comments from my colleagues, I would respectfully ask, on this occasion in particular, that they first check them against the actual papers that they cite, and against the bibliography of my paper with Andrej Bauer:

The Dedekind Reals in Abstract Stone Duality

This paper originally appeared in the proceedings of Computability and Complexity in Analysis, held in Kyoto in August 2005. It has now been accepted for a journal, although we are still tightening the proofs, narrative and bibliography.

Abstract Stone Duality is a new axiomatisation of recursive general topology without set theory. In it, the Dedekind reals satisfy the Heine–Borel property (“finite open subcover” compactness of [0,1]), whereas this fails in traditional recursive analysis based on set theory. How ASD achieves this is, of course, explained in the paper, but I also gave a summary in my posting to the categories mailing list on 18 August 2007.

However, this posting is about different issues, and will be expressed in traditional (set-theoretic) language. Where I have dD here, in ASD it is written δ d.

1  Definition of a Dedekind cut

Dedekind originally defined a cut (D,U) as a partition of the rationals. This means that there are two representations of each rational number q as a cut, namely ({ddq},{uu > q}) and ({dd < q},{uuq}). This is messy, even in the classical situation, so our definition omits q in this case. Who first did this?

The letters D and U stand for “down” and “up”, for which they are conveniently positioned in the alphabet. For the moment, all lower case variables range over ℚ.

We want to say, in a mathematically sensitive way, that D and U have no endpoints:

d ∈ D  ⇐⇒ ∃ e. (d < e) ∧ e ∈ D      and     u ∈ U  ⇐⇒ ∃ t. (t < u) ∧ t ∈ U.

This property is called roundedness, and I learned it in the context of continuous lattices. Where did it come from originally?

D and U also need to be bounded and disjoint:

∃ dd ∈ D,     ∃ uu ∈ U      and     d ∈ D ∧ u ∈ U  ⇒  d < u,

although there are other ways of expressing disjointness.

The final condition is the most interesting one. We want to say that D and U account for the whole line, apart from at most one point, or “almost touch”. There is an order-theoretic way to say this,

d < u  ⇒  d ∈ D   ∨  u ∈ U

and an arithmetical one,

є > 0  ⇒  ∃ d ud ∈ D ∧ u ∈ U ∧ ud < є. 

These properties are called locatedness. Who formulated them, particularly the first one? Was is Brouwer, perhaps?

I shall return to the difference between these two notions of locatedness, but first examine some useful subsystems.

2  One-sided reals

Although Andrej and I only set out to construct the real line itself, we found that both the construction and the applications forced us to consider other structures first.

Consider D on its own, satisfying just the roundedness condition. Classically, D = {dd < a} where a ≡ supD ∈ℝ∪{− ∞,+ ∞}. The collection of all such D or a forms a continuous lattice, which naturally carries the Scott topology.

As the set D grows, so does the extended real number a. Customarily, we use a temporal metaphor for the reals, and a vertical one for lattices, so I call D an ascending real, although it is called (I think) lower elsewhere. Similarly, U is a descending or upper real.

There is a bijection between rounded lower subsets D of the rationals and of the reals, and this respects the other properties, so I shall switch between the two without comment. This bijection is essentially the idempotence of the cut construction in Dedekind's paper.

An (extended) real-valued function f is called lower semicontinuous if f−1(a,+ ∞] is open for all a. I have checked the handedness of this definition with two textbooks and two websites, but who first used it?

A function is lower semicontinuous iff it is continuous in the generic sense as a function to the ascending reals with the Scott topology.

3  Constructive least upper bound

Constructively and computationally, it is not always possible to form supD.

The supremum, if it is a real number a, is represented by a Dedekind cut ({dd < a},{uu > a}). But the lower half of this must be the original D. So, I'm saying that an ascending real D need not have a descending partner U for which (D,U) is a cut.

If, as is essentially the case in ASD, D and U are recursively enumerable sets of rationals, an RE set D need not have an RE complement U.

Let S be any set of real numbers, and suppose that a = supS exists as a real number. Then, as ℝ is totally ordered, for any x < y, either x < a or a < y, and in the latter case, s < y for all sS.

So, if supS exists then, for any x < y, either x < s for some sS or s < y for all sS. This is known as the constructive least upper bound principle.

Douglas Bridges told me that Errett Bishop knew it, and Bas Spitters says that similar ideas occur in Brouwer's work, but cannot pinpoint the statement. Did Brouwer formulate it? If not, who did?

A subset S of a metric space X is called located if the distance d(x,S) ≡ inf{d(x,y) ∣ yS} exists, for all points xX. Who formulated this idea? Classically, using infima of distances like this goes back to Hausdorff.

Is ASD, using the same logical principle, we find that the supremum of any compact subspace of ℝ exists, but it is a descending real number. Similarly, the supremum of an overt subspace is an ascending real number. If the subspace is compact, overt and nonempty, it has a maximum.

4  Intervals

If D and U are rounded and disjoint, ℝ∖(DU) is classically a closed interval, [d,u], where d ≡ supD and u ≡ infU; these are ± ∞ unless D and U are bounded. The systematic study of intervals and their arithmetic was begun by Ramon Moore.

I claim that intervals should be represented by the sets D and U, and not by the numbers d and u or the closed subset [d,u].

At any stage in the computation of a real number, we have so far demonstrated that some numbers are lower bounds, and some are upper ones. These (and their lower and upper closures) form sets D and U.

For example, Archimedes measured the area of a circle by inscribing and circumscribing polygons, and Riemann integration is defined in the same way for general continuous functions.

When dealing with sets of lower and upper bounds like this, it is usually automatic that D and U be rounded and disjoint. Sometimes boundedness may be in doubt, but usually the most difficult part is to demonstrate locatedness. Intervals (possibly infinite ones) separate the easy parts of a proof or computation from the difficult ones.

In the case of Riemann integration of a continuous function on a closed interval, locatedness follows from uniform continuity. It is then much easier to demostrate the linearity properties of the integral using Dedekind cuts that with limits of Cauchy sequences.

Similarly, it is possible to define the derivative of any continuous function, as an interval. This is bounded iff the function is Lipschitz and located iff it is differentiable in the standard sense.

5  Interval arithmetic

Ramon Moore defined the arithmetic operations on intervals as follows:

[d,u] ⊕ [e,t][d+e,u+t
 ⊖ [d,u][−u,−d
[d,u] ⊗ [e,t][min(d e,d t,u e,u t), max(d e,d t,u e,u t)]
   [d,u]−1[u−1,d−1]         if 0 ∉ [d,u], so 0 ∈ DU
 [−∞,+∞].         if 0 ∈ [d,u]

The formula for multiplication is complicated by the need to consider all possible combinations of signs.

However, we have just said that supD and infU need not exist as real numbers, so we have more work to do to define the arithmetic operations for them.

This extension can be defined, using methods from continuous lattices, so long as each operation ⊛ is rounded in the sense that

[d,u] ⊛ [e,t] ⋐ [a,z]   ⇐⇒ ∃ duet'. [d,u] ⋐ [d',u'] ∧ [e,t] ⋐ [e',t'] ∧ [d',u']⊛[e',t']⋐[a',z'], 

where [d,u] ⋐ [d',u'] means d'< du< u'. This is easy to prove for addition, but rather messy for multiplication. It's neither difficult nor deep, but the property is important, so is it stated anywhere?

6  Completeness of interval computation

Standard interval analysis evaluates a function with an interval as its argument by replacing the usual arithmetic operations with Moore's generalisation to intervals. This overestimates the range of a function, often by a long way.

However, we may compute the set-theoretic image of an interval [d,u] to within any є > 0 by sub-dividing [d,u] into sufficiently but finitely many parts, applying the function Moore-wise to each part, and forming the union. Where is this theorem (first) proved in the interval literature? This result is the basis of the construction of ℝ and the proof that [d,u] is compact in ASD.

7  Back-to-front intervals

Bizarrely, it is also possible to underestimate the image by considering back-to-front intervals. This was first done by E.W. Kaucher, but can anyone give me his paper or tell me his first name? It is also used to construct the existential quantifier in our paper.

It does not make sense to represent Kaucher intervals as closed subspaces: we should use D and U instead, although they now overlap.

8  Dedekind completeness and the Archimedean axiom

A total order X is Dedekind complete if every pair of subsets D, UX that satisfies the axioms for a cut is of the form

D = {d ∣ d < a}     and     U = {u ∣ u > a}

for some unique aX.

Dedekind cuts of the rationals form a Dedekind complete field, ie a cut composed of real numbers defines another real number. In other words, Dedekind's construction is idempotent.

ℚ and ℝ are Archimedean:

є > 0 ⇒ ∃ n:ℕ. (n−1) є < x < (n+1) є. 

Classically, ℝ is the only Dedekind complete ordered field, because Dedekind completeness implies the Archimdean principle.

Proof: the sets

D ≡ {d ∣ ∃ n:ℕ. d < n}     and     U ≡ {u ∣ ∀ n:ℕ. u > n}

are rounded, disjoint and order-located. They are bounded (and so form a cut) iff U is nonempty. By Dedekind completeness, the cut (D,U) corresponds to an element ω of the field. This should be the only element that lies between D and U, but ω−1 does so too. Hence U must be empty.

Assuming the Archimedean principle, the two definitions of locatedness are equivalent. Who first proved these results?

John Conway's system of surreal numbers tries to extend the idea of Dedekind completeness to infinities and infinitessimals. The whole system is a proper class, as is U in this proof, and so (D,U) is not a legitimate cut: Conway calls it a gap.

I conjecture (and John Conway considered this plausible when I put it to him) that there is a version of his number system in which sets are replaced by RE sets, and every RE Dedekind cut represents a number, but there are also infinities and infinitessimals.

Andrej and I certainly don't construct such a thing in our paper. However, we have isolated the use of the Archimedean principle, in the hope that someone in future might eliminate it.

9  Locatedness of the arithmetic operations

The most difficult thing to prove is always locatedness.

In computational terms, when we are asked to compute the result of an operation to a certain precision, we have to ask for its arguments to the appropriate precision. For example, if we need x+y within є, it is enough to know x and y within ½є.

However, for multiplication things are more complicated: we need to know x to within є / |y|. This is also formulated and proved constructively in our paper.

Finally, we also have a way of defining inverse functions (reciprocals, roots, powers and logarithms) that the referee considered “rather elegant”.

This document was translated from LATEX by HEVEA.