Constructive Theory of Banach algebras

We present a way to organize a constructive development of the theory of Banach algebras, inspired by works of Cohen, de Bruijn and Bishop. We illustrate this by giving elementary proofs of Wiener's result on the inverse of Fourier series and Wiener's Tauberian Theorem, in a sequel to this paper we show how this can be used in a localic, or point-free, description of the spectrum of a Banach algebra.


Introduction
The applications of the theory of Banach algebras to "concrete" theorems in analysis, such as Wiener's theorem on the inverse of Fourier series, or Wiener's Tauberian Theorem [Wie32], constitute a striking example of the power of abstract methods in mathematics. The abstract argument is short and easy to grasp, when compared to Wiener's explicit constructions. Furthermore, it is highly non constructive and uses Zorn's Lemma. As such, it is a perfect illustration of Hilbert's defense of the use of the law of excluded-middle against Brouwer [Hil26]. A natural question is whether the abstract argument does not contain, in some implicit way, an actual construction. This question has been analysed and answered by P. Cohen [Coh61]. Closely related are later works by de Bruijn and van der Meiden [dBvdM67], and by Bishop and Bridges [BB85]. Here we present a slightly different analysis of the non-constructive argument, thus providing an elementary treatment. Furthermore it allows us to reformulate some key results of Bishop and Bridges [BB85] in a localic, or point-free, setting. This will be done in a sequel to the present paper. This paper is organised as follows. The first section explains informally the main idea of the paper. The next sections, until Section 7, are presented in elementary Bishop style mathematics, in a detailed way so that they can hopefully be readily implemented in type theory using the framework presented in [OS08]. This presentation follows ideas presented in [BS99].
Notation. We use the following conventions, unless otherwise indicated: A, B for Banach algebras. a, b, c, . . . for elements of a Banach algebra. ϕ, ψ for functionals. Γ R denotes the circle with radius R, Γ is the circle with radius 1.

Analysis of the abstract reasoning
Let A be a commutative Banach algebra with unit and let MFn(A) be the compact space of non-zero multiplicative functionals A → C. Let us assume that an element f of A satisfies ϕ(f ) = 0 for all ϕ in MFn(A); and we want to show that f is then invertible in A. To obtain a contradiction, we assume that f is not invertible.
Then the ideal f is proper and hence, using Zorn's Lemma, included in a maximal ideal m. This maximal ideal is necessarily closed. Using the Gelfand-Mazur theorem, the algebra A/m is isomorphic to C and the corresponding multiplicative functional ϕ m : A → A/m is such that ϕ m (f ) = 0. This contradicts the hypothesis on f . For example, the application of this argument to A = l 1 (Z) gives Wiener's theorem on inverse of Fourier series.
Cohen [Coh61] observes that, instead of working with a maximal ideal, one can work just as well with the closure I of ideal generated by f . Moreover, an ideal contains 1 if, and only if, its closure contains 1. This follows directly from the fact that x is invertible if |1 − x| < 1. Finally, the proof of the Gelfand-Mazur Theorem actually gives a more general result: if A is a non trivial Banach algebra, the spectrum of any element u in A is non empty. A combination of these remarks eliminates the use of Zorn's Lemma.
To simplify, we consider the case where A is generated by one element u: for any a in A and any r > 0 there is a polynomial P such that |a − P (u)| < r. This covers for instance the example of the disc algebra [Coh61]. In this case the spectrum MFn(A) can be identified with the spectrum, σ(u), of u: the set of complex λ such that λ − u is not invertible. Any such element λ defines a multiplicative functional ϕ(P (u)) := P (λ) on polynomials in u. If |P (λ)| > |P (u)|, then P (λ) − P (u) is invertible and hence λ − u is invertible. Since this is not the case, |P (λ)| |P (u)|, so ϕ can be extended to a multiplicative functional ϕ(g) := g(λ). Conversely, for ϕ in MFn(A), ϕ(ϕ(u) − u) = ϕ(u) − ϕ(u) = 0 and hence ϕ(u) − u cannot be invertible, i.e. ϕ(u) ∈ σ(u). The hypothesis on f becomes f (λ) = 0 for all λ in the spectrum of u. The spectrum of u is empty in A/I. Indeed, the closure of the ideal generated by λ − u contains f (λ) − f = f (λ) modulo I which is invertible and hence λ − u is always invertible modulo I. Hence 1 belongs to I and hence f is invertible. The maximal ideal of the abstract argument has been replaced by the "big enough" ideal I. Essentially, this is the method followed by de Bruijn and van der Meiden [dBvdM67], who advocated a point-free approach to the description of the spectrum of a Banach algebra.
In this way we eliminate the use of Zorn's Lemma, but the argument is still non-constructive. Cohen shows that we can follow the proof of Gelfand-Mazur and produce an actual computation of the inverse of f . Bishop and Bridges' work is similar [BB85]. Coquand and Stolzenberg [CS91] show that in most cases, we can obtain a relatively short and explicit formula for the inverse.
In this paper we suggest a slightly different constructive argument. It is enough to have a constructive proof of the following result.
Theorem 1. If we have u in A such that for all λ the element λ − u is invertible and furthermore the inverse is uniformly bounded, then 1 = 0 in A.
Notice that, constructively, we have to state explicitly that the inverse is uniformely bounded. At first the conclusion seems purely negative, since it says that a ring is trivial, and not informative. But if we apply the Theorem to a quotient algebra A/I the conclusion shows that 1 is in I. Choosing I to be the closure of f actually builds an inverse of f , provided we work constructively. Such use of "trivial rings" occurs often in constructive algebra [Ric88,LQ08]. In the applications to Wiener's Theorems on Fourier series and on the Tauberian Theorem, we believe that our treatment is simpler than the one in [BB85]. Moreover, this theorem has a rather direct proof, for instance, it does not rely on Cauchy's formula like the treatment in [Coh61,BB85,CS91].
However, one should notice that this constructive reading does not yet have the "elegance" of the classical proofs, since we have to deal explicitely with a bound of the inverse. In the sequel to this paper, we explain how to "hide" explicit mentions to this bound, by using a localic, or pointfree, presentation of the spectrum. In this way we obtain a rather faithful constructive explanation of the classical arguments.

Integration and differentiation with values in a Banach space
We consider a vector space E over the real numbers with an upper real valued semi-norm: if a is in E then |a| is an upper real (open nonempty upper set of rationals) such that |ra| = |r||a| and |a + b| |a| + |b|. This is an important difference with Bishop's treatment, for whom the norm is a Cauchy real 1 .
The equality in E is such that Let E be a Banach space, i.e. a normed vector space which is complete: any Cauchy approximation converges. If g :

Exponential function
A (commutative) Banach algebra A is a Banach space with a ring structure, multiplication ab being such that |ab| |a||b|. In this section, we assume also that A has a unit element 1.
The fact that the norm is not a Cauchy real constructively arises if one wants to quotient a Banach space by a closed subspace which may not be located. (This will indeed occur in the applications.) Another motivation is the interpretation of these results in Banach algebra bundles which we will discuss in Section 8. In a Banach algebra bundle the function x → |ax| is upper semi continuous, but generally not continuous; see [KR94]. Bishop  Proof. We write y = 1 − x. We have x = 1 − y and y n is the inverse of 1 − y since |y| < 1.
We define the exponential e a = a n n! of an element a of A. The map x → e ax defines a function e a : R → A such that e a is differentiable, e ′ a (x) = ae a (x) and e a (0) = 1. Furthermore, we have e a (x + y) = e a (x)e a (y).
On the other hand we have, using It follows that f is differentiable and

Path integration
Let E be a Banach space over of the complex numbers. We say that f : C → E is differentiable iff there exists a (uniformly) continuous function f ′ : C → E such that for all ε > 0 there exists η > 0 such that If γ : [0, 1] → E is a differentiable function and f : C → E is a continuous function, we define γ f (z)dz to be 1 0 f (γ(t))γ ′ (t)dt. We say that γ is a loop iff γ(0) = γ(1).

Inverse function
Let A be a Banach algebra over the complex numbers with a unit element.
Lemma 5. If a is invertible, with inverse b such that |b| M , then a − u is invertible if |u| c/M and c < 1. It inverse is bounded by M/(1 − c).
Proof. The element a − u is invertible iff the element 1 − ub is. Since |u| < 1/M and |b| < M , |ub| c and hence 1 − ub is invertible by Lemma 2. The inverse is bounded by (1 − c) −1 . Hence the inverse of a − u is bounded by M/(1 − c).

It follows from this lemma that the set of invertible elements A × of A is an open subset of A.
Theorem 2. Let u be in A. If for all z in C the element z − u is invertible, with inverse f (z), and f is uniformly bounded, then 1 = 0 in A.
Notice that classically, the hypothesis that f is bounded is not needed 2 .
On the other hand, for R big enough f (z) is equal to u n /z n+1 over Γ R and so ΓR f (z)dz is equal to ΓR dz/z = 2πi. So, we have 1 = 0 in A. Before giving the applications we remark that if I is an ideal of a Banach algebra A, we can define a new Banach algebra A/I by taking the new norm to be: |a| I < r iff there exists b in I such that |a − b| < r. Then a = 0 in A/I iff a belongs to the closure of I. It follows from Lemma 2 that 1 = 0 in A/I iff 1 belongs to I.
We deduce the following method from the previous theorem: To prove that an element g of A is invertible, it is enough to find u in A such that for all z in C the element z − u is invertible of bounded inverse in A/ g . We emphasize that classically the inverse is always uniformly bounded. Constructively, there is a meta-theorem, the fan rule, that allows us to always find a bound in concrete case; see e.g. [TvD88]. In the examples below we make an effort to be explicit about the bound.

Application 1: Wiener's Theorem on Fourier series
We present Wiener's theorem on Fourier series; see e.g. [Loo53]. The Banach algebra B = l 1 (Z) is the completion of the algebra of sequences in C Z of finite support with the convolution product (a * b) n = a i b n−i and the norm |a| = |a n |. This algebra has unit δ 0 . We write u for δ 1 . Then u −1 = δ −1 and for every a, a = a n u n and B is simply an algebra of infinite series under formal multiplication.
2 Classically, this theorem implies the Gelfand-Mazur Theorem: a Banach algebra A which is also a field is isomorphic to C. Indeed, from Theorem 2, for each u in A, there exists λ such that u − λ is not invertible and this implies u = λ.
Theorem 3. If f is an element of B such that |f (λ)| ε for all λ in Γ, then f is invertible.
Proof. We let A be the quotient B/ f . We show that λ − u is invertible for all λ in C and furthermore that the inverse is uniformly bounded. We can then apply Theorem 2.

Application 2: Wiener's Tauberian Theorem
Let C(R) be the space of continuous functions of compact support and let L = L 1 (R) be the completion of this space for the L 1 norm. If we define T x (g)(y) = g(x + y), for g in C(R), we have |T x (g) − T x (h)| = |g − h| and, by extension, we have a continuous function x → T x (g), R → L 1 (R) for any g in L such that |T x (g)| = |g| for all x. For f in C(R) and g in L, we define It follows that we have |f * g| |f | ∞ |g| 1 . Hence the product f * g can be defined by extension for f and g both in L. If both f, g are in C(R), so is f * g.
Since we have for f, g in C(R) this holds also for f, g in L and the algebra L with convolution product is a commutative Banach algebra. (It can be shown that it does not have any unit element.) If g is a continuous function with compact support we definê We then haveĝ in C 0 (R), that is,ĝ is continuous on R and tends to 0 at infinity. This is first proved for g ∈ C 1 (R) by integration by parts and then for general g ∈ C(R) by density. Moreover, |ĝ(p)| |g| for all p. It follows that we can extend the map g →ĝ and defineĝ in C 0 (R) for g in L.
We  Proof. We let I be the ideal of all elements k * h − k and B be the Banach algebra L/I. Notice that B has h as unit element. We claim that f is invertible in B.
The map α : R → B defined by α(x) = T x (h) is continuous and satisfies α(0) = h and α(x + y) = α(x) * α(y). By Lemma 3 there exists an element u in B such that T x (h) = e ux for all x in R. In B we have We show that u − λ is invertible mod f and of bounded inverse. We can then apply Theorem 2 and deduce that f is invertible in B.
1. We claim that ip − u is invertible modulo f if |p| M . Indeed, if we take N such that Since, this element is invertible mod f , by hypothesis its inverse is bounded by 1/ε. By Lemma 5, dx is invertible mod f . Let g be its inverse. Then |g| 1/ε(1 − 1 2 ) = 2/ε. Since this integral is divisible by ip − u, it remains to find an explicit bound for the inverse. We have e −ipx − e −ux = e −ipx (1 − e (ip−u)x ) and for all r, r|(1 − e rx ) and | 1−e rx r | 1 + e |r|x , as a simple power series argument shows. Consequently, the inverse of ip − u is bounded by 2 ǫ N −N |f (x)|(1 + e rx )dx, where r > |ip − u|.
2. We claim that ip − u is invertible with bounded inverse for all p such that |p| M + δ. Indeed, if we take η > 0 such that |ĥ(p)| 1 − η and N such that dx is invertible with inverse bounded by 2/η. As before, this element is divisible by ip − u and its inverse is bounded by 2 η N −N |h(x)||1 + e rx |dx, where r > |ip − u|. We conclude that ip − u is invertible for all p in R and its inverse is bounded by d := 2 ǫ (|h(x)| + |f (x)|)(1 + e rx )dx, where r > |ip − u|. Using Lemma 5 it follows that we can find δ > 0 such that r + ip − u is invertible for all p in R and r such that |r| < 1/2d, the inverse is bounded by 2d.
We have finished the proof that for each λ, u − λ is invertible mod f and of bounded inverse.
It follows that we have k such that f * k = h mod I. If we assume g * h = g we have g * I = 0 and then f * k * g = g so that f divides g.
Proposition 1. [Kat04, VI.1.13] If g in L 1 (R) is such thatĝ has compact support then there exists h in L 1 (R) such that h * g = g. In fact, we can use a De la Vallée Poussin's kernel for h.
We deduce the following version of Wiener's Tauberian Theorem.
Corollary 2. If f in L is such thatf never takes the value 0, then every function g such thatĝ is of compact support is in the ideal generated by f . Consequently, this ideal is dense in L.
Since f * g = g(x)T −x (f )dx another way to state this Corollary is the following one.
Corollary 3. If f in L is suchf never takes the value 0 then the vector space generated by the functions T x (f ) is dense in L.
Constructively, the hypothesis thatf never takes the value 0 should be read as: f is bounded away from 0 on any compact.

Conclusions and Future work
To have a point-free description of the spectrum of a Banach algebra was the goal of the work of de Bruijn and van der Meiden [dBvdM67]. The complexity of this description and the process of finding the proof of the compactness of the spectrum is cited by de Bruijn as one inspiration for his AUTOMATH project [dBvdM67,dB94].
It would be interesting to have an actual implementation of our work, following already existing work formalising basic analysis [OS08].
The work of Krivine [Kri64] contains several examples similar to Wiener's results, but in a 'real' framework. One considers respectively l 1 (N) instead of l 1 (Z) and L 1 (R + ) instead of L 1 (R). It might be interesting to give a constructive interpretation these results using the technique presented in [Coq05,CS05].
Constructive, and choice-free, results on Banach spaces can be interpreted as results on Banach sheaves, or equivalently, Banach bundles [Mul80]. Similarly, constructive results on Banach algebras can be interpreted as results on Banach algebra bundles [KR94].