betaFeedback

Reading v1View latest version

WorksMathematics

The Gleason–Kahane–Zelazko theorem: a gentle introduction

Abstract

After some personal recollections of Jean-Pierre Kahane, this paper gives a gentle introduction to the Gleason-Kahane-Zelazko theorem. The theorem states that a normalised complex-linear functional on a unital complex Banach algebra is multiplicative whenever it does not vanish on invertible elements. Neither continuity of the functional nor commutativity of the algebra is assumed. We give a detailed account of three proofs: the original argument of Gleason and, independently, Kahane and Zelazko; the polynomial proof of Roitman and Sternfeld; and a proof by the author. We explain the complex analysis behind each: a special case of Hadamard's factorisation theorem, a polynomial argument using the fundamental theorem of algebra, and the Pólya-Saxer theorem on zero-free entire functions and their derivatives. The elementary entire-function lemma needed for the first proof is proved in full; the Pólya-Saxer theorem is used without proof. We give full details of the algebraic argument common to the three proofs, which deduces multiplicativity from the preservation of squares, including the non-commutative case. We then discuss applications and counterexamples, and present an extension to Ransford's generalised spectra using the Roitman–Sternfeld argument. We conclude with a brief guide to the more recent literature.

Introduction

Conversations with Kahane

I am very proud to have been Jean-Pierre Kahane’s PhD student—his last thésard, as we say in colloquial French.

I first met him at one of the Monday-afternoon harmonic analysis seminars in Orsay. I was then studying for the Diplôme d’études approfondies de Mathématiques Pures, a postgraduate research qualification broadly equivalent to today’s French Master 2. The programme was jointly run by Université Paris-XI (Orsay), the École normale supérieure de Cachan and the École polytechnique. When I told Kahane that I hoped to go on to a PhD in analysis, he made it clear that he would not be taking on any more doctoral students. He nevertheless found time for occasional conversations with me, despite his busy schedule.

We spoke first about Paul Erdős and other Hungarian mathematicians. Although Kahane had never written a joint paper with Erdős,[1] he regarded him as a friend and shared with me some stories about their encounters. Kahane later wrote about Erdős in [32]. He also explained several results that Erdős had proved using the probabilistic method. Later, he introduced me to some of his own applications of probability to analysis.

My own connection with Erdős went back to one of his open problems, which I had solved when I was nineteen. He had asked about the arithmetic nature of the sum

θ=∑n=1∞1F2n+1,\theta= \sum_{n=1}^{\infty} \frac{1}{F_{2n+1}},

where (Fn)(F_n) is the Fibonacci sequence; the question appears, with related problems, in his book with Ronald Graham [23]. I found an elementary proof that θ\theta is irrational. The result was subsequently published in [7].

Erdős himself sent me a very kind letter, even suggesting where I might submit the paper. His advice probably helped it find its way into print: my exposition was very poor. He also suggested further problems, which I could not solve at the time, and invited me to visit him in Budapest,2 an invitation that, unfortunately, I never took up. He encouraged me to pursue an academic career, which, with some good fortune, I eventually did.

Elementary though my proof was, I think it helped Kahane to see me as an interlocuteur mathématique—someone with whom he could discuss mathematics. To my surprise, he wanted to hear the details of my irrationality argument. By then, however, I had forgotten my original proof. Trying to reconstruct it at the blackboard in his office gave me the idea that led to my later paper [9]. I return to these irrationality questions in § 1.3.

It was during our early mathematical conversations that Kahane introduced me to the Gleason–Kahane–Żelazko theorem. I managed to prove a few modest results, some of which were announced, without proofs, in a short note in the Comptes rendus de l’Académie des sciences [8], presented to the Academy by Jean-Pierre Kahane. A later account of the third proof presented here appeared in [11]; further papers related to GKZ include [10, 12]. I believe that this early work on GKZ also helped persuade him to reconsider his decision about taking on another doctoral student.

We also spoke about the history of mathematics. I had read several books and articles about the Lwów school and the Scottish Café, and knew a few stories that were new to him. Kahane was preparing a talk on the Polish mathematical school before the Second World War, and some of the material we discussed found its way into his lecture at the Polish Academy of Sciences’ Scientific Centre in Paris on 9 April 1992 [31].

Kahane kindly invited me to attend his talk. My recollection of the lecture itself is rather hazy. What I do remember is sitting between two relatively young Polish guests whose interest in the history of mathematics seemed rather limited. They appeared to be waiting for the lecture to end and the reception to begin. There was a generous supply of Polish vodka and French wine.

A souvenir from our first meetings

Figure 1 reproduces the first page of an offprint of Kahane and Żelazko’s paper in Studia Mathematica, with a handwritten dedication from Kahane to me.

The first page of Kahane and Żelazko’s paper, with Jean-Pierre Kahane’s handwritten dedication to the author

Figure 1. The first page of Kahane and Żelazko’s paper [33], with Jean-Pierre Kahane’s handwritten dedication to the author.

The dedication takes me back to the period before I began my PhD. Mme Josette Dumas, the secretary of the harmonic analysis group in Orsay, had given me the offprint. Kahane added the dedication in his office, writing very quickly, just before leaving for a meeting. I remember the scene as vividly as if it were yesterday. Soon afterwards, he agreed to supervise my PhD, just in time for me to renew my scholarship and begin doctoral work as

a Boursier du Gouvernement Français, a French government scholarship holder. I remain deeply grateful to Jean-Pierre Kahane for that opportunity and for all that he taught me, both during my PhD and in the years that followed.

My thoughts also turn to Wiesław Żelazko, who died in November 2025. I fondly remember our conversations many years ago in Orsay, Warsaw and Będlewo—about Banach algebras, Solidarność and much else—and the kindness he always showed me. One evening, he invited me to dinner at his home in Warsaw and, at the end of the meal, asked each of his guests to write a few words in his visitors’ book. As I leafed through its pages, I recognised several famous names. Having only recently completed my PhD, I was surprised to find myself adding my own alongside them.

A digression on the irrationality of some Fibonacci series

Erdős’s problem mentioned above belongs to a family of questions about the irrationality or transcendence of the sums of certain infinite series of rational numbers. To recall the background, write

F0=0,F1=1,Fn+2=Fn+1+Fn(n≥0).F_0 = 0,\qquad F_1 = 1,\qquad F_{n+2} = F_{n+1} + F_n\quad(n \ge0).

Good [27], Hoggatt and Bicknell [29], and probably many others, rediscovered Lucas’s identity [37]

∑n=0∞1F2n=7−52.\sum_{n=0}^{\infty} \frac{1}{F_{2^n}} = \frac{7-\sqrt{5}}{2}.

This sum is therefore algebraic and irrational. By contrast, the sum

∑n=0∞1n!F2n\sum_{n=0}^{\infty} \frac{1}{n!F_{2n}}

is transcendental. This was proved by Mignotte [39] and subsequently obtained by Mahler using a different method [38].

The series θ=∑n≥11/F2n+1\theta= \sum_{n\ge1} 1/F_{2n+1} of § 1.1 has a general term of the same order of growth as 1/F2n1/F_{2n}, but no closed form comparable to Lucas’s identity is available for its sum. Erdős asked about the arithmetic nature of θ\theta, together with related problems, in [23]. I proved in 1987 that θ\theta is irrational [7]. In 1994, Becker and Töpfer [13] proved that θ\theta is transcendental, which gives a complete answer to Erdős’s question.

One of the further questions in Erdős’s letter, also recorded in [23], concerned the more general series

∑k=1∞1Fnk,(1)\sum_{k=1}^{\infty} \frac{1}{F_{n_k}}, \tag*{(1)}

where (nk)(n_k) is a strictly increasing sequence of positive integers. Must its sum be irrational whenever nk+1/nk≥cn_{k+1}/n_k \ge c for some fixed c>1c > 1? In my 1993 paper in Acta Arithmetica [9], I answered this affirmatively for c≥2c \ge2. In 2022, Nguyen [40] proved that the sum in (1) is transcendental under the assumption nk+1/nk≥c>2n_{k+1}/n_k \ge c > 2. The strict inequality is necessary: for nk=2k−1n_k = 2^{k-1}, the successive ratios are exactly 2, yet the sum is algebraic by Lucas’s identity displayed above. Thus c>2c > 2 cannot be replaced by c≥2c \ge2 in a general transcendence statement.

The general irrationality question is listed as Problem 267 on the Erdős Problems website [15]. More recently, the Star Fleet project has made available an AI-assisted proof covering the full range c>1c > 1, together with a Lean formalisation [45]. As of 25 October 2026, the Formal Conjectures repository records this case as solved and links to the formal proof [24].

I also discussed irrationality questions with Richard André-Jeannin when he visited me in Orsay. His 1989 proof of the irrationality of

∑k=1∞1Fk,\sum_{k=1}^{\infty} \frac{1}{F_k},

using Apéry’s method, is a particularly striking result [2]. Whether this sum is transcendental appears to be an open problem.

What this paper is about

The aim of these notes is to give a pedagogical account of the Gleason–Kahane–Żelazko (GKZ) theorem. We present three proofs and explain precisely which results from complex analysis each requires. We also include comments and observations, together with material from the work underlying the 1991 Comptes rendus note [8] that did not appear in the published version.

The exposition is intended to be largely self-contained. After the proofs, we give applications and examples showing the role of the hypotheses. § 10 returns to one of the results announced in the 1991 note: the extension of GKZ to Ransford’s generalised spectra. We present an elementary proof based on the Roitman–Sternfeld argument. A short guide to further developments, including recent results for reproducing kernel Hilbert spaces, concludes the paper (§ 11).

Acknowledgements

I would like to thank Hervé Queffélec for the opportunity to write these notes, and for his help and patience. I am also grateful to Persi Diaconis for a conversation about Gleason, Kahane and the Gleason–Kahane–Żelazko theorem, which prompted me to look at possible applications of GKZ to quantum mechanics. Although I usually write mathematics in English, I first wrote this article in French as a tribute to Jean-Pierre Kahane, before translating it into English. The French version is available from the author on request. The author acknowledges the support of the CDP C2EMPI and its institutional partners for the project R-CDP-24-004-C2EMPI, and of the EU COST Actions.

The theorem and the three routes to it

Conventions. A unital Banach algebra is a complex algebra with a non-zero identity 11, complete for a submultiplicative norm. We assume that ∥1∥=1\lVert1\rVert= 1. The main characters in these notes are, literally, characters: a character on a unital complex algebra AA is a non-zero complex-linear map φ:A→C\varphi: A \to\mathbb{C} satisfying

φ(ab)=φ(a)φ(b)(a,b∈A).\varphi(ab) = \varphi(a)\varphi(b) \qquad(a,b \in A).

A character necessarily satisfies φ(1)=1\varphi(1) = 1, since φ(a)=φ(a)φ(1)\varphi(a) = \varphi(a)\varphi(1) and φ(a)≠0\varphi(a) \ne0 for some a∈Aa \in A. Moreover, if uu is invertible, then

φ(u)φ(u−1)=φ(1)=1.\varphi(u)\varphi(u^{-1}) = \varphi(1) = 1.

Thus a character cannot vanish on an invertible element. The remarkable fact is that, for a normalised linear functional on a complex Banach algebra, this necessary condition is also sufficient for multiplicativity.

Theorem 2.1 (Gleason–Kahane–Żelazko). Let AA be a unital complex Banach algebra and let φ:A→C\varphi: A \to\mathbb{C} be complex-linear. Suppose that

φ(1)=1,φ(u)≠0(u∈A−1).(2)\varphi(1) = 1, \qquad\varphi(u) \ne0 \qquad(u \in A^{-1}). \tag*{(2)}

Then

φ(ab)=φ(a)φ(b)(a,b∈A).(3)\varphi(ab) = \varphi(a)\varphi(b) \qquad(a,b \in A). \tag*{(3)}

The functional is automatically continuous. With the normalisation ∥1∥=1\lVert1\rVert= 1, it satisfies ∥φ∥=1\lVert\varphi\rVert= 1.

The normalisation of the functional is needed for this precise conclusion. If a linear functional Λ\Lambda is merely assumed to be non-zero on every invertible element, then Λ(1)≠0\Lambda(1) \ne0, and the theorem applies to

φ=ΛΛ(1).\varphi= \frac{\Lambda}{\Lambda(1)}.

The conclusion in this case is

Λ(ab)Λ(1)=Λ(a)Λ(b).(4)\Lambda(ab)\Lambda(1) = \Lambda(a)\Lambda(b). \tag*{(4)}

The theorem was proved independently by Gleason [26] and by Kahane and Żelazko [33]. Gleason’s paper appeared in 1967, and that of Kahane and Żelazko a year later. Wiesław Żelazko once told me how their proof came about, and Kahane later confirmed the story. During a seminar talk in Orsay, Żelazko mentioned the result as a conjecture. Kahane immediately saw how to apply Hadamard’s factorisation theorem. Gleason independently found the same approach. Sometimes an idea is in the air, and several mathematicians arrive at it at much the same time.

In a recent conversation about the Gleason–Kahane–Żelazko theorem, Persi Diaconis remarked that it is not a result one would immediately associate with the names of Kahane or Gleason. I do not know for certain why Gleason became interested in maximal ideals (or characters). One conjecture, possibly unfounded, is that the interest came from quantum mechanics. Using the Gleason–Kahane–Żelazko theorem, Bach [6] discussed the theoretical assumptions on which Bell’s theorem on local hidden variables is based. Andrew M. Gleason is also famous for a result of 1957, answering a question posed by George W. Mackey, which played a historically significant role in showing that wide classes of hidden-variable theories are inconsistent with quantum physics. Gleason’s theorem [25] shows that the Born rule, the rule used to calculate probabilities in quantum physics, can be derived from the usual mathematical representation of measurements together with the assumption of non-contextuality.

How the proofs fit together

Linearity relates sums and scalar multiples; multiplicativity relates products. The hypothesis in (2) says nothing directly about a pair of elements a,ba,b: it tests only whether the functional vanishes on the group of invertible elements. Nevertheless, this condition determines its behaviour on every product.

Each of the three proofs first establishes the apparently weaker identity

φ(a2)=φ(a)2(a∈A).(5)\varphi(a^2) = \varphi(a)^2 \qquad(a \in A). \tag*{(5)}

For a commutative algebra, expanding (a+b)2(a+b)^2 then gives multiplicativity. In a non-commutative algebra, this expansion initially provides information only about the Jordan product 12(ab+ba)\frac{1}{2}(ab+ba), so a further argument is needed. We give that argument once, in full, in § 4.

The three routes to (5) use different properties of complex functions.

  1. Gleason–Kahane–Żelazko: growth of a zero-free entire function. For fixed aa, the function z↦φ(eza)z \mapsto\varphi(e^{za}) is entire, zero-free and of at most exponential growth. Such a function is the exponential of an affine function. Comparing its first two derivatives at the origin gives (5).

  1. Roitman–Sternfeld: roots of polynomials. For φ(x)=0\varphi(x)=0, the roots of the monic polynomial φ((λ1−x)n)\varphi((\lambda_1-x)^n) lie in σA(x)\sigma_A(x), the spectrum of xx. Its first two coefficients

show that φ(x2)≠0\varphi(x^2) \ne0 would force some of these roots to escape to infinity as n→∞n \to\infty. Boundedness of σA(x)\sigma_A(x) rules this out.

  1. Non-vanishing of derivatives [11]. If xx is invertible, every derivative of z↦φ(ezx)z \mapsto\varphi(e^{zx}) is zero-free. The Pólya–Saxer theorem identifies this function as an exponential. A scalar shift then transfers the square identity from invertible elements to arbitrary elements.

The remark following Lemma 4.1 explains why the three different starting points all lead to the square identity for every element of AA.

The original proofs of Gleason and of Kahane and Želazko use Hadamard’s factorisation theorem. Only a special case is needed, and it can be proved by elementary complex analysis. Siddiqi [44] gave an early proof along these lines; the first proof below follows the same principle. We establish the required entire-function lemma directly, using a holomorphic logarithm and a coefficient estimate for its real part.

The polarisation argument in Kahane and Želazko’s paper assumes commutativity. A footnote added in proof records the extension to non-commutative algebras, referring to Želazko [50]. Combining their analytic argument with the algebraic lemma in Section 4 gives the theorem in its full non-commutative form. Gleason’s note [26] likewise treats only the commutative case.

Preliminaries: invertibility, spectra, and continuity

Unless an assertion is explicitly stated for an arbitrary unital complex algebra, AA is a unital complex Banach algebra with a submultiplicative norm satisfying ∥1∥=1\lVert1\rVert= 1.

The spectrum needs no topology in its definition

For a unital complex algebra AA and a∈Aa \in A, define

σA(a)={λ∈C:a−λ1∉A−1}.\sigma_A(a) = \{\lambda\in\mathbb{C} : a - \lambda1 \notin A^{-1}\}.

This definition is purely algebraic. A norm and completeness enter the standard Banach-algebra proofs that spectra are bounded, closed and non-empty. The usual proof of non-emptiness of the spectrum uses Liouville’s theorem. The GKZ proofs below do not require this general non-emptiness result.

Lemma 3.1 (The spectral reformulation). For a complex-linear map φ:A→C\varphi: A \to\mathbb{C} on a unital complex algebra, the following conditions are equivalent:

(i) φ(1)=1\varphi(1) = 1 and φ(u)≠0\varphi(u) \ne0 for every u∈A−1u \in A^{-1};

(ii) φ(a)∈σA(a)\varphi(a) \in\sigma_A(a) for every a∈Aa \in A.

Proof. Assume (i). For any a∈Aa \in A,

φ(a−φ(a)1)=0.\varphi(a - \varphi(a)1) = 0.

The element a−φ(a)1a - \varphi(a)1 therefore cannot be invertible. By definition this is exactly φ(a)∈σA(a)\varphi(a) \in\sigma_A(a).

Conversely, σA(1)={1}\sigma_A(1) = \{1\}, so (ii) gives φ(1)=1\varphi(1) = 1. If uu is invertible, then 0∉σA(u)0 \notin\sigma_A(u); hence φ(u)≠0\varphi(u) \ne0.

A functional satisfying these equivalent conditions may be called a linear spectral selection. The theorem says that, on a complex Banach algebra, a linear choice of a spectral value must be multiplicative, that is, a character. The spectral-selection point of view leads to a different interpretation of the Gleason–Kahane–Żelazko theorem, as a kind of local-to-global principle. When the Banach algebra AA is commutative, classical Gelfand theory shows that the spectrum σA(a)\sigma_A(a) of aa is exactly the set of values γ(a)\gamma(a), where γ\gamma ranges over the characters of AA. Consequently, if a normalised linear functional φ\varphi on a commutative Banach algebra AA has the property that, for every a∈Aa \in A, there is a character γa\gamma_a, depending on aa, with φ(a)=γa(a)\varphi(a) = \gamma_a(a), then φ\varphi is itself a character.

A remarkable non-linear extension of the Gleason–Kahane–Żelazko theorem is the Kowalski–Słodkowski theorem [34]. If φ\varphi is a complex-valued function on a unital Banach algebra AA, not necessarily linear, satisfying φ(0)=0\varphi(0) = 0 and the spectral condition

φ(x)−φ(y)∈σA(x−y)(x,y∈A),\varphi(x) - \varphi(y) \in\sigma_A(x-y) \qquad(x,y \in A),

then φ\varphi is automatically linear and multiplicative. Thus linearity is now part of the conclusion. For a linear φ\varphi, the spectral condition reduces to φ(a)∈σA(a)\varphi(a) \in\sigma_A(a) for every a∈Aa \in A, so the statement contains the classical Gleason–Kahane–Żelazko theorem.

The Kowalski–Słodkowski theorem implies the following ‘2-local-to-global’ principle. If a complex-valued function φ\varphi on AA has the property that, for every pair a,b∈Aa,b \in A, there is a character γa,b\gamma_{a,b}, depending on aa and bb, with φ(a)=γa,b(a)\varphi(a) = \gamma_{a,b}(a) and φ(b)=γa,b(b)\varphi(b) = \gamma_{a,b}(b), then φ\varphi is itself a character, that is, linear and multiplicative. Indeed, under this hypothesis,

φ(a)−φ(b)=γa,b(a−b)∈σA(a−b)(a,b∈A),\varphi(a) - \varphi(b) = \gamma_{a,b}(a-b) \in\sigma_A(a-b) \qquad(a,b \in A),

and φ(0)=γ0,0(0)=0\varphi(0) = \gamma_{0,0}(0) = 0, so the Kowalski–Słodkowski theorem applies.

The Neumann series and the spectral bound

We first recall exactly where completeness is used to bound the spectrum. The series in the following lemma is named after Carl Neumann.

Lemma 3.2 (Neumann series). If b∈Ab \in A and ∥b∥<1\lVert b\rVert< 1, then 1−b1-b is invertible, with

(1−b)−1=∑n=0∞bn,∥(1−b)−1∥≤11−∥b∥.(1-b)^{-1} = \sum_{n=0}^{\infty} b^n,\qquad\lVert(1-b)^{-1}\rVert\leq\frac{1}{1-\lVert b\rVert}.

Consequently,

σA(a)⊂{λ∈C:∣λ∣≤∥a∥}.(6)\sigma_A(a) \subset\{\lambda\in\mathbb{C}: \lvert\lambda\rvert\leq\lVert a\rVert\}. \tag*{(6)}

Proof. The series is absolutely convergent because ∥bn∥≤∥b∥n\lVert b^n\rVert\leq\lVert b\rVert^n for n≥0n \geq0, using ∥1∥=1\lVert1\rVert= 1 when n=0n = 0. Completeness ensures that its sum belongs to AA. For its partial sums,

(1−b)∑n=0Nbn=∑n=0Nbn(1−b)=1−bN+1.(1-b)\sum_{n=0}^{N} b^n = \sum_{n=0}^{N} b^n(1-b) = 1-b^{N+1}.

Since ∥bN+1∥→0\lVert b^{N+1}\rVert\to0, continuity of multiplication gives a two-sided inverse after passage to the limit. The norm estimate follows from the geometric series. If ∣λ∣>∥a∥\lvert\lambda\rvert> \lVert a\rVert, apply this to b=a/λb = a/\lambda and use 1−a=λ(1−a/λ)1-a = \lambda(1-a/\lambda). Thus λ\lambda lies outside the spectrum, which proves (6). ∎

Proposition 3.3 (Automatic continuity before multiplicativity). Let φ:A→C\varphi:A\to\mathbb{C} be complex-linear and satisfy (2). Then

∣φ(a)∣≤∥a∥(a∈A).|\varphi(a)|\leq\|a\| \qquad(a\in A).

In particular, φ\varphi is continuous and, with our normalisation ∥1∥=1\|1\|=1, satisfies ∥φ∥=1\|\varphi\|=1.

Proof. Lemma 3.1 places φ(a)\varphi(a) in the spectrum, and Lemma 3.2 bounds its modulus. This proves ∥φ∥≤1\|\varphi\|\leq1. The equality φ(1)=1\varphi(1)=1 gives the reverse inequality when ∥1∥=1\|1\|=1.

This order of reasoning matters. We shall apply φ\varphi to an infinite power series. An arbitrary algebraic linear functional need not commute with a norm limit. Proposition 3.3 supplies precisely the missing justification, without using the conclusion of GKZ.

Notice also that we have not used the general theorem asserting that every Banach-algebra spectrum is non-empty. Under the GKZ hypothesis, the explicit scalar φ(a)\varphi(a) already belongs to σA(a)\sigma_A(a). For the arguments below, the upper bound provided by the Neumann series is sufficient.

Exponentials in a Banach algebra

For a∈Aa\in A and z∈Cz\in\mathbb{C}, define

eza=∑n=0∞znann!.e^{za}=\sum_{n=0}^{\infty}\frac{z^n a^n}{n!}.

The series converges absolutely and uniformly on bounded subsets of the zz-plane, since ∥an∥≤∥a∥n\|a^n\|\leq\|a\|^n for all n≥0n\geq0 and

∑n=0∞∣z∣n∥a∥nn!=e∣z∣∥a∥.\sum_{n=0}^{\infty}\frac{|z|^n\|a\|^n}{n!}=e^{|z|\|a\|}.

It follows in particular that

∥eza∥≤e∣z∣∥a∥.(7)\|e^{za}\|\leq e^{|z|\|a\|}. \tag*{(7)}

We may multiply the two absolutely convergent series for ezae^{za} and e−zae^{-za} and collect terms of equal degree. Their constant term is 11, while for n≥1n\geq1 the coefficient multiplying znanz^n a^n is

∑k=0n(−1)n−kk!(n−k)!=(1−1)nn!=0.\sum_{k=0}^{n}\frac{(-1)^{n-k}}{k!(n-k)!}=\frac{(1-1)^n}{n!}=0.

Thus

ezae−za=e−zaeza=1.(8)e^{za}e^{-za}=e^{-za}e^{za}=1. \tag*{(8)}

This calculation is valid in a non-commutative algebra: all terms here are powers of the same element, so they commute with one another.

By continuity of φ\varphi,

Fa(z):=φ(eza)=∑n=0∞φ(an)n!zn.(9)F_a(z):=\varphi(e^{za})=\sum_{n=0}^{\infty}\frac{\varphi(a^n)}{n!}z^n. \tag*{(9)}

Since ∣φ(an)∣≤∥a∥n|\varphi(a^n)|\leq\|a\|^n, the series has infinite radius of convergence, and it can be differentiated term by term. In particular,

Fa(k)(z)=φ(akeza),k≥0,(10)F_a^{(k)}(z)=\varphi(a^k e^{za}),\qquad k\geq0, \tag*{(10)}

and

Fa(0)=1,Fa′(0)=φ(a),Fa′′(0)=φ(a2).(11)F_a(0)=1,\qquad F_a'(0)=\varphi(a),\qquad F_a''(0)=\varphi(a^2). \tag*{(11)}

These elementary observations support both analytic proofs.

The common algebraic conclusion

The square defect

For a complex-linear map φ:A→C\varphi: A \to\mathbb{C} with φ(1)=1\varphi(1) = 1, the square defect of φ\varphi is

dφ(a)=φ(a2)−φ(a)2(a∈A).d_{\varphi}(a) = \varphi(a^2) - \varphi(a)^2 \qquad(a \in A).

Each of the three proofs is a way of showing that dφd_{\varphi} vanishes. The following elementary lemma explains how the three arguments fit together.

Lemma 4.1 (Properties of the square defect). Let AA be a unital complex algebra and let φ:A→C\varphi: A \to\mathbb{C} be complex-linear with φ(1)=1\varphi(1) = 1. For a,b∈Aa,b \in A and λ,μ∈C\lambda,\mu\in\mathbb{C},

(i) dφ(a+λ1)=dφ(a)d_{\varphi}(a+\lambda1)=d_{\varphi}(a);

(ii) dφ(μa)=μ2dφ(a)d_{\varphi}(\mu a)=\mu^2d_{\varphi}(a);

(iii) dφ(a+b)=dφ(a)+dφ(b)+φ(ab+ba)−2φ(a)φ(b)d_{\varphi}(a+b)=d_{\varphi}(a)+d_{\varphi}(b)+\varphi(ab+ba)-2\varphi(a)\varphi(b).

Proof. (i) Expanding both terms,

dφ(a+λ1)=φ(a2)+2λφ(a)+λ2−(φ(a)+λ)2=φ(a2)−φ(a)2.\begin{aligned} d_{\varphi}(a+\lambda1) &= \varphi(a^2)+2\lambda\varphi(a)+\lambda^2-(\varphi(a)+\lambda)^2\\ &=\varphi(a^2)-\varphi(a)^2. \end{aligned}

(ii) is immediate from the definition. For (iii), expand (a+b)2=a2+ab+ba+b2(a+b)^2=a^2+ab+ba+b^2 and (φ(a)+φ(b))2(\varphi(a)+\varphi(b))^2.

Remark 4.2. Part (iii) shows that, when dφ≡0d_{\varphi}\equiv0, one obtains the polarisation identity φ(ab+ba)=2φ(a)φ(b)\varphi(ab+ba)=2\varphi(a)\varphi(b), stated as (12) below. Part (i) shows that dφ≡0d_{\varphi}\equiv0 as soon as dφd_{\varphi} vanishes on a set that meets every coset a+C1a+\mathbb{C}1. This explains the different starting points of the three proofs. Proof I shows that dφ(a)=0d_{\varphi}(a)=0 for every a∈Aa\in A. Proof II works on ker⁡φ\ker\varphi, which meets every coset because a−φ(a)1∈ker⁡φa-\varphi(a)1\in\ker\varphi; there dφ(x)=φ(x2)d_{\varphi}(x)=\varphi(x^2). Proof III works on the invertible elements, which meet every coset in a Banach algebra because a−λ1∈A−1a-\lambda1\in A^{-1} whenever ∣λ∣>∥a∥|\lambda|>\lVert a\rVert.

Why polarisation alone is insufficient in the non-commutative case

Suppose that a complex-linear map φ:A→C\varphi: A \to\mathbb{C} preserves squares. Expansion of (a+b)2(a+b)^2 gives

φ(ab+ba)=2φ(a)φ(b).(12)\varphi(ab+ba)=2\varphi(a)\varphi(b). \tag*{(12)}

If ab=baab=ba, this already gives multiplicativity for the pair a,ba,b. Otherwise, (12) determines only φ(ab)+φ(ba)\varphi(ab)+\varphi(ba); one cannot replace ab+baab+ba by 2ab2ab.

For scalar-valued maps, however, square preservation still implies multiplicativity. The following purely algebraic lemma appears in Roitman–Sternfeld [43 Lemma 1]; we spell out its non-commutative step.

Lemma 4.3 (Scalar-valued Jordan maps are multiplicative). Let AA be a unital complex algebra and let φ:A→C\varphi: A \to\mathbb{C} be complex-linear with φ(1)=1\varphi(1)=1. The following are equivalent:

(i) If φ(x)=0\varphi(x)=0, then φ(x2)=0\varphi(x^2)=0.

(ii) φ(x2)=φ(x)2\varphi(x^2)=\varphi(x)^2 for every x∈Ax\in A.

(iii) If φ(x)=0\varphi(x)=0, then φ(xy)=0\varphi(xy)=0 for every y∈Ay\in A.

(iv) φ(xy)=φ(x)φ(y)\varphi(xy)=\varphi(x)\varphi(y) for all x,y∈Ax,y\in A.

Proof. Assume (i). For arbitrary a∈Aa\in A, put x=a−φ(a)1x=a-\varphi(a)1. Then φ(x)=0\varphi(x)=0, and hence

0=φ(x2)=φ(a2)−2φ(a)φ(a)+φ(a)2φ(1)=φ(a2)−φ(a)2.0=\varphi(x^2)=\varphi(a^2)-2\varphi(a)\varphi(a)+\varphi(a)^2\varphi(1) =\varphi(a^2)-\varphi(a)^2.

This proves (ii).

Assume (ii). Polarisation gives (12) for all a,ba,b. Fix x,y∈Ax,y\in A with φ(x)=0\varphi(x)=0, and set

s=xy+yx,c=xy−yx.s=xy+yx,\qquad c=xy-yx.

Equation (12) gives φ(s)=0\varphi(s)=0, and (ii) then gives φ(s2)=0\varphi(s^2)=0. The identity

c2=2(x(yxy)+(yxy)x)−s2(13)c^2=2\bigl(x(yxy)+(yxy)x\bigr)-s^2 \tag*{(13)}

is verified by expanding both sides. Applying (12) to xx and yxyyxy, we obtain

φ(c2)=4φ(x)φ(yxy)−φ(s2)=0.\varphi(c^2)=4\varphi(x)\varphi(yxy)-\varphi(s^2)=0.

Again by (ii), φ(c)2=0\varphi(c)^2=0. The range is the scalar field, so φ(c)=0\varphi(c)=0. Therefore

2φ(xy)=φ(s)+φ(c)=0,2\varphi(xy)=\varphi(s)+\varphi(c)=0,

which proves (iii).

Assume (iii), and take arbitrary a,b∈Aa,b\in A. Since a−φ(a)1a-\varphi(a)1 belongs to the kernel,

0=φ((a−φ(a)1)b)=φ(ab)−φ(a)φ(b).0=\varphi\bigl((a-\varphi(a)1)b\bigr)=\varphi(ab)-\varphi(a)\varphi(b).

Thus (iv) holds. Finally, (iv) immediately implies (i).

Remark 4.4 (Where the scalar target is used). The implication φ(c)2=0⇒φ(c)=0\varphi(c)^2=0\Rightarrow\varphi(c)=0 uses the fact that a complex number whose square is zero must itself be zero. Polarisation also uses the commutativity of scalar multiplication. The analogous assertion that square-preserving maps into an arbitrary algebra are multiplicative is false. For instance, matrix transposition is linear and preserves squares, but (ab)T=bTaT(ab)^{\mathrm{T}}=b^{\mathrm{T}}a^{\mathrm{T}}.

First proof: zero-free entire functions of exponential type

The basic observation is that Fa(z)=φ(eza)F_a(z)=\varphi(e^{za}) is zero-free: ezae^{za} is invertible, so the hypothesis (2) excludes Fa(z)=0F_a(z)=0. On the other hand,

∣Fa(z)∣≤∥φ∥∥eza∥≤e∣z∣∥a∥.(14)\lvert F_a(z)\rvert\leq\lVert\varphi\rVert\lVert e^{za}\rVert\leq e^{\lvert z\rvert\lVert a\rVert}. \tag*{(14)}

The analytic content of the proof is the assertion that these two facts force FaF_a to be the exponential of a linear function.

Lemma 5.1 (Zero-free entire functions of exponential type). Let FF be an entire function with F(0)=1F(0)=1 and no zeros. Suppose that there exist constants C≥1C\geq1 and T≥0T\geq0 such that

∣F(z)∣≤CeT∣z∣(z∈C).(15)\lvert F(z)\rvert\leq Ce^{T\lvert z\rvert}\qquad(z\in\mathbb{C}). \tag*{(15)}

Then there is α∈C\alpha\in\mathbb{C} such that

F(z)=eαz(z∈C).F(z)=e^{\alpha z}\qquad(z\in\mathbb{C}).

Moreover, α=F′(0)\alpha=F'(0) and ∣α∣≤T\lvert\alpha\rvert\leq T.

This is the special zero-free case of the factorisation theory of entire functions that is needed in the Kahane–Żelazko argument. Their paper invokes a standard result on entire functions, taken from Titchmarsh [46]. The full Hadamard factorisation theorem is considerably more general: it describes an entire function in terms of its zeros and its growth. Here there are no zeros and the growth is at most exponential, so we can prove the required special case directly.

Step 1: constructing a global logarithm

Since FF has no zeros, F′/FF'/F is entire, so it is the sum of a power series with infinite radius of convergence, F′(z)/F(z)=∑n≥0cnznF'(z)/F(z)=\sum_{n\geq0}c_nz^n. Integrating term by term gives an entire function

h(z)=∑n≥0cnn+1zn+1,h(0)=0,h′=F′F.(16)h(z)=\sum_{n\geq0}\frac{c_n}{n+1}z^{n+1},\qquad h(0)=0,\qquad h'=\frac{F'}{F}. \tag*{(16)}

Differentiating Fe−hFe^{-h} gives

ddz(F(z)e−h(z))=e−h(z)(F′(z)−F(z)h′(z))=0.\frac{d}{dz}\left(F(z)e^{-h(z)}\right)=e^{-h(z)}\left(F'(z)-F(z)h'(z)\right)=0.

Hence Fe−hFe^{-h} is constant; its value at 00 is 11, and consequently

F(z)=eh(z),h(0)=0.(17)F(z)=e^{h(z)},\qquad h(0)=0. \tag*{(17)}

Thus hh is an entire logarithm of FF, normalised by h(0)=0h(0)=0. From (15),

Re⁡h(z)=log⁡∣F(z)∣≤log⁡C+T∣z∣.(18)\operatorname{Re}h(z)=\log|F(z)|\leq\log C+T|z|. \tag*{(18)}

Step 2: a coefficient estimate from the real part

We prove the elementary form of Carathéodory’s coefficient estimate needed here.

Lemma 5.2 (A real-part coefficient estimate). Let r>0r>0, and let hh be holomorphic on a neighbourhood of the closed disc {∣z∣≤r}\{|z|\leq r\}, with h(0)=0h(0)=0. Write h(z)=∑n≥1anznh(z)=\sum_{n\geq1}a_nz^n. If M∈RM\in\mathbb{R} and Re⁡h(z)≤M\operatorname{Re}h(z)\leq M on ∣z∣=r|z|=r, then

∣an∣≤2Mrn(n≥1).(19)|a_n|\leq\frac{2M}{r^n}\qquad(n\geq1). \tag*{(19)}

Proof. Put u(t)=Re⁡h(reit)u(t)=\operatorname{Re}h(re^{it}). The power series converges uniformly on the circle, so the Fourier coefficients of uu can be computed by termwise integration. For $n\geq 1,

12π∫02πu(t)e−int dt=anrn2.(20)\frac{1}{2\pi}\int_0^{2\pi}u(t)e^{-int}\,dt=\frac{a_nr^n}{2}. \tag*{(20)}

Indeed, in

u(t)=12(∑k≥1akrkeikt+∑k≥1ak‾rke−ikt),u(t)=\frac{1}{2}\left(\sum_{k\geq1}a_kr^ke^{ikt}+\sum_{k\geq1}\overline{a_k}r^ke^{-ikt}\right),

only the term with k=nk=n in the first sum has a non-zero integral after multiplication by e−inte^{-int}. Also,

12π∫02πu(t) dt=Re⁡h(0)=0.\frac{1}{2\pi}\int_0^{2\pi}u(t)\,dt=\operatorname{Re}h(0)=0.

The function M−u(t)M-u(t) is non-negative and has mean MM; in particular, M≥0M \ge0. The constant MM has zero nnth Fourier coefficient for n≥1n \ge1, hence (20) implies

∣an∣rn2=∣12π∫02π(M−u(t))e−int dt∣≤12π∫02π(M−u(t)) dt=M.\frac{|a_n|r^n}{2}=\left|\frac{1}{2\pi}\int_0^{2\pi}(M-u(t))e^{-int}\,\mathrm{d}t\right| \le\frac{1}{2\pi}\int_0^{2\pi}(M-u(t))\,\mathrm{d}t=M.

This proves (19).

The point of this estimate is that it controls all coefficients from an upper bound on the real part. A direct Cauchy estimate on hh would require a bound on ∣h∣|h|, which (18) does not immediately provide.

Step 3: the logarithm is linear

Proof of Lemma 5.1. Construct hh as in (17). Apply Lemma 5.2 on every circle of radius r>0r>0, with M=log⁡C+TrM=\log C+Tr. For the Taylor coefficients of hh we obtain

∣an∣≤2(log⁡C+Tr)rn.|a_n|\le\frac{2(\log C+Tr)}{r^n}.

For each n≥2n\ge2, the right-hand side tends to zero as r→∞r\to\infty. Thus an=0a_n=0 for all n≥2n\ge2, and h(z)=a1zh(z)=a_1z. Taking α=a1\alpha=a_1 gives F(z)=eαzF(z)=e^{\alpha z}. Differentiation at 00 gives α=F′(0)\alpha=F'(0).

If α≠0\alpha\ne0, choose z=rα‾/∣α∣z=r\overline{\alpha}/|\alpha|. Then Re⁡(αz)=r∣α∣\operatorname{Re}(\alpha z)=r|\alpha|, and (15) gives r∣α∣≤log⁡C+Trr|\alpha|\le\log C+Tr. Divide by rr and let r→∞r\to\infty to obtain ∣α∣≤T|\alpha|\le T. The case α=0\alpha=0 is immediate.

Completion of the first proof of GKZ

Proof I of Theorem 2.1. Proposition 3.3 gives automatic continuity. Fix any a∈Aa\in A and define FaF_a by (9). It is entire, Fa(0)=1F_a(0)=1, it has no zeros, and it satisfies (14). Lemma 5.1 gives

Fa(z)=eαz.F_a(z)=e^{\alpha z}.

By (11), α=Fa′(0)=φ(a)\alpha=F_a'(0)=\varphi(a), and therefore

φ(eza)=ezφ(a)(z∈C).(21)\varphi(e^{za})=e^{z\varphi(a)}\qquad(z\in\mathbb{C}). \tag*{(21)}

Comparing the second derivatives at zero yields

φ(a2)=Fa′′(0)=φ(a)2.\varphi(a^2)=F_a''(0)=\varphi(a)^2.

Since aa was arbitrary, Lemma 4.3 implies multiplicativity.

In fact, comparison of all coefficients gives

φ(an)=φ(a)n(n≥0).(22)\varphi(a^n)=\varphi(a)^n\qquad(n\ge0). \tag*{(22)}

Only the case n=2n=2 was needed to finish the proof.

Remark 5.3 (Normalising the logarithm). If F(z)=eαz+βF(z)=e^{\alpha z+\beta} and F(0)=1F(0)=1, the exact conclusion is eβ=1e^\beta=1, or β∈2πiZ\beta\in2\pi i\mathbb{Z}. It follows that F(z)=eαzF(z)=e^{\alpha z}, and one may choose the logarithm with constant term zero. It does not follow that an arbitrarily chosen β\beta equals zero. Our construction h(0)=0h(0)=0 removes this ambiguity from the start.

A finite-radius version of the coefficient estimate

Lemma 5.2 does not require an entire function. Let R>0R>0, and suppose that FF is holomorphic and zero-free on the disc ∣z∣<R|z|<R, with F(0)=1F(0)=1. Assume that C≥1C\ge1 and T≥0T\ge0 satisfy

∣F(z)∣≤CeT∣z∣(∣z∣<R).|F(z)| \le Ce^{T|z|} \qquad(|z|<R).

Since F′/FF'/F is holomorphic on the disc, integrating its Taylor series term by term gives a holomorphic function hh on the disc with F=ehF=e^h and h(0)=0h(0)=0, exactly as in (16)–(17). Write h(z)=c1z+c2z2+⋯h(z)=c_1z+c_2z^2+\cdots. For every 0<r<R0<r<R, Lemma 5.2 gives

∣c2∣≤2(log⁡C+Tr)r2.|c_2|\le\frac{2(\log C+Tr)}{r^2}.

Since

h′′(0)=F′′(0)−F′(0)2=2c2,h''(0)=F''(0)-F'(0)^2=2c_2,

we obtain

∣F′′(0)−F′(0)2∣≤4(log⁡C+Tr)r2(0<r<R).(23)|F''(0)-F'(0)^2|\le\frac{4(\log C+Tr)}{r^2}\qquad(0<r<R). \tag*{(23)}

For F(z)=φ(eza)F(z)=\varphi(e^{za}), the left-hand side is ∣dφ(a)∣|d_\varphi(a)|, the modulus of the square defect. Under the GKZ hypotheses, FaF_a is entire and zero-free on the whole plane, and the same growth bound holds on every disc. We may therefore let r→∞r\to\infty and obtain dφ(a)=0d_\varphi(a)=0; this recovers Proof I. The finite-radius inequality is relevant when only local zero-freeness or local growth information is available.

One might try to weaken the GKZ hypothesis by assuming only local non-vanishing near the identity. More precisely, suppose that φ\varphi is complex-linear, φ(1)=∥φ∥=1\varphi(1)=\|\varphi\|=1, and

φ(u)≠0whenever∥u−1∥<δ,δ>0.\varphi(u)\ne0\qquad\text{whenever}\qquad\|u-1\|<\delta,\qquad\delta>0.

We retain ∥1∥=1\|1\|=1. If a=0a=0, then dφ(a)=0d_\varphi(a)=0, so assume a≠0a\ne0. The exponential estimate gives

∥eza−1∥≤e∣z∣∥a∥−1<δfor∣z∣<R:=log⁡(1+δ)∥a∥.\|e^{za}-1\|\le e^{|z|\|a\|}-1<\delta\qquad\text{for}\qquad|z|<R:=\frac{\log(1+\delta)}{\|a\|}.

Thus FaF_a is zero-free on this disc. Applying (23) with C=1C=1 and T=∥a∥T=\|a\|, then letting r↑Rr\uparrow R, yields

∣dφ(a)∣≤4∥a∥2log⁡(1+δ).|d_\varphi(a)|\le\frac{4\|a\|^2}{\log(1+\delta)}.

However, since φ(0)=0\varphi(0)=0, the open ball cannot contain 00, so δ≤1\delta\le1. The right-hand side is therefore at least 4∥a∥2/log⁡2>5∥a∥24\|a\|^2/\log2>5\|a\|^2, whereas the elementary estimate

∣dφ(a)∣≤∣φ(a2)∣+∣φ(a)∣2≤2∥a∥2|d_\varphi(a)|\le|\varphi(a^2)|+|\varphi(a)|^2\le2\|a\|^2

is already better. Indeed, the local non-vanishing assumption adds no information when δ≤1\delta\le1: every normalised norm-one functional satisfies

∣φ(u)−1∣=∣φ(u−1)∣≤∥u−1∥<1|\varphi(u)-1|=|\varphi(u-1)|\le\|u-1\|<1

on the open unit ball centred at 11.

This calculation therefore does not yield a quantitative version of GKZ under a local non-vanishing hypothesis. What it does show is exactly what the real-part estimate gives on a fixed disc, and why letting r→∞r\to\infty is decisive in Proof I.

Why the exponential function?

The preceding proof yields a stronger formulation: it suffices to assume that φ\varphi does not vanish on

exp⁡(A)={ea:a∈A},\exp(A)=\{e^a:a\in A\},

the set of exponentials in AA. Indeed, once continuity is known, Proof I uses the non-vanishing hypothesis only at the exponentials ezae^{za}; continuity itself, through Proposition 3.3, uses it only at the elements λ1−a\lambda1-a with ∣λ∣>∥a∥|\lambda|>\lVert a\rVert. These elements are exponentials too, since 1−b1-b has the logarithm −∑n≥1bn/n-\sum_{n\geq1}b^n/n whenever ∥b∥<1\lVert b\rVert<1.

In 1987, Richard Arens [3] asked whether the exponential function could be replaced by other non-constant entire functions. This problem has been studied, and solved, by Krzysztof Jarosz [30].

Second proof: polynomial roots and bounded spectra

This proof does not construct an exponential and does not need the functional to be continuous. Its key device is a polynomial of increasing degree whose roots are forced into one fixed spectrum.

The theorem in its natural algebraic setting

Theorem 6.1 (Roitman–Sternfeld, bounded-spectrum form). Let AA be a unital complex algebra. Suppose that σA(x)\sigma_A(x) is bounded for every x∈Ax\in A. If a complex-linear functional φ:A→C\varphi:A\to\mathbb{C} satisfies (2), then φ\varphi is multiplicative.

This is the corollary to Lemmas 1 and 2 of Roitman–Sternfeld [43 p. 113]. There is no topology in the statement: the algebra need not be normed or complete. Boundedness is a property of the subsets σA(x)⊂C\sigma_A(x)\subset\mathbb{C}, and the bound may depend on xx. In a Banach algebra it follows from the Neumann series, so Theorem 6.1 immediately implies GKZ. Although boundedness alone allows an empty spectrum in an arbitrary algebra, (2) rules this out: Lemma 3.1 gives φ(x)∈σA(x)\varphi(x)\in\sigma_A(x) for every xx.

The polynomials and the location of their roots

Fix x∈ker⁡φx\in\ker\varphi, and for each integer n≥2n\geq2 consider

Pn(λ)=φ((λ1−x)n).(24)P_n(\lambda)=\varphi\left((\lambda1-x)^n\right). \tag*{(24)}

Since λ1\lambda1 commutes with xx, the binomial formula is valid even when AA is non-commutative. Therefore

Pn(λ)=λn−nφ(x)λn−1+(n2)φ(x2)λn−2+⋯ .(25)P_n(\lambda)=\lambda^n-n\varphi(x)\lambda^{n-1}+\binom{n}{2}\varphi(x^2)\lambda^{n-2}+\cdots. \tag*{(25)}

The leading coefficient is φ(1)=1\varphi(1)=1, so PnP_n is monic of degree exactly nn. By the fundamental theorem of algebra, write

Pn(λ)=∏j=1n(λ−λn,j),(26)P_n(\lambda)=\prod_{j=1}^{n}(\lambda-\lambda_{n,j}), \tag*{(26)}

where the roots are counted with algebraic multiplicity. If Pn(λ)=0P_n(\lambda)=0 and λ1−x\lambda1-x were invertible, then its power (λ1−x)n(\lambda1-x)^n would be invertible, contradicting (2). Hence

λn,j∈σA(x)(1≤j≤n).(27)\lambda_{n,j}\in\sigma_A(x)\qquad(1\leq j\leq n). \tag*{(27)}

This step uses only the fact that powers of invertible elements are invertible; it does not assume multiplicativity of φ\varphi.

The first two Newton identities

Comparing coefficients in (25) and (26), and using φ(x)=0\varphi(x)=0, gives

∑j=1nλn,j=0,(28)\sum_{j=1}^{n}\lambda_{n,j}=0, \tag*{(28)}
∑1≤i<j≤nλn,iλn,j=(n2)φ(x2).(29)\sum_{1\le i<j\le n}\lambda_{n,i}\lambda_{n,j}=\binom{n}{2}\varphi(x^{2}). \tag*{(29)}

Squaring (28) yields

0=∑j=1nλn,j2+2∑1≤i<j≤nλn,iλn,j.0=\sum_{j=1}^{n}\lambda_{n,j}^{2}+2\sum_{1\le i<j\le n}\lambda_{n,i}\lambda_{n,j}.

Thus

∑j=1nλn,j2=−n(n−1)φ(x2).(30)\sum_{j=1}^{n}\lambda_{n,j}^{2}=-n(n-1)\varphi(x^{2}). \tag*{(30)}

These are the first two Newton relations in elementary form. We do not need the general Newton identities.

Choose a finite number Rx≥0R_x\ge0 such that ∣λ∣≤Rx|\lambda|\le R_x for every λ∈σA(x)\lambda\in\sigma_A(x). By (27) and (30),

n(n−1)∣φ(x2)∣=∣∑j=1nλn,j2∣≤∑j=1n∣λn,j∣2≤nRx2.(31)n(n-1)|\varphi(x^{2})|=\left|\sum_{j=1}^{n}\lambda_{n,j}^{2}\right| \le\sum_{j=1}^{n}|\lambda_{n,j}|^{2}\le nR_x^{2}. \tag*{(31)}

Consequently,

∣φ(x2)∣≤Rx2n−1(n≥2).(32)|\varphi(x^{2})|\le\frac{R_x^{2}}{n-1}\qquad(n\ge2). \tag*{(32)}

The element xx and the bound RxR_x are fixed, while nn is arbitrary. Letting n→∞n\to\infty gives φ(x2)=0\varphi(x^{2})=0.

Proof of Theorem 6.1, and Proof II of GKZ. The preceding argument proves condition (i) of Lemma 4.3: φ(x)=0\varphi(x)=0 implies φ(x2)=0\varphi(x^{2})=0. The lemma therefore gives multiplicativity. In a Banach algebra every spectrum is bounded by (6); this proves Theorem 2.1. Automatic continuity is supplied separately by Proposition 3.3.

Within this proof, the only theorem about complex polynomials is the fundamental theorem of algebra: a monic polynomial of degree nn has nn complex roots counted with multiplicity. The coefficient comparisons and estimates are elementary algebra.

The fundamental theorem of algebra is often proved using complex analysis, for example by Liouville’s theorem. Other proofs exist. It is therefore accurate to say that the Roitman–Sternfeld argument needs no complex-analytic apparatus beyond the fundamental theorem of algebra. Completeness enters only if one first derives the bounded-spectrum hypothesis from the Banach-algebra axioms.

Third proof: Pólya–Saxer and zero-free derivatives

The first proof controls an entire function by a bound on its growth. The proof given in [11] instead controls it through the absence of zeros of its derivatives. The starting exponential is the same, but the rigidity theorem is different.

The exponential function and absence of zeros

Theorem 7.1 (Pólya–Saxer). Let FF be entire. If FF, F′F', and F′′F'' have no zeros in C\mathbb{C}, then

F(z)=eαz+βF(z) = e^{\alpha z+\beta}

for constants α∈C∖{0}\alpha\in\mathbb{C} \setminus\{0\} and β∈C\beta\in\mathbb{C}.

The statement and the historical account below are recalled in [11 Section 2], which refers to Burckel [17 Chapter 12] for the history and complex-analytic background. The theorem was stated by Pólya in the 1920s and proved by Saxer, Varopoulos and Csillag. It is considerably deeper than the coefficient estimate of § 5. We use it without proof; no growth hypothesis appears in its statement.

In the application below, every derivative of FF is zero-free. It would therefore suffice to quote the weaker version of the theorem, in which the derivatives of all orders, including order zero, are assumed to be zero-free; this is the version used explicitly in the proof in [11]. The stronger statement, which involves only FF and its first two derivatives, makes clear that the remaining derivatives are not needed for this route to GKZ.

In fact, fewer non-vanishing conditions suffice. The following theorem of Tumura and Clunie is also stated in [11 Section 2]; for n=2n = 2, a short proof was given by Yang [48].

Theorem 7.2 (Tumura–Clunie). Let FF be entire and let n≥2n \ge2. If FF and F(n)F^{(n)} have no zeros in C\mathbb{C}, then

F(z)=eαz+βF(z) = e^{\alpha z+\beta}

for constants α∈C∖{0}\alpha\in\mathbb{C} \setminus\{0\} and β∈C\beta\in\mathbb{C}.

The case n=2n = 2 strengthens Theorem 7.1, since the hypothesis on F′F' is unnecessary. Proof III therefore only requires that FF and F′′F'' be zero-free. The example below shows that a condition on a higher derivative cannot simply be omitted: it is not enough that FF and F′F' be zero-free.

Why non-vanishing of F and F' alone is insufficient

Consider

F(z)=exp⁡(ez−1).(33)F(z) = \exp(e^z - 1). \tag*{(33)}

Then F(0)=1F(0) = 1, FF has no zeros, and

F′(z)=ezexp⁡(ez−1)F'(z) = e^z\exp(e^z - 1)

also has no zeros. Nevertheless, FF is not eαze^{\alpha z} for a constant α\alpha, since F′/F=ezF'/F = e^z is not constant. Its second derivative is

F′′(z)=ez(1+ez)exp⁡(ez−1),F''(z) = e^z(1 + e^z)\exp(e^z - 1),

which vanishes at z=(2k+1)πiz = (2k + 1)\pi i, k∈Zk \in\mathbb{Z}. Thus the second-derivative condition imposes a genuine restriction.

The example also has much faster growth than allowed by Lemma 5.1: along the positive real axis, log⁡∣F(r)∣=er−1\log|F(r)| = e^r - 1, which cannot be bounded by log⁡C+Tr\log C + Tr. The two analytic lemmas therefore exclude it for different reasons.

Step 1: start with an invertible algebra element

Assume (2), and let x∈A−1x \in A^{-1}. Set

Fx(z)=φ(ezx).F_x(z)=\varphi(e^{zx}).

Automatic continuity and the identities (9) and (10) show that this function is entire and that

Fx(k)(z)=φ(xkezx)≠0(k≥0, z∈C).(34)F_x^{(k)}(z)=\varphi(x^k e^{zx})\ne0 \qquad(k\ge0,\ z\in\mathbb{C}). \tag*{(34)}

Indeed, both xkx^k and ezxe^{zx} are invertible. Their product is invertible as well, with inverse e−zxx−ke^{-zx}x^{-k}. The hypothesis on φ\varphi now gives (34).

The restriction x∈A−1x\in A^{-1} is essential at this step. For arbitrary a∈Aa\in A, ezae^{za} is still invertible, but aezaae^{za} need not be. Thus the hypothesis guarantees Fa≠0F_a\ne0, but it does not automatically guarantee Fa′≠0F_a'\ne0.

Step 2: apply Pólya–Saxer

By (34), Theorem 7.1 gives Fx(z)=eαz+βF_x(z)=e^{\alpha z+\beta}. Since Fx(0)=1F_x(0)=1, we have eβ=1e^\beta=1, and hence Fx(z)=eαzF_x(z)=e^{\alpha z}. Differentiation at zero yields

α=Fx′(0)=φ(x),φ(x2)=Fx′′(0)=φ(x)2.\alpha=F_x'(0)=\varphi(x),\qquad\varphi(x^2)=F_x''(0)=\varphi(x)^2.

We have therefore proved

φ(x2)=φ(x)2(x∈A−1).(35)\varphi(x^2)=\varphi(x)^2\qquad(x\in A^{-1}). \tag*{(35)}

The non-vanishing hypothesis also gives φ(x)≠0\varphi(x)\ne0, consistently with α≠0\alpha\ne0 in Theorem 7.1. If one prefers to quote Theorem 7.2 with n=2n=2, only the cases k=0k=0 and k=2k=2 of (34) are used here.

Step 3: a scalar shift removes the invertibility restriction

Take any a∈Aa\in A. Choose λ∈C\lambda\in\mathbb{C} with ∣λ∣>∥a∥|\lambda|>\|a\|. The Neumann series shows that x=a−λ1x=a-\lambda1 is invertible. Apply (35) to xx:

φ((a−λ1)2)=φ(a−λ1)2.\varphi((a-\lambda1)^2)=\varphi(a-\lambda1)^2.

The left side is

φ(a2)−2λφ(a)+λ2,\varphi(a^2)-2\lambda\varphi(a)+\lambda^2,

and the right side is

(φ(a)−λ)2=φ(a)2−2λφ(a)+λ2.(\varphi(a)-\lambda)^2=\varphi(a)^2-2\lambda\varphi(a)+\lambda^2.

Cancelling the last two terms on each side gives

φ(a2)=φ(a)2.(36)\varphi(a^2)=\varphi(a)^2. \tag*{(36)}

In the language of Lemma 4.1(i), this says that dφ(a)=dφ(a−λ1)=0d_\varphi(a)=d_\varphi(a-\lambda1)=0.

Proof III of Theorem 2.1. Proposition 3.3 gives continuity. For invertible xx, the derivative argument and Pólya–Saxer give (35). For arbitrary aa, the scalar-shift argument gives (36). Lemma 4.3 now proves multiplicativity.

What the theorem is useful for

Hyperplanes consisting of non-invertible elements

Corollary 8.1. Let AA be a unital complex Banach algebra, and let H⊂AH \subset A be a linear subspace of algebraic codimension one. The following are equivalent:

  • (i) H∩A−1=∅H \cap A^{-1} = \varnothing;

  • (ii) H=ker⁡φH = \ker\varphi for a character φ\varphi;

  • (iii) HH is a proper two-sided ideal.

In this case HH is closed, A/HA/H is isomorphic to C\mathbb{C}, and HH is a maximal proper two-sided ideal. It is also a maximal proper left ideal and a maximal proper right ideal.

Proof. Assume (i). Since 11 is invertible, 1∉H1 \notin H. Codimension one gives the algebraic direct sum A=H⊕C1A = H \oplus\mathbb{C}1. Define φ(h+λ1)=λ\varphi(h+\lambda1)=\lambda. This is linear, normalised, and has kernel HH. If uu is invertible, then u∉Hu \notin H, so φ(u)≠0\varphi(u) \ne0. GKZ gives (ii), including continuity of φ\varphi; its kernel is therefore closed.

A character’s kernel is a proper two-sided ideal, so (ii) implies (iii). A proper two-sided ideal cannot contain an invertible element, since u∈H∩A−1u \in H \cap A^{-1} would imply 1=u−1u∈H1=u^{-1}u \in H. Thus (iii) implies (i). The character induces an algebra isomorphism A/H→CA/H \to\mathbb{C}. Finally, any linear subspace properly containing a codimension-one subspace equals AA; this proves all the maximality assertions.

For commutative Banach algebras, this is the hyperplane characterisation of maximal ideals referred to in the title of Kahane and Żelazko’s paper [33]. The codimension-one hypothesis is essential. For example, in A=C2A=\mathbb{C}^{2} with coordinatewise operations, the zero subspace H={0}H=\{0\} satisfies (i) and (iii) but not (ii): it is not the kernel of a character, and it is not a maximal ideal, since it is properly contained in C×{0}\mathbb{C}\times\{0\}.

A non-vanishing test for point evaluation on C(X)C(X)

Let XX be a non-empty compact Hausdorff space. In C(X)C(X), with the supremum norm and pointwise multiplication, the invertible elements are exactly the nowhere-zero functions.

Corollary 8.2. Let Λ:C(X)→C\Lambda:C(X)\to\mathbb{C} be complex-linear, with Λ(1)=1\Lambda(1)=1. If Λ(f)≠0\Lambda(f)\ne0 whenever ff is nowhere zero, then there is a unique x0∈Xx_0\in X such that

Λ(f)=f(x0)(f∈C(X)).\Lambda(f)=f(x_0) \qquad(f\in C(X)).

Proof. By GKZ, Λ\Lambda is a character. To show that it is a point evaluation, put H=ker⁡ΛH=\ker\Lambda and suppose, for a contradiction, that the functions in HH have no common zero. For each x∈Xx\in X, choose fx∈Hf_x\in H with fx(x)≠0f_x(x)\ne0. The open sets where these functions are non-zero cover XX. Compactness supplies finitely many f1,…,fm∈Hf_1,\ldots,f_m\in H with no common zero. Then

g=∑j=1mfj‾fjg=\sum_{j=1}^{m}\overline{f_j}f_j

is strictly positive everywhere, hence invertible in C(X)C(X). Since HH is an ideal, g∈Hg\in H, a contradiction. There is therefore a point x0x_0 at which every element of HH vanishes.

For any f∈C(X)f\in C(X), the function f−Λ(f)1f-\Lambda(f)1 lies in HH. Evaluation at x0x_0 gives f(x0)=Λ(f)f(x_0)=\Lambda(f). Uniqueness follows because continuous functions separate distinct points of a compact Hausdorff space.

Thus the non-vanishing condition singles out point evaluations among normalised linear functionals. In particular, a non-trivial convex combination of evaluations at distinct points must vanish on some nowhere-zero complex-valued function.

Linear maps preserving invertibility

Proposition 8.3. Let AA, BB be unital complex Banach algebras, with BB commutative and semisimple. Suppose T:A→BT:A\to B is complex-linear, T(1A)=1BT(1_A)=1_B, and T(u)T(u) is invertible in BB whenever uu is invertible in AA. Then TT is multiplicative.

Proof. For each character χ\chi of BB, the functional χ∘T\chi\circ T is linear, normalised, and non-zero on every invertible element of AA. By GKZ,

χ(T(ab))=χ(T(a))χ(T(b))=χ(T(a)T(b)).\chi(T(ab))=\chi(T(a))\chi(T(b))=\chi(T(a)T(b)).

Thus every character of BB vanishes on T(ab)−T(a)T(b)T(ab)-T(a)T(b). In a commutative semisimple Banach algebra the characters separate points, so this difference is zero.

The proof illustrates a useful method: compose an algebra-valued map with enough characters to separate points in its range. If BB is commutative but not semisimple, the same argument shows only that T(ab)−T(a)T(b)T(ab)-T(a)T(b) belongs to its Jacobson radical, the intersection of the kernels of its characters. For a non-commutative target, characters need not separate points and may not exist at all. The argument therefore does not settle the corresponding questions for non-commutative target algebras.

The Kaplansky–Aupetit conjecture asks for the following extension [16]. Let AA, BB be unital complex Banach algebras, with BB semisimple, and let T:A→BT:A\to B be a surjective complex-linear map satisfying T(1A)=1BT(1_A)=1_B. If T(u)T(u) is invertible whenever uu is invertible, must TT be a Jordan homomorphism, that is, must

T(a2)=T(a)2(a∈A)?T(a^2)=T(a)^2 \qquad(a\in A)?

When BB is commutative, Proposition 8.3 gives the stronger conclusion of multiplicativity without requiring surjectivity. For a non-commutative target, matrix transposition shows why multiplicativity itself cannot be expected.

Weighted composition on spaces of continuous functions

Applying Corollary 8.2 at each point of the target space gives a representation for linear maps that preserve nowhere-zero functions.

Corollary 8.4. Let XX, YY be non-empty compact Hausdorff spaces. Suppose a complex-linear map T:C(X)→C(Y)T:C(X)\to C(Y) takes every nowhere-zero function to a nowhere-zero function. Then there are a nowhere-zero function ψ∈C(Y)\psi\in C(Y) and a continuous map θ:Y→X\theta:Y\to X such that

(Tf)(y)=ψ(y)f(θ(y))(f∈C(X), y∈Y).(Tf)(y)=\psi(y)f(\theta(y)) \qquad(f\in C(X),\ y\in Y).

Proof. Set ψ=T1\psi=T1, which is nowhere zero. For each y∈Yy\in Y, the map

Λy(f)=(Tf)(y)ψ(y)\Lambda_y(f)=\frac{(Tf)(y)}{\psi(y)}

is linear, normalised, and non-zero on every nowhere-zero function. Corollary 8.2 gives a unique point θ(y)∈X\theta(y)\in X with Λy(f)=f(θ(y))\Lambda_y(f)=f(\theta(y)). This proves the displayed formula. For every f∈C(X)f\in C(X), the function f∘θ=Tf/ψf\circ\theta=Tf/\psi is continuous. Since the topology of the compact Hausdorff space XX is the weakest topology making every f∈C(X)f\in C(X) continuous, θ\theta is continuous.

Continuity of TT follows from the representation: ∥Tf∥∞≤∥ψ∥∞∥f∥∞\lVert T f\rVert_\infty\le\lVert\psi\rVert_\infty\lVert f\rVert_\infty. In fact, ∥T∥=∥ψ∥∞\lVert T\rVert= \lVert\psi\rVert_\infty, since equality is attained at f=1f=1.

The study of composition operators and their weighted counterparts, particularly on spaces of holomorphic functions, is an active area of research in operator theory. For a recent authoritative monograph, see [36].

Why the hypotheses matter

The real-field version fails

Example 9.1. Take the real Banach algebra A=C([0,1],R)A=C([0,1],\mathbb{R}) and define

φ(f)=∫01f(t) dt\varphi(f)=\int_0^1 f(t)\,\mathrm{d}t

Then φ(1)=1\varphi(1)=1. Every invertible f∈Af\in A is continuous and nowhere zero; connectedness of [0,1][0,1] forces it to have a fixed sign. Consequently φ(f)≠0\varphi(f)\ne0. Nevertheless, for a(t)=ta(t)=t,

φ(a2)=13≠14=φ(a)2.\varphi(a^2)=\frac{1}{3}\ne\frac{1}{4}=\varphi(a)^2.

Thus the real analogue of GKZ is false.

This example appears in the Kahane–Żelazko paper [33]. On the complex algebra C([0,1],C)C([0,1],\mathbb{C}) the same integral does not satisfy the hypothesis: f(t)=e2πitf(t)=e^{2\pi i t} is nowhere zero but has integral zero. The values of a complex-valued nowhere-zero function can therefore cancel under integration, whereas those of a continuous real-valued function of fixed sign cannot.

An incomplete normed algebra where the conclusion fails

Example 9.2. Let A=C[t]A=\mathbb{C}[t], the polynomial algebra, equipped with

∥p∥=max⁡∣z∣≤1∣p(z)∣.\lVert p\rVert=\max_{|z|\le1}|p(z)|.

This is a unital normed algebra, but it is not complete. Its only invertible elements are the non-zero constant polynomials. Define

φ(p)=p(1)+p(−1)2.\varphi(p)=\frac{p(1)+p(-1)}{2}.

Then φ(1)=1\varphi(1)=1, ∥φ∥=1\lVert\varphi\rVert=1, and φ\varphi is non-zero on all invertible elements of AA. Yet

φ(t)=0,φ(t2)=1,\varphi(t)=0,\qquad\varphi(t^2)=1,

so φ\varphi is not multiplicative.

For a non-constant polynomial pp, every p−λp-\lambda remains non-constant and is non-invertible in C[t]\mathbb{C}[t]. Hence

σC[t](p)=C.\sigma_{\mathbb{C}[t]}(p)=\mathbb{C}.

Thus the bounded-spectrum hypothesis of Theorem 6.1 fails. Completeness can be replaced by a suitable spectral hypothesis, as Roitman and Sternfeld showed; see also Corollary 9.3 below. The example shows why some such replacement is needed.

Why passage to the completion needs care

The completion in Example 9.2 is the disc algebra, consisting of the functions that are continuous on the closed unit disc and holomorphic in its interior. The bounded functional φ\varphi extends to it by the same averaging formula. However, the completion contains many new invertible elements. For example,

f(z)=eiπz/2f(z)=e^{i\pi z/2}

belongs to the disc algebra and has an inverse there, whereas

f(1)+f(−1)2=i−i2=0.\frac{f(1)+f(-1)}{2}=\frac{i-i}{2}=0.

Thus the extended functional does not satisfy the GKZ non-vanishing hypothesis. Continuity of a functional is not enough to transfer that hypothesis from an incomplete algebra to its completion. An argument using the completion must therefore account for the relation between the two sets of invertible elements, or between their spectra.

A positive counterpart: normed Q-algebras and spectral norms

Theorem 6.1 applies whenever the spectra are bounded, even if completeness is unavailable. The following two conditions ensure this boundedness.

Corollary 9.3. Let AA be a unital complex algebra and let φ:A→C\varphi:A\to\mathbb{C} be a linear functional satisfying (2.1). Suppose that one of the following holds.

(a) AA carries a norm, not necessarily complete or submultiplicative, such that A−1A^{-1} contains a ball {u:∥u−1∥<ε}\{u:\lVert u-1\rVert<\varepsilon\} for some ε>0\varepsilon>0. This is the case for every normed QQ-algebra, that is, whenever A−1A^{-1} is open, and for every Banach algebra (with ε=1\varepsilon=1, by Lemma 3.2).

(b) AA carries a function p:A→[0,∞)p:A\to[0,\infty) with ∣λ∣≤p(x)|\lambda|\leq p(x) for all x∈Ax\in A and λ∈σA(x)\lambda\in\sigma_A(x); for example, a spectral seminorm in the sense of [41].

Then φ\varphi is multiplicative.

Proof. In case (b), every spectrum is bounded by hypothesis. In case (a), let x∈Ax\in A and ∣λ∣>∥x∥/ε|\lambda|>\lVert x\rVert/\varepsilon. Then ∥(1−x/λ)−1∥=∥x∥/∣λ∣<ε\lVert(1-x/\lambda)-1\rVert=\lVert x\rVert/|\lambda|<\varepsilon, so 1−x/λ∈A−11-x/\lambda\in A^{-1}, and therefore λ1−x=λ(1−x/λ)∈A−1\lambda1-x=\lambda(1-x/\lambda)\in A^{-1}. Hence σA(x)⊂{λ:∣λ∣≤∥x∥/ε}\sigma_A(x)\subset\{\lambda:|\lambda|\leq\lVert x\rVert/\varepsilon\}. Theorem 6.1 therefore applies in either case.

The GKZ theorem for normed QQ-algebras is discussed by Roitman and Sternfeld [43], and its extension to spectral algebras by Palmer [41]. A version for bounded functionals on algebras with a spectral norm is proved in [11 Section 3] by passing to the completion: the algebra is a full subalgebra of its completion, and the upper semicontinuity of the spectrum [5] transfers the relation φ(x)∈σA(x)\varphi(x)\in\sigma_A(x) to all elements of the completion. In Corollary 9.3 boundedness of φ\varphi is not assumed, because Theorem 6.1 does not need it.

Zero-free does not by itself mean exponential-linear

The entire function F(z)=ez2F(z)=e^{z^2} is zero-free but cannot be written as eαz+βe^{\alpha z+\beta}. It violates the exponential-type bound, because log⁡∣F(r)∣=r2\log|F(r)|=r^2 on the positive real axis. Since F′(0)=0F'(0)=0, it also fails the hypotheses of the Pólya–Saxer theorem. Likewise, the function in (33) is zero-free with zero-free first derivative, but its second derivative has zeros. These examples explain the roles of the growth and derivative hypotheses in the two analytic proofs.

A scalar selection need not exist

For n≥2n \ge2, the matrix algebra Mn(C)M_n(\mathbb{C}) has no characters. Here is a direct proof. Let EijE_{ij} be the usual matrix units, and suppose that χ\chi were a character. Each χ(Eii)\chi(E_{ii}) is idempotent in C\mathbb{C}, hence is 0 or 1, and their sum is χ(I)=1\chi(I)=1. There is an index ii with χ(Eii)=1\chi(E_{ii})=1, and an index j≠ij \ne i with χ(Ejj)=0\chi(E_{jj})=0. But

EijEji=Eii,EjiEij=Ejj,E_{ij}E_{ji}=E_{ii}, \qquad E_{ji}E_{ij}=E_{jj},

and multiplicativity would give both

χ(Eij)χ(Eji)=1andχ(Eji)χ(Eij)=0,\chi(E_{ij})\chi(E_{ji})=1 \qquad\text{and} \qquad\chi(E_{ji})\chi(E_{ij})=0,

a contradiction. By GKZ, no normalised linear functional on Mn(C)M_n(\mathbb{C}) can avoid zero on all invertible matrices. Thus, although each matrix has an eigenvalue, there is no linear functional that selects an eigenvalue from the spectrum of every matrix.

GKZ for Ransford’s generalised spectra

Of the three proofs, the polynomial argument of Roitman and Sternfeld offers the most elementary route to extensions involving Ransford’s generalised spectra. The key is to retain the two ingredients used in Proof II: suitable powers stay in the set on which the functional does not vanish, and the polynomial roots remain bounded. This leads to the extension announced in my 1991 note [8], which we present in the algebraic form given by Kulkarni and Sukumar [35]. Throughout this section, AA is an arbitrary unital complex algebra with 1≠01 \ne0; a norm will be introduced only where explicitly stated.

Replacing the invertible elements by a distinguished set

Let Ω⊂A\Omega\subset A satisfy the following three conditions:

  1. (i) 0∉Ω0 \notin\Omega;

  2. (ii) 1∈Ω1 \in\Omega;

  3. (iii) λΩ⊂Ω\lambda\Omega\subset\Omega for every non-zero λ∈C\lambda\in\mathbb{C}.

The third condition actually gives λΩ=Ω\lambda\Omega=\Omega: apply it also to λ−1\lambda^{-1}. In particular, −Ω=Ω-\Omega=\Omega.

Define the Ω\Omega-spectrum of a∈Aa \in A by

σΩ(a):={λ∈C:a−λ1∉Ω}.(37)\sigma_{\Omega}(a):=\{\lambda\in\mathbb{C}:a-\lambda1\notin\Omega\}. \tag*{(37)}

This abstract construction is due to Tom Ransford [42]. The definition and the first three conditions make sense even in a complex vector space with a distinguished non-zero vector 11; multiplication is not yet needed. They imply

σΩ(c1)={c}(c∈C),(38)\sigma_{\Omega}(c1)=\{c\} \qquad(c \in\mathbb{C}), \tag*{(38)}

because (c−λ)1(c-\lambda)1 belongs to Ω\Omega exactly when λ≠c\lambda\ne c.

Choosing Ω=A−1\Omega=A^{-1} recovers the ordinary spectrum σA(a)\sigma_A(a). For a general Ω\Omega, however, membership in Ω\Omega is a separate condition from invertibility. We assume no inclusion between Ω\Omega and A−1A^{-1}, and shall not use the ordinary spectrum to locate the roots in the argument below.

By imposing geometric conditions on Ω\Omega, notably pseudoconvexity, Ransford recovers a surprisingly large part of classical spectral theory in the setting of a complex normed space. Much of the usual theory thus survives without multiplication or completeness. As Ransford puts it [42 p. 306]: “Still included in the ‘grin after the Cheshire cat has disappeared’ are generalisations of many of the classical results of spectral theory, including non-emptiness of the spectrum, the Beurling-Gelfand spectral radius formula, Newburgh’s continuity theorem, and even an analogue of B. E. Johnson’s uniqueness-of-norm theorem in which all mention of representation theory has been eliminated.”

In a related direction, Zalduendo [49] found a geometric condition equivalent to commutativity in Banach algebras.

When multiplication is available, we impose one further condition:

(iv) there is a strictly increasing sequence of positive integers n1<n2<⋯n_1 < n_2 < \cdots such that

u∈Ω⟹unj∈Ω(j≥1).(39)u \in\Omega\quad\Longrightarrow\quad u^{n_j} \in\Omega\qquad(j \ge1). \tag*{(39)}

The same sequence is required to work for every u∈Ωu \in\Omega.

Only powers along this unbounded sequence are required; closure of Ω\Omega under arbitrary products is a stronger sufficient condition.

Lemma 10.1 (A sufficient condition for the power property). If Ω\Omega is closed under multiplication, then un∈Ωu^n \in\Omega for every u∈Ωu \in\Omega and every positive integer nn. In particular, (39) holds with nj=jn_j = j.

Proof. Starting with u∈Ωu \in\Omega, induction gives un+1=unu∈Ωu^{n+1} = u^n u \in\Omega whenever un∈Ωu^n \in\Omega.

Thus the ordinary group of invertible elements satisfies the required power property. The proof below only uses the implication in (39); it does not require its converse or closure under products of distinct elements.

In fact, the power condition is equivalent to the existence of a single integer m≥2m \ge2 such that u∈Ωu \in\Omega implies um∈Ωu^m \in\Omega for every uu. One direction follows by choosing any nj≥2n_j \ge2; for the other, iteration gives umj∈Ωu^{m^j} \in\Omega for all j≥1j \ge1. The exponent mm, like the sequence in (39), must work for every uu.

Boundedness and non-emptiness

Boundedness of the Ω\Omega-spectra will be a separate hypothesis in the algebraic theorem. A simple sufficient condition is available when AA has a norm, even without completeness or a submultiplicative norm.

Proposition 10.2 (An elementary spectral bound). Assume the first three conditions on Ω\Omega. If a norm on AA and a number ε>0\varepsilon> 0 satisfy

{u∈A:∥u−1∥<ε}⊂Ω,\{u \in A : \lVert u - 1 \rVert< \varepsilon\} \subset\Omega,

then

σΩ(a)⊂{λ∈C:∣λ∣≤∥a∥/ε}(a∈A).(40)\sigma_{\Omega}(a) \subset\{\lambda\in\mathbb{C} : |\lambda| \le\lVert a \rVert/ \varepsilon\} \qquad(a \in A). \tag*{(40)}

If Ω\Omega is open, every Ω\Omega-spectrum is also closed, and hence compact, possibly empty.

Proof. For ∣λ∣>∥a∥/ε|\lambda| > \lVert a \rVert/ \varepsilon,

∥(1−a/λ)−1∥=∥a∥∣λ∣<ε.\lVert(1-a/\lambda)-1\rVert= \frac{\lVert a\rVert}{|\lambda|} < \varepsilon.

Thus 1−a/λ∈Ω1-a/\lambda\in\Omega. Scalar invariance gives a−λ1=−λ(1−a/λ)∈Ωa-\lambda1=-\lambda(1-a/\lambda) \in\Omega, so λ∉σΩ(a)\lambda\notin\sigma_{\Omega}(a). This proves the bound. If Ω\Omega is open, then A∖ΩA \setminus\Omega is closed, and σΩ(a)\sigma_{\Omega}(a) is its inverse image under the continuous map λ↦a−λ1\lambda\mapsto a-\lambda1; hence σΩ(a)\sigma_{\Omega}(a) is closed. Moreover, openness and 1∈Ω1 \in\Omega supply a suitable ε>0\varepsilon> 0, so this closed set is also bounded.

Openness does not ensure non-emptiness. For example, give A=C2A = \mathbb{C}^2 coordinatewise multiplication, with 1=(1,1)\mathbf{1} = (1,1), and put Ω=A∖{0}\Omega= A \setminus\{0\}. This set is open and satisfies all four conditions, since a non-zero coordinate remains non-zero under positive integer powers. Nevertheless,

(1,0)−λ(1,1)=(1−λ,−λ)≠(0,0)(λ∈C),(1,0) - \lambda(1,1) = (1-\lambda,-\lambda) \ne(0,0) \qquad(\lambda\in\mathbb{C}),

so σΩ((1,0))=∅\sigma_{\Omega}((1,0)) = \varnothing. More generally, the Ω\Omega-spectrum of (z1,z2)(z_1,z_2) is empty if z1≠z2z_1 \ne z_2 and is {z1}\{z_1\} otherwise. Thus all these spectra are bounded. This example also shows that the power property need not imply multiplicative closure: (1,0),(0,1)∈Ω(1,0),(0,1) \in\Omega, but their product is zero.

Non-vanishing and spectral selection

Lemma 10.3 (The Ω\Omega-spectral reformulation). Assume the first three conditions on Ω\Omega. For a complex-linear functional φ:A→C\varphi:A \to\mathbb{C} with φ(1)=1\varphi(1)=1, the following conditions are equivalent:

  1. φ(u)≠0\varphi(u) \ne0 for every u∈Ωu \in\Omega;

  2. φ(a)∈σΩ(a)\varphi(a) \in\sigma_{\Omega}(a) for every a∈Aa \in A.

Proof. If (i) holds, then φ(a−φ(a)1)=0\varphi(a-\varphi(a)1)=0 forces a−φ(a)1∉Ωa-\varphi(a)1 \notin\Omega. By definition, φ(a)∈σΩ(a)\varphi(a) \in\sigma_{\Omega}(a). Conversely, if u∈Ωu \in\Omega, then 0∉σΩ(u)0 \notin\sigma_{\Omega}(u). Condition (ii) places φ(u)\varphi(u) in this spectrum, so φ(u)≠0\varphi(u) \ne0.

In particular, the existence of such a functional forces every Ω\Omega-spectrum to be non-empty. It also rules out the preceding example. Notice that spectral selection itself forces the normalisation: applying it to 11 and using (38) gives φ(1)=1\varphi(1)=1.

The theorem and its polynomial proof

Theorem 10.4 ([8, 35]). Let AA be a unital complex algebra, and let Ω⊂A\Omega\subset A satisfy the four conditions above. Suppose that σΩ(a)\sigma_{\Omega}(a) is bounded for every a∈Aa \in A. If a complex-linear functional φ:A→C\varphi:A \to\mathbb{C} satisfies

φ(1)=1,φ(u)≠0(u∈Ω),\varphi(1)=1,\qquad\varphi(u)\ne0 \quad(u\in\Omega),

then φ\varphi is multiplicative. Equivalently, every complex-linear selection of the Ω\Omega-spectra is a character.

Proof. By Lemma 4.3, it suffices to prove φ(x2)=0\varphi(x^2)=0 for x∈ker⁡φx \in\ker\varphi. Fix such an xx and use the polynomials PnP_n from (24), with n=nj≥2n=n_j \ge2. If λ∉σΩ(x)\lambda\notin\sigma_{\Omega}(x), then λ1−x∈Ω\lambda1-x \in\Omega, since −Ω=Ω-\Omega=\Omega. The power condition (39) gives (λ1−x)nj∈Ω(\lambda1-x)^{n_j}\in\Omega, and non-vanishing of φ\varphi on Ω\Omega yields Pnj(λ)≠0P_{n_j}(\lambda)\ne0. Thus every root belongs to σΩ(x)\sigma_{\Omega}(x), a bounded set. Choose a finite bound MxM_x for the moduli of these spectral values. The coefficient identity (30) and the estimate (31) from Proof II now apply with MxM_x in place of RxR_x:

∣φ(x2)∣≤Mx2nj−1⟶0(j→∞).\lvert\varphi(x^2)\rvert\le\frac{M_x^2}{n_j-1}\longrightarrow0 \qquad(j\to\infty).

The Roitman–Sternfeld argument therefore extends by replacing ordinary spectra with Ω\Omega-spectra and using only the exponents njn_j. It requires no continuity, completeness or analytic growth theorem; the role of the fundamental theorem of algebra is exactly as in Section 6. For a fixed φ\varphi, boundedness of σΩ(x)\sigma_{\Omega}(x) is needed only for x∈ker⁡φx\in\ker\varphi.

Corollary 10.5 (The normed case and automatic continuity). Suppose AA has a norm and Ω\Omega satisfies the four conditions above. If

{u:∥u−1∥<ε}⊂Ωfor some ε>0,\{u : \lVert u-1\rVert< \varepsilon\} \subset\Omega\qquad\text{for some }\varepsilon>0,

then every normalised complex-linear functional that is non-zero on Ω\Omega is multiplicative and continuous, with

∣φ(a)∣≤∥a∥ε,∥φ∥≤1ε.|\varphi(a)| \le\frac{\lVert a\rVert}{\varepsilon}, \qquad\lVert\varphi\rVert\le\frac{1}{\varepsilon}.

Proof. Proposition 10.2 supplies bounded spectra, so Theorem 10.4 gives multiplicativity. The same bound, together with φ(a)∈σΩ(a)\varphi(a)\in\sigma_{\Omega}(a) from Lemma 10.3, gives the stated norm estimate. Thus continuity could also be established before multiplicativity. If Ω\Omega is open, a suitable ε\varepsilon exists because 1∈Ω1\in\Omega.

Why the converse need not hold

For the ordinary spectrum, every character is a spectral selection. For an arbitrary Ω\Omega, this converse can fail. On A=C2A=\mathbb{C}^{2} with coordinatewise multiplication and 1=(1,1)\mathbf{1}=(1,1), put

Ω={(z1,z2):z1≠0}.\Omega=\{(z_{1},z_{2}):z_{1}\ne0\}.

This set is open, satisfies the first three conditions and is closed under multiplication, so it satisfies the power condition as well. Directly from the definition,

σΩ((z1,z2))={z1}.\sigma_{\Omega}((z_{1},z_{2}))=\{z_{1}\}.

The second coordinate functional ψ(z1,z2)=z2\psi(z_{1},z_{2})=z_{2} is a character, but (1,0)∈Ω(1,0)\in\Omega and ψ(1,0)=0\psi(1,0)=0. Thus ψ\psi neither avoids zero on Ω\Omega nor selects its spectra. The first coordinate functional does both. In this example,

A−1={(z1,z2):z1z2≠0}⊊Ω,σA((z1,z2))={z1,z2},A^{-1}=\{(z_{1},z_{2}):z_{1}z_{2}\ne0\}\subsetneq\Omega,\qquad\sigma_{A}((z_{1},z_{2}))=\{z_{1},z_{2}\},

which makes the distinction between the two spectra explicit.

Extensions and further reading

The following references provide starting points for further reading.

  • Other proofs. In addition to Siddiqi’s elementary analytic proof [44], discussed earlier, Wille [47] gives two proofs for commutative Banach algebras with a symmetric involution. Palmer’s book [41] provides a broader account of the theorem.

  • Beyond Banach algebras. For the extensions to normed Q-algebras and spectral algebras illustrated by Corollary 9.3, see [43, 41] and [11 Section 3].

  • Maps between algebras. Proposition 8.3 introduces the study of linear maps that preserve invertibility. Related generalisations of GKZ were proved by Aupetit [4] and Christensen [18].

  • Weaker hypotheses. The article [22] in the Encyclopaedia of Mathematics lists further directions, including versions in which linearity of the functional is weakened and extensions to maps between topological algebras.

  • Recent work on reproducing kernel Hilbert spaces. Chu, Hartz, Mashreghi and Ransford [19] extend GKZ to a reproducing kernel Hilbert space with a normalised complete Pick kernel. More recently, Bhattacharyya, Jaikishan, Kumar and Sahu [14] give tests involving only reproducing kernels.

For further references on the Gleason–Kahane–Żelazko theorem, we encourage the reader to consult MathSciNet [1] or its freely accessible European counterpart, zbMATH Open.

References

  1. [1]MathSciNet, American Mathematical Society. The Collaboration Distance tool is available at https://mathscinet.ams.org/mathscinet/freetools/collab-dist (accessed 25 October 2026).
  2. [2]R. André-Jeannin, Irrationalité de la somme des inverses de certaines suites récurrentes, Comptes rendus de l’Académie des sciences, Série I, Mathématique 308 (1989), no. 19, 539–541.
  3. [3]R. Arens, On a theorem of Gleason, Kahane and Żelazko, Studia Mathematica 87 (1987), no. 2, 193–196.DOI
  4. [4]B. Aupetit, Une généralisation du théorème de Gleason–Kahane–Żelazko pour les algèbres de Banach, Pacific Journal of Mathematics 85 (1979), no. 1, 11–17.DOI
  5. [5]B. Aupetit, Propriétés spectrales des algèbres de Banach, Lecture Notes in Mathematics 735, Springer-Verlag, Berlin, 1979.DOI
  6. [6]A. Bach, Quantum-mechanical aspects of the theorem of Gleason, Kahane, Zelazko, Lettere al Nuovo Cimento 35 (1982), 377–380.DOI
  7. [7]C. Badea, The irrationality of certain infinite series, Glasgow Mathematical Journal 29 (1987), no. 2, 221–228.DOI
  8. [8]C. Badea, The Gleason–Kahane–Żelazko theorem and Ransford’s generalised spectra, C. R. Acad. Sci. Paris Sér. I Math. 313 (1991), no. 10, 679–683.
  9. [9]C. Badea, A theorem on irrationality of infinite series and applications, Acta Arithmetica 63 (1993), no. 4, 313–323.DOI
  10. [10]C. Badea, The Gleason–Kahane–Żelazko theorem, Rendiconti del Circolo Matematico di Palermo (2), Suppl. no. 33 (1993), 177–188.
  11. [11]C. Badea, Sur les caractères d’une algèbre de Banach, Canadian Mathematical Bulletin 40 (1997), no. 2, 129–132.DOI
  12. [12]C. Badea and I. Raşa, The Gleason–Kahane–Żelazko property and Korovkin systems in symmetric involutive Banach algebras, Archiv der Mathematik 61 (1993), no. 2, 163–169.DOI
  13. [13]P.-G. Becker and T. Töpfer, Transcendency results for sums of reciprocals of linear recurrences, Mathematische Nachrichten 168 (1994), 5–17.DOI
  14. [14]T. Bhattacharyya, Jaikishan, P. Kumar and C. K. Sahu, Multiplicative linear functionals on reproducing kernel Hilbert spaces, preprint, arXiv:2605.21808v2 [math.FA], 2 September 2026.arxiv.org/abs/2605.21808
  15. [15]T. F. Bloom, Erdős Problem #267, https://www.erdosproblems.com/267. See also the accompanying discussion: https://www.erdosproblems.com/forum/thread/267.
  16. [16]M. Brešar and P. Šemrl, An extension of the Gleason–Kahane–Żelazko theorem: a possible approach to Kaplansky’s problem, Expositiones Mathematicae 26 (2008), no. 3, 269–277.DOI
  17. [17]R. B. Burckel, An Introduction to Classical Complex Analysis, Vol. I, Academic Press, New York, 1979.DOI
  18. [18]E. Christensen, Two generalizations of the Gleason–Kahane–Żelazko theorem, Pacific Journal of Mathematics 177 (1997), no. 1, 27–32.DOI
  19. [19]C. Chu, M. Hartz, J. Mashreghi and T. Ransford, A Gleason–Kahane–Żelazko theorem for reproducing kernel Hilbert spaces, Bulletin of the London Mathematical Society 54 (2022), no. 3, 1120–1130.arxiv.org/abs/2011.03360
  20. [20]J. Dixmier, P. Erdős and J.-L. Nicolas, Sur le nombre d’invariants fondamentaux des formes binaires, C. R. Acad. Sci. Paris Sér. I Math. 305 (1987), no. 8, 319–322.
  21. [21]J. Dixmier, J.-P. Kahane and J.-L. Nicolas, Un exemple de non-dérivabilité en géométrie du triangle, L’Enseignement Mathématique (2) 53 (2007), nos. 3–4, 369–428.
  22. [22]Encyclopaedia of Mathematics, article “Gleason–Kahane–Żelazko theorem”. encyclopaediaofmath.org.
  23. [23]P. Erdős and R. L. Graham, Old and new problems and results in combinatorial number theory, Monographies de L’Enseignement Mathématique, no. 28, L’Enseignement Mathématique, Université de Genève, Geneva, 1980, especially pp. 64–65.
  24. [24]The Formal Conjectures Authors, Erdős Problem 267, Formal Conjectures, Lean statement and external proof record, https://github.com/google-deepmind/formal-conjectures/blob/main/FormalConjectures/ErdosProblems/267.lean (checked 25 October 2026).
  25. [25]A. M. Gleason, Measures on the closed subspaces of a Hilbert space, Journal of Mathematics and Mechanics 6 (1957), 885–893.
  26. [26]A. M. Gleason, A characterization of maximal ideals, Journal d’Analyse Mathématique 19 (1967), 171–172.DOI
  27. [27]I. J. Good, A reciprocal series of Fibonacci numbers, Fibonacci Quarterly 12 (1974), no. 4, 346.DOI
  28. [28]J. W. Grossman, The Erdős Number Project, https://sites.google.com/oakland.edu/grossman/home/the-erdos-number-project (accessed 25 October 2026).
  29. [29]V. E. Hoggatt, Jr. and M. Bicknell, A reciprocal series of Fibonacci numbers with subscripts 2^k, Fibonacci Quarterly 14 (1976), no. 5, 453–455.DOI
  30. [30]K. Jarosz, Multiplicative functionals and entire functions, I and II, Studia Math. 119 (1996), 289–297 and Studia Math. 124 (1997), 193–198.DOI
  31. [31]J.-P. Kahane, Aperçu sur l’influence de l’école mathématique polonaise 1918–1939, Centre scientifique de l’Académie polonaise des sciences, Paris, 1992. Polish translation by R. Duda: Próba oceny wpływu polskiej szkoły matematycznej lat 1918–1939, Wiadomości Matematyczne 31 (1995), 163–175.
  32. [32]J.-P. Kahane, Bernoulli convolutions and self-similar measures after Erdős, in Reflections on Paul Erdős on His Birth Centenary, Notices of the American Mathematical Society 62 (2015), no. 2, 136–140.
  33. [33]J.-P. Kahane and W. Żelazko, A characterization of maximal ideals in commutative Banach algebras, Studia Mathematica 29 (1968), 339–343.DOI
  34. [34]S. Kowalski and Z. Słodkowski, A characterization of multiplicative linear functionals in Banach algebras, Studia Mathematica 67 (1980), no. 3, 215–223.
  35. [35]S. H. Kulkarni and D. Sukumar, Gleason–Kahane–Żelazko theorem for spectrally bounded algebra, International Journal of Mathematics and Mathematical Sciences 2005 (2005), no. 15, 2447–2460.DOI
  36. [36]P. Lefèvre and H. Queffélec, A Primer, and Beyond, on Composition Operators on the Unit Disk, New Mathematical Monographs 53, Cambridge University Press, Cambridge, 2026.DOI
  37. [37]É. Lucas, Théorie des fonctions numériques simplement périodiques, American Journal of Mathematics 1 (1878), 184–240.
  38. [38]K. Mahler, On the transcendence of the solutions of a special class of functional equations, Bulletin of the Australian Mathematical Society 13 (1975), no. 3, 389–410. corrigendum, 14 (1976), no. 3, 477–478.
  39. [39]M. Mignotte, Quelques problèmes d’effectivité en théorie des nombres, doctoral thesis, Université Paris XIII, 1974.
  40. [40]K. D. Nguyen, Transcendental series of reciprocals of Fibonacci and Lucas numbers, Algebra & Number Theory 16 (2022), no. 7, 1627–1654.arxiv.org/abs/2009.02446
  41. [41]T. W. Palmer, Banach Algebras and the General Theory of *-Algebras, Vol. I: Algebras and Banach Algebras, Encyclopedia of Mathematics and its Applications 49, Cambridge University Press, Cambridge, 1994.
  42. [42]T. J. Ransford, Generalised spectra and analytic multivalued functions, Journal of the London Mathematical Society (2) 29 (1984), no. 2, 306–322.DOI
  43. [43]M. Roitman and Y. Sternfeld, When is a linear functional multiplicative?, Transactions of the American Mathematical Society 267 (1981), no. 1, 111–124.DOI
  44. [44]J. A. Siddiqi, On a characterization of maximal ideals, Canadian Mathematical Bulletin 13 (1970), no. 2, 219–220.DOI
  45. [45]Star Fleet Math, Irrationality of lacunary Fibonacci reciprocal sums, online report and Lean proof for Erdős Problem 267, 2026. https://www.starfleetmath.com/, section “Erdős Problem #267” (accessed 25 October 2026).
  46. [46]E. C. Titchmarsh, The Theory of Functions, Oxford University Press, Oxford, 1939.
  47. [47]R. Wille, The theorem of Gleason–Kahane–Żelazko in a commutative symmetric Banach algebra, Mathematische Zeitschrift 190 (1985), 301–304.DOI
  48. [48]C.-C. Yang, On the zeros of an entire function and its second derivative, Atti della Accademia Nazionale dei Lincei, Rendiconti, Classe di Scienze Fisiche, Matematiche e Naturali (8) 49 (1970), nos. 1–2, 27–29.
  49. [49]I. Zalduendo, A geometric condition equivalent to commutativity in Banach algebras, Studia Mathematica 94 (1989), no. 2, 187–192.DOI
  50. [50]W. Żelazko, A characterization of multiplicative linear functionals in complex Banach algebras, Studia Mathematica 30 (1968), 83–85.DOI

Paper details

vlnc:W_hagrm6w3
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents