betaFeedback

WorksMathematics

On extremal problems related to the Krzyż conjecture

Abstract

We study general extremal problems similar or closely related to the Krzyż conjecture and establish the structural properties of extremal functions, pointing out some differences with the original Krzyż case. We also introduce a family of more specific problems of this kind, interpolating between the Krzyż problem and the problem of maximizing the constant term, and show that they are closely related to the Krzyż conjecture: the conjecture implies the analogous statement for all of them and, conversely, if one of them has the same extremal value as the Krzyż problem, then this value is 2/e2/e. This may represent a new approach to attacking the original problem. Finally, we show that for these problems an extremal function whose Taylor coefficients a1,…,a⌊n/3⌋a_1,\ldots,a_{\lfloor n/3\rfloor} vanish must be the conjectured one; as a special case, we obtain a new proof of a recent result of MacDonald.

Introduction

We study extremal problems for bounded non-vanishing analytic functions in the unit disc D\mathbb{D} that generalize the well-known Krzyż problem. Denote by B∗\mathcal{B}_* the class of all analytic functions ff in the unit disc D\mathbb{D} such that 0<∣f(z)∣≤10<|f(z)|\leq1 for all zz in D\mathbb{D}. If f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j in D\mathbb{D}, let

Mn=sup⁡{Re⁡an:f∈B∗},n≥1.(1)M_n=\sup\{\operatorname{Re}a_n:f\in\mathcal{B}_*\},\qquad n\geq1. \tag*{(1)}

Standard arguments involving normal families show that the supremum is attained for each n≥1n\geq1, hence an extremal function always exists. Since instead of f(z)f(z) we can consider μf(λz)\mu f(\lambda z) with ∣λ∣=∣μ∣=1|\lambda|=|\mu|=1, after such rotations we may assume from the start that both a0a_0 and ana_n are real and positive. The celebrated Krzyż conjecture, posed in 1968 in [13], states that Mn=2/eM_n=2/e and the only extremal function normalized in this way is

fn(z)=ezn−1zn+1=1e+2ezn+⋯(2)f_n(z)=e^{\frac{z^n-1}{z^n+1}}=\frac{1}{e}+\frac{2}{e}z^n+\cdots \tag*{(2)}

This problem provides the main inspiration for our study and is still open. Upper bounds, valid for all n≥1n\geq1 simultaneously and strictly smaller than 11 (with approximate values of 0.999…0.999\ldots), were obtained in [10] and [9]. The conjecture has been proved for n≤5n\leq5 (cf. [13], [11], [5], [24], [23] and [21]), including a recent computer-assisted proof [26] for n=6n=6, but remains open to this date for all larger values of nn. For further information and some closely related questions, we mention [11], [15], [23], [19], and [2], as well as the recent preprints [14] and [17], and [8].

We will use the following known facts about normalized extremal functions, which we collect here for convenience. The three parts of the statement can be found respectively in [11], in both [11] and [19], and in [19], Theorem 1.

Theorem A. Let n≥1n \ge1 and let f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j be a normalized extremal function for (1), that is, a0>0a_0>0 and an=Mna_n=M_n. Then:

(i) f(z)=exp⁡[∑j=1Nrjαjz+1αjz−1]f(z)=\exp\left[\sum_{j=1}^{N}r_j\frac{\alpha_jz+1}{\alpha_jz-1}\right] for some NN with 1≤N≤n1\le N\le n, some rj>0r_j>0, and distinct points αj\alpha_j with ∣αj∣=1|\alpha_j|=1;

(ii) an≥2a0a_n\ge2a_0;

(iii) an=2a0a_n=2a_0 if and only if Mn=2/eM_n=2/e and f=fnf=f_n.

Our purpose is to consider general coefficient extremal problems for non-vanishing bounded analytic functions. A useful survey of extremal problems of this kind is [3]. A comprehensive treatment of various linear and non-linear coefficient extremal problems can be found in the classical reference [12].

Let f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j be analytic in D\mathbb{D}. If ∣f(z)∣≤1|f(z)|\le1 for all z∈Dz\in\mathbb{D}, the Schwarz–Pick lemma yields ∣a1∣≤1−∣a0∣2|a_1|\le1-|a_0|^2. The F. Wiener transform WnfW_nf of ff [19 Section 1.7] is defined by

Wnf(z)=1n∑j=0n−1f(e(2πji)/nz)=∑j=0∞ajnzjn=G(zn),(3)W_nf(z)=\frac{1}{n}\sum_{j=0}^{n-1}f\left(e^{(2\pi ji)/n}z\right)=\sum_{j=0}^{\infty}a_{jn}z^{jn}=G(z^n), \tag*{(3)}

where GG is also analytic in D\mathbb{D} and bounded by 1. Since G(0)=a0G(0)=a_0 and G′(0)=anG'(0)=a_n, the Schwarz–Pick lemma applied to GG yields ∣an∣≤1−∣a0∣2|a_n|\le1-|a_0|^2 for all n≥1n\ge1. Moreover, if f∈B∗f\in\mathcal{B}_*, then clearly ∣an∣<1|a_n|<1 for all n≥1n\ge1 and ∣a0∣=1|a_0|=1 if and only if ff is a unimodular constant. In view of this, if we assume that h:D‾n+1→Ch:\overline{\mathbb{D}}^{n+1}\to\mathbb{C} is holomorphic in Dn+1\mathbb{D}^{n+1}, continuous in D‾n+1\overline{\mathbb{D}}^{n+1}, and non-constant, then the coefficient functional of the form Φ(f)=h(a0,a1,…,an)\Phi(f)=h(a_0,a_1,\ldots,a_n) is well defined for every f∈B∗f\in\mathcal{B}_*. We study the associated extremal problem of finding

Fn=sup⁡{Re⁡Φ(f):f∈B∗}.(4)F_n=\sup\{\operatorname{Re}\Phi(f): f\in\mathcal{B}_*\}. \tag*{(4)}

The Krzyż problem corresponds to the special case h(a0,a1,…,an)=anh(a_0,a_1,\ldots,a_n)=a_n. Our Theorem 4, formulated in § 2, shows that the maximum in (4) is always attained and, moreover, every extremal function is either a unimodular constant or a singular inner function with at most nn atoms; that is, a unimodular multiple of a function like ff in Theorem A (i). This has a flavor of other known results but we have not been able to find a complete proof of the result in this generality in the literature. We also show that if the function hh defining the functional Φ\Phi does not depend on the variable a0a_0, then all extremal functions are non-constant and, thus, have the same structure as the ones for the original Krzyż problem. On the other hand, we give specific non-trivial examples (where Φ\Phi also depends on other variables) showing that an extremal function can actually be constant. It should be mentioned that, for a fixed value of log⁡a0\log a_0, Romanova [20] studied the extremal problem for bounded non-vanishing analytic functions for the linear functional of the form Φ(f)=c1a1+⋯+cnan\Phi(f)=c_1a_1+\cdots+c_na_n and derived a similar structure using an extremal problem for functions with positive real part and Loewner chains.

We also discuss some issues related to normalization, since for certain problems we cannot freely assume that f(0)>0f(0)>0 because of the risk of losing the structure proved in Theorem 4. Examples are given to illustrate some differences in such cases. For a smaller class of functionals Φ\Phi that only depend on the variables a0a_0 and ana_n, in § 2 we show that if a normalized extremal function is unique, it must be a function of znz^n (Theorem 9).

This last situation, together with some known results for the Krzyż conjecture, motivates our § 3, where we study a special family of such functionals Φ\Phi depending only on the coefficients a0a_0 and ana_n of ff. In view of parts (ii) and (iii) of Theorem A, it seems quite appropriate to study convex combinations of the functionals Φ0(f)=an\Phi_0(f)=a_n and Φ1(f)=2a0\Phi_1(f)=2a_0. It also makes sense to restrict our attention to a smaller class containing all candidates for extremal functions for the Krzyż problem. For n≥1n \ge1 we define Bn\mathcal{B}_n as the set of all functions f∈B∗f \in\mathcal{B}_* which are singular inner functions with at most nn atoms and for which an≥2a0>0a_n \ge2a_0 > 0.

Next, let pp be a holomorphic function on D2\mathbb{D}^2 such that for xx, yy real and 0<2x<y<10 < 2x < y < 1, the value p(x,y)p(x,y) is real with 0<p(x,y)≤10 < p(x,y) \le1. By continuity, for 0<x<1/20 < x < 1/2 we have 0≤p(x,2x)≤10 \le p(x,2x) \le1. Constant functions pp are allowed. Under these restrictions on pp, consider the functionals

Φp(f)=2p(a0,an)a0+(1−p(a0,an))an,f∈Bn,(5)\Phi_p(f) = 2p(a_0,a_n)a_0 + (1-p(a_0,a_n))a_n,\qquad f \in\mathcal{B}_n, \tag*{(5)}

where a0a_0 and ana_n are the usual coefficients of the function ff and consider

Gn=Gn(p)=sup⁡{Re⁡Φp(f):f∈Bn},(6)G_n = G_n(p) = \sup\{\operatorname{Re}\Phi_p(f): f \in\mathcal{B}_n\}, \tag*{(6)}

(We write Gn(p)G_n(p) when the dependence on pp needs to be emphasized.) This is of interest since it forces the inequalities 2a0≤Φp(f)≤an2a_0 \le\Phi_p(f) \le a_n for every f∈Bnf \in\mathcal{B}_n; moreover, Φp(f)=an\Phi_p(f)=a_n holds if and only if an=2a0a_n=2a_0, which is precisely the condition known to imply f=fnf=f_n for an extremal function ff of the Krzyż problem. The following result from § 3 connects the problems (1) and (6).

Theorem 1. Let n≥1n \ge1 and let Φp\Phi_p be an admissible functional of the form (5).

(a) If the function fn(z)=ezn−1zn+1f_n(z)=e^{\frac{z^{n}-1}{z^{n+1}}} is extremal for the Krzyż problem (1), that is, if Mn=2/eM_n=2/e, then Gn(p)=2/eG_n(p)=2/e and fnf_n is the only extremal function for (6).

(b) If Gn(p)=MnG_n(p)=M_n, then Mn=2/e=Gn(p)M_n=2/e=G_n(p), the function fnf_n is extremal for (1), and it is the only extremal function for (6). In particular, if

sup⁡{2a0:f∈Bn}=sup⁡{an:f∈Bn},\sup\{2a_0:f \in\mathcal{B}_n\}=\sup\{a_n:f \in\mathcal{B}_n\},

then Mn=2/eM_n=2/e and fnf_n is extremal for the Krzyż problem (1).

(c) If there is only one extremal function for (6), then Gn(p)=2/eG_n(p)=2/e and this function is fnf_n.

Note that the original Krzyż problem can be viewed as the limit case of (6) when pp is constant and tends to zero. Thus, investigating this general problem may indicate a new way of attacking the original conjecture. In particular, (6) is confirmed to be true for n≤5n \le5, as expressed by Corollary 12, and it seems legitimate to conjecture that max⁡{Φp(f):f∈Bn}=2/e\max\{\Phi_p(f):f \in\mathcal{B}_n\}=2/e for all nn, with the same expected extremal function fnf_n as for the normalized Krzyż problem.

Section 4 concerns extremal functions for the special functionals Φp\Phi_p whose initial coefficients are assumed to vanish. There are several statements of this type in the literature. For the Krzyż problem, Brown [4] showed that if aj=0a_j=0 for 1≤j<(n+1)/21 \le j < (n+1)/2 for an extremal function, then the function must be fnf_n. He actually proved a more general version for HpH^p spaces posed in [11]. In the Krzyż setting, a general theorem of this type involving the so-called inductive sets was later given in [19]. Recently, MacDonald [17] improved this result by proving that if an extremal function ff for the Krzyż problem it suffices to assume only a1=⋯=a⌊n/3⌋=0a_1=\cdots=a_{\lfloor n/3\rfloor}=0 to conclude that f=fnf=f_n (see Corollary 23 below). Here ⌊x⌋=max⁡{k∈Z:k≤x}\lfloor x\rfloor=\max\{k \in\mathbb{Z}: k \le x\} denotes the integer part of a real number xx. Note that [17] uses the notation l=⌊(n−2)/3⌋l=\lfloor(n-2)/3\rfloor, which gives the same integer as ⌊n/3⌋\lfloor n/3\rfloor.

Theorem 2. Let n≥6n \ge6, let pp be an admissible weight or p≡0p \equiv0, and let f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j be a fixed extremal function for (6) (for p≡0p \equiv0, this means that ff is a normalized extremal function for the Krzyż problem (1)). Then the following statements are equivalent:

(a) a1=a2=⋯=al=0a_1=a_2=\cdots=a_l=0 for l=⌊n/3⌋l=\lfloor n/3\rfloor.

(b) f(z)=ezn−1zn+1f(z)=e^{\frac{z^n-1}{z^{n+1}}}.

If any of the above conditions is satisfied, then Gn(p)=2eG_n(p)=\frac{2}{e}.

Corollary 23 also covers the case p≡0p \equiv0 of Theorem 2, thus giving a new proof of MacDonald’s theorem. Our proof differs a great deal from MacDonald’s. It is based on the determinant estimates of Efraimidis [7] (generalizing some classical results by Livingston [16]).

Structural results for general non-linear extremal problems

Existence and structure of extremal functions

Recall that, by a theorem of Fatou, every bounded analytic function in D\mathbb{D} has radial limits f~(eiθ)=lim⁡r→1−f(reiθ)\widetilde{f}(e^{i\theta})=\lim_{r\to1-}f(re^{i\theta}) for almost every θ∈[0,2π)\theta\in[0,2\pi). A function bounded by one and such that ∣f~(eiθ)∣=1|\widetilde{f}(e^{i\theta})|=1 almost everywhere is said to be inner. For a positive integer NN, a singular inner function with NN atoms is a function of the form

f(z)=βexp⁡(∑j=1Nrjαjz+1αjz−1),rj>0,∣αj∣=∣β∣=1,1≤j≤N.(7)f(z)=\beta\exp\left(\sum_{j=1}^{N}r_j\frac{\alpha_jz+1}{\alpha_jz-1}\right),\qquad r_j>0,\quad|\alpha_j|=|\beta|=1,\quad1\leq j\leq N. \tag*{(7)}

where the points α1,…,αN\alpha_1,\ldots,\alpha_N are distinct. It is known that every extremal function for the Krzyż problem (1) must have this form with N≤nN\leq n; see Theorem A (i). In spite of this information, proving that the only normalized extremal function is the function fnf_n defined in Section 1 has remained elusive up to the time of writing this paper.

In this section we show that all extremal functions for the general problem (4) must have a structure very similar to (7), with one difference: constant functions of modulus one can also be extremal in some cases. Our proof follows an idea for the original Krzyż problem suggested by Donald E. Marshall [18], which in this case requires a careful further discussion in view of the generality of the problem. To this end, the following auxiliary statement will be very useful.

Lemma 3. Let n≥1n\geq1 and let hh be a non-constant holomorphic function on Dn+1\mathbb{D}^{n+1}. Then there exists F∈B∗F\in\mathcal{B}_* such that

Re⁡Φ(F)>Re⁡h(0,…,0).\operatorname{Re}\Phi(F)>\operatorname{Re}h(0,\ldots,0).

Proof. This statement can be deduced in more than one way. Here is one simple proof.

The restriction of hh defined by g(z)=h(z,0,…,0)g(z)=h(z,0,\ldots,0), z∈Dz\in\mathbb{D}, is clearly analytic in D\mathbb{D}. By the mean value property of the harmonic function Re⁡g\operatorname{Re}g,

Re⁡g(0)=12π∫02πRe⁡g(25eiθ) dθ,\operatorname{Re}g(0)=\frac{1}{2\pi}\int_{0}^{2\pi}\operatorname{Re}g\left(\frac{2}{5}e^{i\theta}\right)\,d\theta,

hence there exists c0c_0 with ∣c0∣=25|c_0|=\frac{2}{5} (in particular, 13<∣c0∣<12\frac{1}{3}<|c_0|<\frac{1}{2}) such that Re⁡g(c0)≥Re⁡g(0)\operatorname{Re}g(c_0)\geq\operatorname{Re}g(0).

Since hh is not constant in Dn+1\mathbb{D}^{n+1}, it cannot be constant on the open set

Ω={(a0,a1,…,an):13<∣a0∣<12, ∣a1∣,…,∣an∣<13n}⊂Dn+1\Omega=\left\{(a_0,a_1,\ldots,a_n):\frac{1}{3}<|a_0|<\frac{1}{2},\ |a_1|,\ldots,|a_n|<\frac{1}{3n}\right\}\subset\mathbb{D}^{n+1}

by the identity principle for holomorphic functions of several complex variables. Now for the value c0c_0 such that 13<∣c0∣<12\frac{1}{3}<|c_0|<\frac{1}{2} and Re⁡g(c0)≥Re⁡g(0)\operatorname{Re}g(c_0)\geq\operatorname{Re}g(0) whose existence we have established, choose a connected open neighborhood U⊂ΩU\subset\Omega of the point (c0,0,…,0)∈Ω(c_0,0,\ldots,0)\in\Omega. Since hh is not constant on UU, the open mapping theorem shows that h(U)h(U) is an open set, hence there exists a point (a0,a1,…,an)∈U(a_0,a_1,\ldots,a_n)\in U such that

Re⁡h(a0,a1,…,an)>Re⁡h(c0,0,…,0),\operatorname{Re}h(a_0,a_1,\ldots,a_n)>\operatorname{Re}h(c_0,0,\ldots,0),

hence

Re⁡h(a0,a1,…,an)>Re⁡h(0,…,0).\operatorname{Re}h(a_0,a_1,\ldots,a_n)>\operatorname{Re}h(0,\ldots,0).

It only remains to check that for any given (a0,a1,…,an)∈Ω(a_0,a_1,\ldots,a_n) \in\Omega there exists a function FF in B∗\mathcal{B}_* such that Φ(F)=h(a0,a1,…,an)\Phi(F)=h(a_0,a_1,\ldots,a_n). This is easily seen by considering the polynomial

F(z)=a0+a1z+…+anznF(z)=a_0+a_1z+\ldots+a_nz^n

By our choice of Ω\Omega, we have F∈B∗F\in\mathcal{B}_* since for all z∈Dz\in\mathbb{D} we have ∣F(z)∣≤∣a0∣+∑j=1n∣aj∣<12+13<1|F(z)|\leq|a_0|+\sum_{j=1}^n|a_j|<\frac{1}{2}+\frac{1}{3}<1 and ∣F(z)∣≥∣a0∣−∑j=1n∣aj∣>13−13=0|F(z)|\geq|a_0|-\sum_{j=1}^n|a_j|>\frac{1}{3}-\frac{1}{3}=0.

We are now ready to prove the existence of extremal functions for (4) and describe their structure. Note that there are some natural additional complications which do not appear in the study of more special problems such as the basic Krzyż problem.

Theorem 4. Let n≥1n\geq1 and let h:D‾n+1→Ch:\overline{\mathbb{D}}^{n+1}\to\mathbb{C} be a non-constant function, holomorphic in Dn+1\mathbb{D}^{n+1} and continuous in D‾n+1\overline{\mathbb{D}}^{n+1}. Then there exists an extremal function for (4). Every such function ff is either a constant function of modulus one or a singular inner function as in (7), where 1≤N≤n1\leq N\leq n. Moreover, if hh does not depend on its first variable a0a_0, then all extremal functions must be of the form (7).

Proof. Existence. Denote by ak(f)a_k(f) the kk-th Taylor coefficient of a function ff analytic in D\mathbb{D}. In order to show that an extremal function for (4) exists, let (gj)j(g_j)_j be a sequence of functions in B∗\mathcal{B}_* such that Fn=lim⁡j→∞Re⁡Φ(gj)F_n=\lim_{j\to\infty}\operatorname{Re}\Phi(g_j) (note that Fn<∞F_n<\infty since hh is bounded on the compact set D‾n+1\overline{\mathbb{D}}^{n+1}). An application of Montel’s theorem yields a subsequence, say (fj)j(f_j)_j, that converges locally uniformly to some function ff. By Weierstrass’ theorem, ff is analytic in D\mathbb{D} and lim⁡j→∞ak(fj)=ak(f)\lim_{j\to\infty}a_k(f_j)=a_k(f) for all k≥0k\geq0. In view of the observations made in § 1, we have ∣ak(fj)∣≤1|a_k(f_j)|\leq1 for all k≥0k\geq0 and all jj, hence ∣ak(f)∣≤1|a_k(f)|\leq1 for all k≥0k\geq0. Moreover, since 0<∣fj∣≤10<|f_j|\leq1 in D\mathbb{D} for all jj, by Hurwitz’s theorem either ff is identically zero or f∈B∗f\in\mathcal{B}_* (in which case it may still be a constant function). The assumption on continuity of hh in D‾n+1\overline{\mathbb{D}}^{n+1} yields

Re⁡h(a0(f),a1(f),…,an(f))=lim⁡j→∞Re⁡h(a0(fj),a1(fj),…,an(fj))=Fn.\operatorname{Re}h(a_0(f),a_1(f),\ldots,a_n(f))=\lim_{j\to\infty}\operatorname{Re}h(a_0(f_j),a_1(f_j),\ldots,a_n(f_j))=F_n.

Thus, ff will be extremal for (4) if it is not identically zero. It is clear from Lemma 3 that at least for the function FF found there Re⁡Φ(F)\operatorname{Re}\Phi(F) yields a strictly greater value than the function identically equal to zero, hence the zero function cannot be extremal, and we are done.

Structure. First, we show that if ff is constant: f≡a0f\equiv a_0, then ∣a0∣=1|a_0|=1. By Lemma 3, for the function FF obtained there we must have

Re⁡h(a0,0,…,0)=Re⁡Φ(f)≥Re⁡Φ(F)>Re⁡h(0,…,0).\operatorname{Re}h(a_0,0,\ldots,0)=\operatorname{Re}\Phi(f)\geq\operatorname{Re}\Phi(F)>\operatorname{Re}h(0,\ldots,0).

In particular, the function g(z)=h(z,0,…,0)g(z)=h(z,0,\ldots,0), holomorphic in D\mathbb{D} and continuous in D‾\overline{\mathbb{D}}, is not constant. Since every constant function ww with 0<∣w∣≤10<|w|\leq1 belongs to B∗\mathcal{B}_*, we have Re⁡g(w)≤Fn=Re⁡g(a0)\operatorname{Re}g(w)\leq F_n=\operatorname{Re}g(a_0) for 0<∣w∣≤10<|w|\leq1, and also for w=0w=0 by continuity. Thus, Re⁡g\operatorname{Re}g attains its maximum over D‾\overline{\mathbb{D}} at the point a0a_0, and the maximum principle yields ∣a0∣=1|a_0|=1, as claimed. We now prove (2) in the case when ff is not constant. It suffices to prove the statement in the special case when f(0)>0f(0)>0. Otherwise we may consider

f0=a0‾∣a0∣f,h0(w0,…,wn)=h(a0∣a0∣w0,…,a0∣a0∣wn).f_0=\frac{\overline{a_0}}{|a_0|}f,\qquad h_0(w_0,\ldots,w_n)=h\left(\frac{a_0}{|a_0|}w_0,\ldots,\frac{a_0}{|a_0|}w_n\right).

Note that multiplication by a0‾∣a0∣\frac{\overline{a_0}}{|a_0|} is a bijective mapping of B∗\mathcal{B}_* onto itself, hence f0∈B∗f_0\in\mathcal{B}_* if and only if f∈B∗f\in\mathcal{B}_*. Also, f0(0)=∣a0∣>0f_0(0)=|a_0|>0 and f0f_0 is extremal for the functional corresponding to the function h0h_0 if and only if ff is extremal for the functional corresponding to the function hh. Thus, it suffices to prove that f0f_0 (as an extremal function for h0h_0) has the structure (7) claimed, and it will immediately follow that ff also has the same structure.

Since ff is not constant, we have 0<f(0)<10<f(0)<1, so we may write f=egf=e^g, where g(z)=∑j=0∞bjzjg(z)=\sum_{j=0}^{\infty}b_jz^j is analytic in D\mathbb{D}, Re⁡g<0\operatorname{Re}g<0 in D\mathbb{D}, and b0=log⁡f(0)b_0=\log f(0) is real and negative. In what follows, we will consider a fixed value b0b_0 and hence also a0=eb0a_0=e^{b_0} will be fixed. As we know from [15] or [19], every coefficient of ff can be written as

ak=a0Pk(b1,…,bk),a_k=a_0P_k(b_1,\ldots,b_k),

where PkP_k is a polynomial (see also Lemma 22 below), hence

Φ(f)=h(a0,a1,…,an)=h(a0,a0P1(b1),…,a0Pn(b1,…,bn))\Phi(f)=h(a_0,a_1,\ldots,a_n)=h(a_0,a_0P_1(b_1),\ldots,a_0P_n(b_1,\ldots,b_n))

is again a holomorphic function of the variables b1,…,bnb_1,\ldots,b_n, which can be written as another holomorphic function

Φ(f)=H(ζ1,…,ζn)\Phi(f)=H(\zeta_1,\ldots,\zeta_n)

of the variables ζ1=−b1,…,ζn=−bn\zeta_1=-b_1,\ldots,\zeta_n=-b_n. Note that the function

−g(z)=−b0−b1z−…−bnzn−…-g(z)=-b_0-b_1z-\ldots-b_nz^n-\ldots

is analytic in D\mathbb{D} and has positive real part there, so we can apply to it the theorems of Carathéodory and Toeplitz which can be found in [25 Chapter IV], among other sources. The set

Rn={(w1,…,wn)∈Cn:∣−2b0w1w‾1−2b0∣>0,…,∣−2b0w1⋯wnw‾1−2b0⋯wn−1⋮⋮⋱⋮w‾nw‾n−1⋯−2b0∣>0}\mathcal{R}_n=\left\{(w_1,\ldots,w_n)\in\mathbb{C}^n: \begin{vmatrix} -2b_0 & w_1\\ \overline{w}_1 & -2b_0 \end{vmatrix}>0,\ldots, \begin{vmatrix} -2b_0 & w_1 & \cdots& w_n\\ \overline{w}_1 & -2b_0 & \cdots& w_{n-1}\\ \vdots& \vdots& \ddots& \vdots\\ \overline{w}_n & \overline{w}_{n-1} & \cdots& -2b_0 \end{vmatrix}>0\right\}

is open, being a finite intersection of open sets. By Sylvester’s criterion, Rn\mathcal{R}_n is the set of vectors for which the corresponding Hermitian Toeplitz matrix is positive definite; hence Rn\mathcal{R}_n is also convex, and it contains the zero vector (w1,…,wn)=(0,…,0)(w_1,\ldots,w_n)=(0,\ldots,0). By the Carathéodory–Toeplitz theorem, the quadratic form associated to the Hermitian Toeplitz matrix corresponding to the coefficient vector (ζ1,…,ζn)(\zeta_1,\ldots,\zeta_n) of −g-g is non-negative. Therefore, for 0<r<10<r<1, the vectors (rζ1,…,rζn)(r\zeta_1,\ldots,r\zeta_n) belong to Rn\mathcal{R}_n, since their Toeplitz matrices are obtained by adding the positive definite matrix −2(1−r)b0I-2(1-r)b_0I to rr times the former one. In particular, (ζ1,…,ζn)(\zeta_1,\ldots,\zeta_n) belongs to the closure of Rn\mathcal{R}_n. If HH were constant in Rn\mathcal{R}_n, continuity would give H(ζ1,…,ζn)=H(0,…,0)H(\zeta_1,\ldots,\zeta_n)=H(0,\ldots,0). Thus, the constant function f≡a0=eb0<1f\equiv a_0=e^{b_0}<1 would also be extremal, which is impossible by the first part of the proof. Hence, HH is not constant in Rn\mathcal{R}_n.

Suppose now, to reach a contradiction, that (ζ1,…,ζn)∈Rn(\zeta_1,\ldots,\zeta_n)\in\mathcal{R}_n. Since Rn\mathcal{R}_n is open and connected and HH is not constant there, the open mapping theorem for holomorphic functions of several complex variables implies the existence of (w1,…,wn)∈Rn(w_1,\ldots,w_n)\in\mathcal{R}_n such that

Re⁡H(ζ1,…,ζn)<Re⁡H(w1,…,wn).\operatorname{Re}H(\zeta_1,\ldots,\zeta_n)<\operatorname{Re}H(w_1,\ldots,w_n).

On the other hand, by a theorem of Carathéodory [25 Theorem IV.24], there exists a function −g∗-g^*, analytic in D\mathbb{D} and with positive real part, such that

−g∗(z)=−b0+w1z+…+wnzn+…-g^*(z)=-b_0+w_1z+\ldots+w_nz^n+\ldots

The function f∗=eg∗f^*=e^{g^*} belongs to B∗\mathcal{B}_* and satisfies f∗(0)=a0f^*(0)=a_0 and Φ(f∗)=H(w1,…,wn)\Phi(f^*)=H(w_1,\ldots,w_n), hence Re⁡Φ(f)<Re⁡Φ(f∗)\operatorname{Re}\Phi(f)<\operatorname{Re}\Phi(f^*), which contradicts the assumption that ff is extremal.

Therefore, (ζ1,…,ζn)∉Rn(\zeta_1,\ldots,\zeta_n) \notin\mathbb{R}_n. Since all the Toeplitz determinants associated with −g-g are non-negative and the first one equals −2b0>0-2b_0>0, there is a smallest integer NN with 1≤N≤n1\leq N\leq n such that

∣−2b0−b1⋯−bN−b1−2b0⋯−bN−1⋮⋮⋱⋮−bN−bN−1⋯−2b0∣=0.\begin{vmatrix} -2b_0 & -b_1 & \cdots& -b_N \\ -b_1 & -2b_0 & \cdots& -b_{N-1} \\ \vdots& \vdots& \ddots& \vdots\\ -b_N & -b_{N-1} & \cdots& -2b_0 \end{vmatrix}=0.

The Carathéodory–Toeplitz theorem [25 Theorem IV.22] implies that, for some rj>0r_j>0 and distinct points αj\alpha_j with ∣αj∣=1|\alpha_j|=1,

−g(z)=∑j=1Nrj1+αjz1−αjz.-g(z)=\sum_{j=1}^{N}r_j\frac{1+\alpha_jz}{1-\alpha_jz}.

which proves the structural formula (7) for ff.

Finally, suppose that hh does not depend on its first variable. If f≡a0f\equiv a_0 is a constant function, then a1=⋯=an=0a_1=\cdots=a_n=0, hence Φ(f)=h(a0,0,…,0)=h(0,0,…,0)\Phi(f)=h(a_0,0,\ldots,0)=h(0,0,\ldots,0). Thus, for the function FF given by Lemma 3 we have

Fn≥Re⁡Φ(F)>Re⁡h(0,…,0)=Re⁡Φ(f),F_n\geq\operatorname{Re}\Phi(F)>\operatorname{Re}h(0,\ldots,0)=\operatorname{Re}\Phi(f),

hence no constant function can be extremal. □

Uniqueness and normalizations

It is important to notice that the structure of extremal functions is not quite the same in the most general case as in the Krzyż problem. Also, for the sake of trying to obtain uniqueness, in some cases it may be convenient to consider our extremal problems over a smaller class instead of over the entire class B∗\mathcal{B}_*. For example, for the original Krzyż problem we can use rotations as above to see that the maximum MnM_n defined in (1) is actually equal to

Mn=sup⁡{Re⁡an:f∈B∗, f(0)=a0>0}.M_n=\sup\{\operatorname{Re}a_n:f\in\mathcal{B}_*,\ f(0)=a_0>0\}.

We will refer to this formulation as the normalized Krzyż problem. Three basic questions arise here:

  1. Can a constant function of modulus one actually be extremal for some choice of Φ\Phi in the general problem (4)? A trivial example would be Φ(f)=a0\Phi(f)=a_0, but it is desirable to have an example where Φ\Phi does not depend only on a0a_0.

  2. How many extremal functions can we have for a given problem? Is it possible to have only one such function, or maybe some other small finite number?

  3. Can we impose further normalizations as in the Krzyż problem, such as f(0)>0f(0)>0? How does this affect the extremal functions, and do we get uniqueness?

In what follows, we answer all these questions by providing appropriate examples.

Our first example answers our question (1) raised above and, regarding question (2), shows that an extremal problem can have a unique solution.

Example 5. The extremal problem associated to the linear functional Φ(f)=2a0+an\Phi(f)=2a_0+a_n is maximized only by the constant function 11. Recalling that for f∈B∗f\in\mathcal{B}_* we have ∣an∣≤1−∣a0∣2|a_n|\leq1-|a_0|^2 (see § 1), this can be seen as follows:

Re⁡(2a0+an)≤2Re⁡a0+(1−∣a0∣2)≤1+2Re⁡a0−(Re⁡a0)2≤2,\operatorname{Re}(2a_0+a_n)\leq2\operatorname{Re}a_0+(1-|a_0|^2)\leq1+2\operatorname{Re}a_0-(\operatorname{Re}a_0)^2\leq2,

and the maximum value 22 is only attained when Re⁡a0=1\operatorname{Re}a_0=1, which immediately forces ff to be constant: f≡1f\equiv1. Thus, in this case there is only one extremal function and it is constant.

Note that the functionals Φ(f)\Phi(f) which are only functions of a0a_0, that is,

h(a0,a1,…,an)=h(a0,0,…,0),∀(a0,a1,…,an)∈Dn+1h(a_0,a_1,\ldots,a_n)=h(a_0,0,\ldots,0), \qquad\forall(a_0,a_1,\ldots,a_n)\in\mathbb{D}^{n+1}

give rise to some unusual situations. Here is an example.

Example 6. Let Φ(f)=a0−a02\Phi(f)=a_0-a_0^2.

(a) Let us first maximize Re⁡Φ(f)\operatorname{Re}\Phi(f) over B∗\mathcal{B}_*. The value a0=f(0)a_0=f(0) must satisfy 0<∣a0∣≤10<|a_0|\le1. Writing a0=x+yia_0=x+yi with x,y∈Rx,y\in\mathbb{R} and 0<x2+y2≤10<x^2+y^2\le1, we see that

Re⁡Φ(f)=x−x2+y2≤1+x−2x2≤98.\operatorname{Re}\Phi(f)=x-x^2+y^2\le1+x-2x^2\le\frac{9}{8}.

The maximum is attained only when equality holds throughout; that is, only for

x=14,y=±154,x=\frac{1}{4},\qquad y=\pm\frac{\sqrt{15}}{4},

and in this case a0a_0 clearly has modulus 11, forcing f≡a0f\equiv a_0. Thus, there are exactly two extremal functions and they are both constant.

(b) If we want to maximize Re⁡Φ(f)\operatorname{Re}\Phi(f) over the reduced class of all f∈B∗f\in\mathcal{B}_* with a0>0a_0>0, we obtain

Re⁡Φ(f)=a0−a02=a0(1−a0)≤14.\operatorname{Re}\Phi(f)=a_0-a_0^2=a_0(1-a_0)\le\frac{1}{4}.

with equality only for a0=12a_0=\frac{1}{2}. However, in this case it is immediate that there are many extremal functions different from those described by Theorem 4, e.g., every function of the form

fλ(z)=12(1+λz),∣λ∣≤1.f_\lambda(z)=\frac{1}{2}(1+\lambda z),\qquad|\lambda|\le1.

All these functions are obviously bounded by one, non-vanishing in D\mathbb{D}, and have the property f(0)=12f(0)=\frac{1}{2}. Even the constant function among them (λ=0\lambda=0) is clearly different from the ones from Theorem 4 and the ones found in the previous analysis.

Both Example 5 and Example 6 (a) are in sharp contrast with the Krzyż problem, for which there are infinitely many extremal functions if we do not assume any normalization. Indeed, at least one extremal function ff for the Krzyż problem exists, and for each λ\lambda with ∣λ∣=1|\lambda|=1 the function λ‾ nf(λz)\overline{\lambda}^{\,n}f(\lambda z) has the same nn-th coefficient as ff; since its value at the origin is λ‾ na0≠0\overline{\lambda}^{\,n}a_0\ne0, there are uncountably many distinct extremal functions.

As our Example 6 (b) shows, the structure obtained in Theorem 4 can be completely ruined under the additional assumption f(0)>0f(0)>0. One way of getting around this obstacle in order to be able to assume that there is an extremal function with a positive value at the origin is as follows. Define two functions hh and h~\widetilde{h} that satisfy the conditions of Theorem 4 to be associated if there exists λ\lambda with ∣λ∣=1|\lambda|=1 such that h~(z0,z1,…,zn)=h(λz0,λz1,…,λzn)\widetilde{h}(z_0,z_1,\ldots,z_n)=h(\lambda z_0,\lambda z_1,\ldots,\lambda z_n). This clearly defines an equivalence relation, so instead of individual functionals we can consider equivalence classes of associated functionals. In each class we can choose one representative, hh, such that there exists an extremal function ff for the functional Φ(f)=h(a0,a1,…,an)\Phi(f)=h(a_0,a_1,\ldots,a_n) with a0>0a_0>0, because we can find λ\lambda with ∣λ∣=1|\lambda|=1 such that λa0>0\lambda a_0>0. This may provide a satisfactory solution (albeit not an ideal one).

The next example shows that the problem encountered in our Example 6 does not come from the fact that the functional depended only on a0a_0. There are other functionals that give rise to the same situation, even though the extremal function fits within the description given in Theorem 4.

Example 7. Let n=1n=1 and Φ(f)=a1−a02\Phi(f)=a_1-a_0^2. We already know that there is an extremal function for this functional but there are none with the property f(0)>0f(0)>0. For a general f∈B∗f\in\mathcal{B}_* we have

Re⁡Φ(f)=Re⁡a1−Re⁡a02≤1−∣a0∣2−Re⁡a02≤1.\operatorname{Re}\Phi(f)=\operatorname{Re}a_1-\operatorname{Re}a_0^2\leq1-\lvert a_0\rvert^2-\operatorname{Re}a_0^2\leq1.

Equality in the Schwarz–Pick lemma for a non-constant function is only possible for a disc automorphism, which has a zero in D\mathbb{D}, so this case must be ruled out. Thus, equality in the above chain can only hold for a constant function f=a0f=a_0 with ∣a0∣=1\lvert a_0\rvert=1 and Re⁡a0=0\operatorname{Re}a_0=0, that is, f≡±if\equiv\pm i. Hence, there are only two extremal functions, both constant functions of modulus one. However, under the assumption a0>0a_0>0 we get

Re⁡Φ(f)=Re⁡a1−a02≤Re⁡a1≤2e,\operatorname{Re}\Phi(f)=\operatorname{Re}a_1-a_0^2\leq\operatorname{Re}a_1\leq\frac{2}{e},

as is well known. Thus, no extremal function has the property f(0)>0f(0)>0.

In order to obtain an associated functional with an extremal function that has a positive value at the origin, instead of Φ\Phi we would have to consider the functional Φi(f)=ia1−(ia0)2=a02+ia1\Phi_i(f)=ia_1-(ia_0)^2=a_0^2+ia_1, whose extremal functions are f≡±1f\equiv\pm1.

Further properties of extremal functions

In what follows, we will refer to any function of the form (7) as a candidate function. We will use the notation CN\mathcal{C}_N for the set of all candidate functions with exactly NN atoms, 1≤N≤n1\leq N\leq n, while C0\mathcal{C}_0 will denote the class of constant functions of modulus one (no atoms in the representation in Theorem 4). Thus, every extremal function for (4) belongs to the set ⋃N=0nCN\bigcup_{N=0}^{n}\mathcal{C}_N. The following fact was observed in [2 Remark 3.2] for the Krzyż problem, and this easily generalizes to our situation. It shows that finding the maximum of Re⁡Φ(f)\operatorname{Re}\Phi(f) over all admissible functions reduces to finding the supremum of the same quantity only over the functions with the full number of atoms.

Remark 8. Let n≥1n\geq1, 0≤N<n0\leq N<n, and let h:D‾n+1→Ch:\overline{\mathbb{D}}^{n+1}\to\mathbb{C} be a non-constant function, holomorphic in Dn+1\mathbb{D}^{n+1} and continuous in D‾n+1\overline{\mathbb{D}}^{n+1}. Then

sup⁡{Re⁡Φ(f):f∈CN}≤sup⁡{Re⁡Φ(f):f∈CN+1}.\sup\{\operatorname{Re}\Phi(f):f\in\mathcal{C}_N\}\leq\sup\{\operatorname{Re}\Phi(f):f\in\mathcal{C}_{N+1}\}.

In particular,

sup⁡{Re⁡Φ(f):f∈⋃N=0n−1CN}≤sup⁡{Re⁡Φ(f):f∈Cn}.\sup\left\{\operatorname{Re}\Phi(f):f\in\bigcup_{N=0}^{n-1}\mathcal{C}_N\right\}\leq\sup\{\operatorname{Re}\Phi(f):f\in\mathcal{C}_n\}.

Hence,

Fn=max⁡{Re⁡Φ(f):f∈B∗}=sup⁡{Re⁡Φ(f):f∈Cn}.F_n=\max\{\operatorname{Re}\Phi(f):f\in\mathcal{B}_*\}=\sup\{\operatorname{Re}\Phi(f):f\in\mathcal{C}_n\}.

To see this, it suffices to show that if Re⁡Φ(f)≤M\operatorname{Re}\Phi(f)\leq M for every candidate function with N+1N+1 atoms, 0≤N<n0\leq N<n, then the same is true for every candidate function with NN atoms. Given f∈CNf\in\mathcal{C}_N, if we choose α\alpha with ∣α∣=1\lvert\alpha\rvert=1 different from the atoms of ff (if N=0N=0, any α\alpha will do). Then the function

gr(z)=f(z)erαz+1αz−1,r>0,g_r(z)=f(z)e^{r\frac{\alpha z+1}{\alpha z-1}},\qquad r>0,

belongs to CN+1\mathcal{C}_{N+1}. It is elementary to see that rαz+1αz−1→0r\frac{\alpha z+1}{\alpha z-1}\to0 uniformly on compact sets, hence erαz+1αz−1→1e^{r\frac{\alpha z+1}{\alpha z-1}}\to1, also uniformly on compact sets. By Weierstrass’ theorem, it is immediate that the Taylor coefficients of grg_r converge to the Taylor coefficients of ff and therefore also Re⁡Φ(gr)→Re⁡Φ(f)\operatorname{Re}\Phi(g_r)\to\operatorname{Re}\Phi(f) as r→0+r\to0^+. Thus, any uniform bound for Re⁡Φ(f)\operatorname{Re}\Phi(f) over the class CN+1\mathcal{C}_{N+1} is also a uniform bound for the functions in CN\mathcal{C}_N.

It has recently been proved by Lei and Zhang [14] (indirectly, by proving the Agler–McCarthy conjecture [2]) that if every extremal function for the original Krzyż problem has exactly nn atoms, then the Krzyż conjecture is true. Note, however, that even for the Krzyż problem it has not yet been established that every extremal function must have the full number of atoms: N=nN=n, nor that there exists an extremal function with this number of atoms.

It is well-known to the experts that if there is a unique extremal functions for the Krzyż conjecture, then the conjecture holds. This fact admits a partial extension to our generalized problem by following the same argument.

Theorem 9. Let h:D‾2→Ch:\overline{\mathbb{D}}^{2}\to\mathbb{C} be a non-constant function holomorphic in D2\mathbb{D}^{2} and continuous in D‾2\overline{\mathbb{D}}^{2}, and consider the functional Φ(f)=h(a0,an)\Phi(f)=h(a_{0},a_{n}). If the associated extremal problem (4) has exactly one extremal function ff with a0>0a_{0}>0, then f(z)=g(zn)f(z)=g(z^{n}) for some g∈B∗g\in\mathcal{B}_{*}.

Proof. If ω\omega is a primitive nn-th root of unity, then fj(z)=f(ωjz)f_{j}(z)=f(\omega^{j}z) has the same coefficients a0a_{0} and ana_{n} as ff, hence it is also extremal for all 0≤j≤n−10\leq j\leq n-1. By uniqueness, f=fjf=f_{j} for all jj and, in particular, f=Wnff=W_{n}f where WnfW_{n}f is the F. Wiener transform of ff defined by (3) in § 1. Thus, f(z)=Wnf(z)=g(zn)f(z)=W_{n}f(z)=g(z^{n}), where gg is analytic in D\mathbb{D}; since f∈B∗f\in\mathcal{B}_{*}, also g∈B∗g\in\mathcal{B}_{*}.

Note that, in spite of the restrictive condition on the coefficients, the above result covers the original Krzyż problem. Moreover, the same argument applies verbatim to extremal problems posed over any subclass of B∗\mathcal{B}_{*} that is invariant under the rotations f↦f(ωz)f\mapsto f(\omega z), such as the class Bn\mathcal{B}_{n} considered in the next section.

A special class of problems

Although our Theorem 4 is valid for quite general functionals, we will now focus on a very specific subclass of functionals which will be useful for our purpose and to which Theorem 9 also applies. As a motivation, we first notice the following.

By Theorem A (i) and (ii), every normalized extremal function ff for the Krzyż problem (1) satisfies an≥2a0>0a_{n}\geq2a_{0}>0 and belongs to one of the classes CN\mathcal{C}_{N} of candidate functions, with 1≤N≤n1\leq N\leq n. Thus, we actually have

Mn=max⁡{Re⁡an:f∈B∗}=max⁡{an:f∈Bn},M_{n}=\max\{\operatorname{Re}a_{n}:f\in\mathcal{B}_{*}\}=\max\{a_{n}:f\in\mathcal{B}_{n}\},

where

Bn={f∈B∗:f(z)=∑j=0∞ajzj in D, 0<2a0≤an, f∈⋃N=1nCN}.(8)\mathcal{B}_{n}=\left\{f\in\mathcal{B}_{*}:f(z)=\sum_{j=0}^{\infty}a_{j}z^{j}\text{ in }\mathbb{D},\ 0<2a_{0}\leq a_{n},\ f\in\bigcup_{N=1}^{n}\mathcal{C}_{N}\right\}. \tag*{(8)}

Clearly, this class encodes all the information needed about possible extremal functions for the Krzyż problem. We note that Bn\mathcal{B}_{n} contains an uncountable collection of functions with exactly nn atoms. Namely, let r>0r>0 and define

gr(z)=−r1−zn1+zn=−r(1−2zn+2z2n−2z3n+…),fr=egr.g_{r}(z)=-r\frac{1-z^{n}}{1+z^{n}}=-r(1-2z^{n}+2z^{2n}-2z^{3n}+\ldots),\qquad f_{r}=e^{g_{r}}.

Then frf_{r} is a singular inner function with exactly nn atoms and

a0=e−r>0,an=2re−r>0.a_{0}=e^{-r}>0,\qquad a_{n}=2re^{-r}>0.

Thus, we have an≥2a0a_{n}\geq2a_{0} whenever r≥1r\geq1. Note that for r=1r=1 we obtain the conjectured extremal function fnf_{n}.

One can also find examples of functions in Bn\mathcal{B}_{n} with fewer than nn atoms when n≥2n\geq2.

Example 10. Let hr(z)=exp⁡(rz−1z+1)h_r(z)=\exp\left(r\frac{z-1}{z+1}\right), r>0r>0, which is a singular inner function with one atom. We can express this function as

hr(z)=e−r∑j=0∞(−1)jLj(−1)(2r)zj,h_r(z)=e^{-r}\sum_{j=0}^{\infty}(-1)^jL_j^{(-1)}(2r)z^j,

where Lj(−1)L_j^{(-1)} are the Laguerre polynomials of order −1-1; cf. [15 p. 258]. Thus, hr∈Bnh_r\in\mathcal{B}_n if and only if (−1)nLn(−1)(2r)≥2(-1)^nL_n^{(-1)}(2r)\ge2. This polynomial can be explicitly computed [1 p. 775] as

Ln(−1)(2r)=∑j=0n(−1)j(n−1n−j)(2r)jj!.L_n^{(-1)}(2r)=\sum_{j=0}^{n}(-1)^j\binom{n-1}{n-j}\frac{(2r)^j}{j!}.

Since the leading coefficient of the polynomial (−1)nLn(−1)(2r)(-1)^nL_n^{(-1)}(2r) in rr is positive, this holds for all sufficiently large rr.

In the rest of this section we consider special functionals of the form (5), which are convex combinations of 2a02a_0 and ana_n, as defined in Section 1. We refer to such functionals as admissible functionals and to the corresponding functions pp as admissible weights. For n≥1n\ge1 and for each admissible weight pp, we consider the extremal problem of finding the supremum defined in (6).

Here are a few examples of such functionals:

Φ(f)=2a0,Φ(f)=2a02+(1−a0)an,Φ(f)=2a0an+(1−an)an.\Phi(f)=2a_0,\qquad\Phi(f)=2a_0^2+(1-a_0)a_n,\qquad\Phi(f)=2a_0a_n+(1-a_n)a_n.

which correspond, respectively, to the weights

p(x,y)=1,p(x,y)=x,p(x,y)=y.p(x,y)=1,\qquad p(x,y)=x,\qquad p(x,y)=y.

Since Φp(f)=an−p(a0,an)(an−2a0)\Phi_p(f)=a_n-p(a_0,a_n)(a_n-2a_0), for all f∈Bnf\in\mathcal{B}_n we have

2a0≤Φp(f)≤an,Φp(f)=an⟺an=2a0.(9)2a_0\le\Phi_p(f)\le a_n,\qquad\Phi_p(f)=a_n\Longleftrightarrow a_n=2a_0. \tag*{(9)}

By Theorem A (iii), if a fixed normalized extremal function ff for the Krzyż problem (1) satisfies a˙n=2a0\dot{a}_n=2a_0, then Mn=2/eM_n=2/e and f=fnf=f_n. This is precisely what makes the functionals Φ˙p\dot{\Phi}_p interesting for our study, in an attempt to “tighten” as much as possible the inequality 2a0≤an2a_0\le a_n and strive for equality.

The classical Krzyż problem corresponds to the weight p˙≡0\dot{p}\equiv0, which is not admissible, since the equivalence in (9) fails for it. Nevertheless, it will be convenient to allow this limit case in some statements, and we will always say so explicitly. Note that for constant weights p≡cp\equiv c with c∈(0,1]c\in(0,1] we easily obtain

(1−c)Mn≤Gn(c)≤Mn,(1-c)M_n\le G_n(c)\le M_n,

so the classical extremal value MnM_n remains the limit of these problems as c→0+c\to0^+. Note, however, that this convergence does not itself establish equality for any c∈(0,1]c\in(0,1].

Recall that Example 5, which concerns the functional 2a0+an=2Φ1/2(f)2a_0+a_n=2\Phi_{1/2}(f) corresponding to the weight p≡1/2p\equiv1/2, shows that constant extremal functions are possible when the maximum is taken over the whole class B∗\mathcal{B}_*. Here we avoid this problem by imposing the structure from Theorem 4 in view of our definition (8) of the restricted class Bn\mathcal{B}_n.

Proposition 11. Let Φp\Phi_p be as in (5). Then Gn≥2/eG_n\ge2/e and the supremum in (6) is attained, that is, there exists an extremal function f∈Bnf\in\mathcal{B}_n (by definition, ff is a singular inner function with at most nn atoms).

Proof. The function fn(z)=ezn−1zn+1=1e+2ezn+⋯f_n(z)=e^{\frac{z^{n}-1}{z^{n}+1}}=\frac{1}{e}+\frac{2}{e}z^n+\cdots belongs to Bn\mathcal{B}_n and Φp(fn)=2/e\Phi_p(f_n)=2/e, hence Gn≥2/eG_n\ge2/e. Let (gj)j(g_j)_j be a sequence in Bn\mathcal{B}_n such that Φp(gj)→Gn\Phi_p(g_j)\to G_n. Since gj(0)>0g_j(0)>0, each gjg_j can be written as

gj(z)=exp⁡(−∑i=1Njrij1+αijz1−αijz),Nj≤n,rij>0,∣αij∣=1.g_j(z)=\exp\left(-\sum_{i=1}^{N_j}r_{ij}\frac{1+\alpha_{ij}z}{1-\alpha_{ij}z}\right),\qquad N_j\le n,\quad r_{ij}>0,\quad|\alpha_{ij}|=1.

Let sj=∑i=1Njrij=−log⁡gj(0)s_j=\sum_{i=1}^{N_j}r_{ij}=-\log g_j(0). If sj→∞s_j\to\infty along a subsequence, then the estimate ∣gj(z)∣≤exp⁡(−sj1−∣z∣1+∣z∣)|g_j(z)|\le\exp\left(-s_j\frac{1-|z|}{1+|z|}\right) shows that gj→0g_j\to0 locally uniformly along that subsequence, so that a0(gj)→0a_0(g_j)\to0, an(gj)→0a_n(g_j)\to0 and, by the continuity of pp at (0,0)(0,0), Φp(gj)→0\Phi_p(g_j)\to0, which contradicts Gn≥2/eG_n\ge2/e. Hence, (sj)j(s_j)_j is bounded and, passing to a subsequence, we may assume that Nj=NN_j=N for all jj, rij→ri≥0r_{ij}\to r_i\ge0 and αij→αi\alpha_{ij}\to\alpha_i for 1≤i≤N1\le i\le N. Then gj→fg_j\to f locally uniformly in D\mathbb{D}, where

f(z)=exp⁡(−∑i=1Nri1+αiz1−αiz)f(z)=\exp\left(-\sum_{i=1}^{N}r_i\frac{1+\alpha_i z}{1-\alpha_i z}\right)

is either a singular inner function with at most nn atoms (some atoms may merge or disappear in the limit) or the constant function 1. Since the Taylor coefficients converge, we have 0<2a0(f)≤an(f)0<2a_0(f)\le a_n(f); in particular, ff is not constant. Thus, f∈Bnf\in\mathcal{B}_n and, by the continuity of pp in D2\mathbb{D}^2,

Φp(f)=lim⁡j→∞Φp(gj)=Gn.\Phi_p(f)=\lim_{j\to\infty}\Phi_p(g_j)=G_n.

By Theorem A (iii), proving the equality an=2a0a_n=2a_0 for a fixed normalized extremal function would prove the statement of the Krzyż conjecture for that specific function. It is also known that uniqueness of the normalized extremal function is equivalent to the full statement of the conjecture. Our Theorem 1, formulated in the Introduction, collects the analogous relations between the problems (1) and (6). Its proof is simple, but it shows that the two problems are closely related and gives us the flexibility to change from one functional to another when convenient. In particular, part (b) shows that the apparently weaker assumption sup⁡{2a0:f∈Bn}=sup⁡{an:f∈Bn}\sup\{2a_0:f\in\mathcal{B}_n\}=\sup\{a_n:f\in\mathcal{B}_n\} already implies that Mn=2/eM_n=2/e. Thus, we believe the result may prove useful for the further study of the Krzyż problem.

Proof of Theorem 1. (a) Let f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j be an extremal function for (6); it exists by Proposition 11. Since fn∈Bnf_n\in\mathcal{B}_n and Φp(fn)=2/e\Phi_p(f_n)=2/e, the inequalities (9) and the inclusion Bn⊂B∗\mathcal{B}_n\subset\mathcal{B}_* yield

2e≤Gn(p)=Φp(f)≤an≤Mn=2e.\frac{2}{e}\le G_n(p)=\Phi_p(f)\le a_n\le M_n=\frac{2}{e}.

Hence, Gn(p)=2/eG_n(p)=2/e and Φp(f)=an=Mn\Phi_p(f)=a_n=M_n. Thus, ff is also a normalized extremal function for the Krzyż problem (1) and, by (9), an=2a0a_n=2a_0. By Theorem A (iii), f=fnf=f_n.

(b) By Proposition 11, there is an extremal function ff for (6). For any such function f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j, the inequalities (9) and the inclusion Bn⊂B∗\mathcal{B}_n\subset\mathcal{B}_* yield

Mn=Gn(p)=Φp(f)≤an≤Mn.M_n=G_n(p)=\Phi_p(f)\le a_n\le M_n.

Hence, Φp(f)=an=Mn\Phi_p(f)=a_n=M_n, so ff is also a normalized extremal function for the Krzyż problem (1) and, by (9), an=2a0a_n=2a_0. By Theorem A (iii), Mn=2/eM_n=2/e and f=fnf=f_n. Since this holds for every extremal function for (6), fnf_n is the only one. The particular case follows by taking p≡1p\equiv1, since Gn(1)=sup⁡{2a0:f∈Bn}G_n(1)=\sup\{2a_0:f\in\mathcal{B}_n\} and Mn=sup⁡{an:f∈Bn}M_n=\sup\{a_n:f\in\mathcal{B}_n\}.

(c) Let ω=e(2πi)/n\omega=e^{(2\pi i)/n}. The class Bn\mathcal{B}_n is invariant under the rotations f↦f(ωjz)f\mapsto f(\omega^jz), which do not change the coefficients a0a_0 and ana_n. Hence, proceeding as in the proof of Theorem 9, we see that the unique extremal function satisfies f(z)=g(zn)f(z)=g(z^n) for some g∈B∗g\in\mathcal{B}_* with g(0)=a0>0g(0)=a_0>0 and g′(0)=ang'(0)=a_n. By the classical case n=1n=1 of the Krzyż conjecture, we have an=g′(0)≤2/ea_n=g'(0)\le2/e, with equality if and only if g(z)=ez−1z+1g(z)=e^{\frac{z-1}{z+1}}. Arguing like in (a), we see that Gn(p)=2/eG_n(p)=2/e, g(z)=ez−1z+1g(z)=e^{\frac{z-1}{z+1}}, and f(z)=g(zn)=ezn−1zn+1f(z)=g(z^n)=e^{\frac{z^n-1}{z^n+1}}.

Since the Krzyž conjecture has been proved for n≤5n\leq5 and a computer-assisted proof for n=6n=6 is currently under scrutiny (see § 1), Theorem 1 (a) yields the following.

Corollary 12. Let 1≤n≤51\leq n\leq5 and let Φp\Phi_p be an admissible functional. Then Gn(p)=2/eG_n(p)=2/e and fnf_n given by (2) is the only extremal function for (6).

We hope that Theorem 1 may provide some new directions in the research on the Krzyž problem. The intermediate values of pp can be seen as a “bridge” between the original Krzyž conjecture and the problem of maximizing the constant term within the class Bn\mathcal{B}_n and may also deserve a closer look.

Remark 13. Comparing the extremal functions for the general problem and the original Krzyž problem allows us to deduce some conclusions about the constant term of an extremal function. For instance, if Gn(p)=2/eG_n(p)=2/e for some admissible weight pp, then every normalized extremal function g(z)=∑j=0∞bjzjg(z)=\sum_{j=0}^{\infty}b_jz^j for the Krzyž problem (1) satisfies b0≤1/eb_0\leq1/e. Indeed, g∈Bng\in\mathcal{B}_n by Theorem A, and (9) yields 2b0≤Φp(g)≤Gn(p)=2/e2b_0\leq\Phi_p(g)\leq G_n(p)=2/e.

Remark 14. The case p≡1p\equiv1 admits a natural reformulation in terms of the singular measure associated with ff. If f∈Bnf\in\mathcal{B}_n is written as

f(z)=exp⁡(−∑j=1Nrj1+αjz1−αjz)f(z)=\exp\left(-\sum_{j=1}^{N}r_j\frac{1+\alpha_jz}{1-\alpha_jz}\right)

then a0=e−sa_0=e^{-s}, where s=∑j=1Nrjs=\sum_{j=1}^{N}r_j is the total mass of this measure. Hence, Gn(1)=2e−s∗G_n(1)=2e^{-s_*}, where s∗s_* is the minimum of the total mass over all f∈Bnf\in\mathcal{B}_n; it is attained by Proposition 11. Since fnf_n has total mass 1, the conjecture formulated in Subsection 4.2 for p≡1p\equiv1 states that every singular inner function ff with at most nn atoms, f(0)>0f(0)>0 and an≥2a0a_n\geq2a_0 has total mass at least 1, with equality only for f=fnf=f_n. Moreover, by Theorem 1 (b), the equality 2e−s∗=Mn2e^{-s_*}=M_n would imply that Mn=2/eM_n=2/e.

In the same spirit as above, we can also prove the following result. Observe that the bound in (c) is only of interest when p>1−2/ep>1-2/e, since otherwise the maximum on the right-hand side is zero.

Proposition 15. Let Φp\Phi_p be an admissible functional and let f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j be an arbitrary but fixed extremal function for (6). Let g(z)=∑j=0∞bjzjg(z)=\sum_{j=0}^{\infty}b_jz^j be an arbitrary but fixed normalized extremal function for the Krzyž problem (1). Then the following statements hold.

(a) If p(a0,an)≥p(b0,bn)p(a_0,a_n)\geq p(b_0,b_n) then b0≤a0b_0\leq a_0.

(b) If pp is a positive constant, then b0≤a0b_0\leq a_0.

(c) If pp is a constant with 0<p<10<p<1, then

a0≥max⁡{p−p2−(1−p)(2e−1+p)1−p,0},a_0\geq\max\left\{\frac{p-\sqrt{p^2-(1-p)\left(\frac{2}{e}-1+p\right)}}{1-p},0\right\},

and if p≡1p\equiv1, then a0≥1/ea_0\geq1/e.

Proof. For part (a), we see that

p(a0,an)(2a0−an)+an≥p(b0,bn)(2b0−bn)+bn≥p(a0,an)(2b0−bn)+bn.p(a_0,a_n)(2a_0-a_n)+a_n \ge p(b_0,b_n)(2b_0-b_n)+b_n \ge p(a_0,a_n)(2b_0-b_n)+b_n.

Consequently, it follows that

2p(a0,an)(a0−b0)≥(1−p(a0,an))(bn−an)≥0,2p(a_0,a_n)(a_0-b_0) \ge(1-p(a_0,a_n))(b_n-a_n) \ge0,

which proves that a0≥b0a_0 \ge b_0 if p(a0,an)>0p(a_0,a_n)>0. If p(a0,an)=0p(a_0,a_n)=0, then p(b0,bn)=0p(b_0,b_n)=0 by hypothesis, so the first chain of inequalities gives an≥bna_n \ge b_n; moreover, an=2a0a_n=2a_0, since p(x,y)>0p(x,y)>0 whenever 0<2x<y<10<2x<y<1. Hence, 2b0≤bn≤an=2a02b_0 \le b_n \le a_n=2a_0.

Part (b) follows automatically from part (a) since if pp is a constant, then p(a0,an)=p(b0,bn)p(a_0,a_n)=p(b_0,b_n).

Finally, for part (c), assume that pp is a constant. Since fn∈Bnf_n \in\mathcal{B}_n, we have

2a0p+(1−p)an=Gn≥Φp(fn)=2/e.2a_0p+(1-p)a_n=G_n \ge\Phi_p(f_n)=2/e.

If p=1p=1, this is a0≥1/ea_0 \ge1/e. If 0<p<10<p<1, we use the estimate an≤1−a02a_n \le1-a_0^2 from Section 1 to obtain Q(a0)≤0Q(a_0)\le0, where

Q(x)=(1−p)x2−2px+(2e−1+p).Q(x)=(1-p)x^2-2px+\left(\frac{2}{e}-1+p\right).

The discriminant of QQ is 4(p2−(1−p)(2e−1+p))=4(2p2−2(1−1e)p+1−2e)4\left(p^2-(1-p)\left(\frac{2}{e}-1+p\right)\right)=4\left(2p^2-2\left(1-\frac{1}{e}\right)p+1-\frac{2}{e}\right), which is positive since 4(1−1e)2−8(1−2e)=4(2e+1e2−1)<04\left(1-\frac{1}{e}\right)^2-8\left(1-\frac{2}{e}\right)=4\left(\frac{2}{e}+\frac{1}{e^2}-1\right)<0. Hence, a0a_0 is not smaller than the smaller root of QQ, which is the first expression in the maximum, and the claim follows since a0>0a_0>0. □

Remark 16. Using the crude estimate an<1a_n<1 instead of an≤1−a02a_n\le1-a_0^2 in the proof of part (c), one obtains the simpler bound a0≥1p(1e−1−p2)a_0\ge\frac{1}{p}\left(\frac{1}{e}-\frac{1-p}{2}\right), which is positive exactly when p>1−2/ep>1-2/e.

Problems with further assumptions on coefficients

We continue to study the same type of problems as in the previous section, this time under some additional conditions on the coefficients, in the vein of [19]. But first we need to review a number of technical tools that will be needed.

Some estimates for the Carathéodory class

Throughout this section, we will denote by P\mathcal{P} the Carathéodory class of normalized analytic functions qq in D\mathbb{D} with positive real part:

q(z)=∑j=0∞qjzj,q0=q(0)=1.q(z)=\sum_{j=0}^{\infty}q_jz^j,\qquad q_0=q(0)=1.

It is customary to denote the functions in P\mathcal{P} by pp and their Taylor coefficients by pkp_k but here we use the letter qq to avoid confusion with the admissible weights pp of Section 3. It is well known that every function q∈Pq\in\mathcal{P} has the form

q(z)=∫02π1+e−itz1−e−itz dμ(t),q(z)=\int_0^{2\pi}\frac{1+e^{-it}z}{1-e^{-it}z}\,d\mu(t),

where μ\mu is a Borel probability measure on [0,2π)[0,2\pi) (which can also be viewed as the unit circle with an appropriate identification). This follows directly from [6 Section 1.9] upon imposing the normalization condition q(0)=1q(0)=1. The above representation is known as the Herglotz representation of qq and μ\mu is referred to as its Herglotz measure. Among the Herglotz measures, the ones of special interest to us are the ones supported at NN point masses on the unit circle {αj=e−itj:1≤j≤N}\{\alpha_j=e^{-it_j}:1\le j\le N\}, where 0≤tj<2π0\le t_j<2\pi, tj≠tkt_j\ne t_k for j≠kj\ne k. In this case,

q(z)=∑j=1Nrj1+αjz1−αjz,∑j=1Nrj=1.q(z)=\sum_{j=1}^{N}r_j\frac{1+\alpha_jz}{1-\alpha_jz},\qquad\sum_{j=1}^{N}r_j=1.

For arbitrary q∈Pq\in\mathcal{P}, by Carathéodory’s lemma [6 p. 41], we know that ∣qn∣≤2|q_n|\le2 for all n≥1n\ge1, with equality only if the support of the measure is a subset of the nn-th roots of some unimodular complex number. Livingston Livingston [16] further showed that ∣qn−qkqn−k∣≤2|q_n-q_kq_{n-k}|\le2 for all 0≤k≤n0\le k\le n, and this was later generalized in several directions by Efraimidis Efraimidis [7].

Following [7], for a complex number ww and a function qq in P\mathcal{P} we consider the (k+1)×(k+1)(k+1)\times(k+1) determinants

Ak,m(w)=∣qm+kqm+k−1qm+k−2⋯qm+1qmwq110⋯00wq2wq11⋯00⋮⋮⋮⋱⋮⋮wqk−1wqk−2wqk−3⋯10wqkwqk−1wqk−2⋯wq11∣.A_{k,m}(w)= \begin{vmatrix} q_{m+k} & q_{m+k-1} & q_{m+k-2} & \cdots& q_{m+1} & q_m\\ wq_1 & 1 & 0 & \cdots& 0 & 0\\ wq_2 & wq_1 & 1 & \cdots& 0 & 0\\ \vdots& \vdots& \vdots& \ddots& \vdots& \vdots\\ wq_{k-1} & wq_{k-2} & wq_{k-3} & \cdots& 1 & 0\\ wq_k & wq_{k-1} & wq_{k-2} & \cdots& wq_1 & 1 \end{vmatrix}.

The following was proved by Efraimidis in [7 Theorem 2].

Lemma 17. For qq in P\mathcal{P} and ww in C\mathbb{C} we have

∣Ak,m(w)∣≤2max⁡{1,∣1−2w∣k},|A_{k,m}(w)|\le2\max\{1,|1-2w|^k\},

for all integers k≥0k\ge0 and m≥1m\ge1.

If ∣1−2w∣<1|1-2w|<1, then equality holds if and only if the support of the Herglotz measure of qq is contained in a rotation {ζ:ζm+k=λ}\{\zeta:\zeta^{m+k}=\lambda\}, ∣λ∣=1|\lambda|=1, of the set of (m+k)(m+k)-th roots of unity and $q_1=q_2=\cdots=q_k=0.

In what follows, we will use Efraimidis’ estimates to prove Theorem 2. We will also need the classical Schur complement formula for determinants; see [22 p. 217].

Lemma 18. Let MM be a square matrix of complex numbers written in block form

M=(ABCD).M=\begin{pmatrix} A & B \\ C & D \end{pmatrix}.

where AA and DD are square matrices of sizes m×mm\times m and k×kk\times k, respectively. If DD is invertible, then

det⁡M=det⁡D⋅det⁡(A−BD−1C).\det M=\det D\cdot\det(A-BD^{-1}C).

Proof. The identity follows by taking determinants in the factorization

M(Im0−D−1CIk)=(A−BD−1CB0D).M\begin{pmatrix} I_m & 0 \\ -D^{-1}C & I_k \end{pmatrix} = \begin{pmatrix} A-BD^{-1}C & B \\ 0 & D \end{pmatrix}.

Lemma 19. Fix an integer n≥6n\ge6 and let l=⌊n/3⌋l=\lfloor n/3\rfloor. Let qq be a function in P\mathcal{P}, and let ww be any complex number. Assume that qj=0q_j=0 for all 1≤j≤l1\le j\le l. Then

∣qn−w∑j=l+1n−1−lqn−jqj∣≤2max⁡{1,∣1−2w∣n−l−1}.(10)\left|q_n-w\sum_{j=l+1}^{n-1-l}q_{n-j}q_j\right|\le2\max\{1,|1-2w|^{n-l-1}\}. \tag*{(10)}

If ∣1−2w∣<1|1-2w|<1, then equality holds if and only if the support of the Herglotz measure of qq is contained in a rotation {ζ:ζn=λ}\{\zeta:\zeta^n=\lambda\}, ∣λ∣=1|\lambda|=1, of the set of nn-th roots of unity and q1=q2=⋯=qn−1−l=0q_1=q_2=\cdots=q_{n-1-l}=0.

Proof. Since n≥6n\ge6, observe that n−1−l≥l+1≥3n-1-l\ge l+1\ge3. The idea of the proof is to calculate the determinant An−1−l,l+1(w)A_{n-1-l,l+1}(w) and then apply Lemma 17. If n=6n=6, then (10) becomes

∣q6−wq32∣≤2max⁡{1,∣1−2w∣3}.|q_6-wq_3^2|\le2\max\{1,|1-2w|^3\}.

Note that

A3,3(w)=∣q6q5q4q301000010wq3001∣=q6−wq32.A_{3,3}(w)= \begin{vmatrix} q_6 & q_5 & q_4 & q_3\\ 0 & 1 & 0 & 0\\ 0 & 0 & 1 & 0\\ wq_3 & 0 & 0 & 1 \end{vmatrix} =q_6-wq_3^2.

since q1=q2=0q_1=q_2=0, so there is nothing left to do.

Assume now that n≥7n\ge7. We see that

An−1−l,l+1(w)=det⁡(qnRCT).A_{n-1-l,l+1}(w)=\det\begin{pmatrix}q_n & R\\ C & T\end{pmatrix}.

where R=(qn−1qn−2⋯ql+1)R=(q_{n-1}\quad q_{n-2}\quad\cdots\quad q_{l+1}), CT=w(q1q2⋯qn−l−1)C^{T}=w(q_1\quad q_2\quad\cdots\quad q_{n-l-1}) and the (n−1−l)×(n−1−l)(n-1-l)\times(n-1-l) matrix

T=(100⋯0wq110⋯0wq2wq11⋯0⋮⋮⋮⋱⋮wqn−2−lwqn−3−lwqn−4−l⋯1).T= \begin{pmatrix} 1 & 0 & 0 & \cdots& 0\\ wq_1 & 1 & 0 & \cdots& 0\\ wq_2 & wq_1 & 1 & \cdots& 0\\ \vdots& \vdots& \vdots& \ddots& \vdots\\ wq_{n-2-l} & wq_{n-3-l} & wq_{n-4-l} & \cdots& 1 \end{pmatrix}.

Observe that TT is a triangular Toeplitz matrix and thus we may write

T=I+w∑j=1n−2−lqjJj.T=I+w\sum_{j=1}^{n-2-l}q_jJ^j.

where II is the identity matrix of size n−1−ln-1-l and JJ is the matrix of the same size whose entries are equal to 1 on the first subdiagonal and to 0 elsewhere. Since n≥7n\ge7, then n−2−l≥l+1n-2-l\ge l+1 and the hypothesis q1=q2=⋯=ql=0q_1=q_2=\cdots=q_l=0 implies that

T=I+w∑j=l+1n−2−lqjJj.T=I+w\sum_{j=l+1}^{n-2-l}q_jJ^j.

By definition of ll, there exists a unique s∈{0,1,2}s\in\{0,1,2\} such that n=3l+sn=3l+s. Thus, n−1−l=2l−1+s≤2(l+1)n-1-l=2l-1+s\le2(l+1) and, since Jn−1−l=0J^{n-1-l}=0, we have that

(w∑j=l+1n−2−lqjJj)2=0.\left(w\sum_{j=l+1}^{n-2-l}q_jJ^j\right)^2=0.

Summing up, we have that TT must be invertible and

T−1=I−w∑j=l+1n−2−lqjJj.(11)T^{-1}=I-w\sum_{j=l+1}^{n-2-l}q_jJ^j. \tag*{(11)}

Recall that, from the hypothesis, the first ll components of CC are zero. Since JjCJ^{j}C is obtained from CC by shifting its entries jj positions downwards and 2l+1≥n−1−l2l+1 \ge n-1-l, it follows that JjC=0J^{j}C=0 for all l+1≤j≤n−2−ll+1 \le j \le n-2-l and thus T−1C=CT^{-1}C=C by (11). Then Lemma 18 yields that

An−l−1,l+1(w)=det⁡T⋅det⁡(qn−RT−1C)=qn−RC=qn−w∑j=l+1n−l−1qn−jqj.A_{n-l-1,l+1}(w)=\det T\cdot\det(q_n-RT^{-1}C)=q_n-RC=q_n-w\sum_{j=l+1}^{n-l-1}q_{n-j}q_j.

(recall that det⁡T=1\det T=1). The proof now follows by applying Lemma 17 with k=n−1−lk=n-1-l and m=l+1m=l+1.

Next, we show that the bound on the right-hand side of (10) can be improved when w≥1w\ge1 is real.

Lemma 20. Under the assumptions of Lemma 19, if w≥1w\ge1 is real, then

∣qn−w∑j=l+1n−l−1qn−jqj∣≤2(2w−1).\left|q_n-w\sum_{j=l+1}^{n-l-1}q_{n-j}q_j\right|\le2(2w-1).

Proof. For w=1w=1, (10) is exactly what we want to show. If w>1w>1, we can apply the case w=1w=1, Carathéodory’s lemma, and the triangle inequality to see that

∣qn−w∑j=l+1n−l−1qn−jqj∣=∣−(w−1)qn+w(qn−∑j=l+1n−l−1qn−jqj)∣≤2(w−1)+2w=2(2w−1).\left|q_n-w\sum_{j=l+1}^{n-l-1}q_{n-j}q_j\right| =\left|-(w-1)q_n+w\left(q_n-\sum_{j=l+1}^{n-l-1}q_{n-j}q_j\right)\right| \le2(w-1)+2w=2(2w-1).

Equivalent formulations of the generalized conjecture

Since the Krzyż conjecture is believed to be true by many experts, we expect the following more general conjecture for problem (6) to be true.

Conjecture. Let n≥1n\ge1 and let Φp\Phi_p be an admissible functional. Then Gn(p)=2/eG_n(p)=2/e and the only extremal function for (6) is fn(z)=eznzn+1f_n(z)=e^{\frac{z^n}{z^{n+1}}}.

By Corollary 12, the conjecture holds for n≤5n\le5, and by Theorem 1 (a), it holds for every nn for which the Krzyż conjecture is true. This raises the natural question as to whether some parts of the main theorem from [19] (see Theorem A) can also be formulated in this more general context.

For a non-constant function ff in B∗\mathcal{B}^* we may write f=egf=e^g where Re⁡g(z)<0\operatorname{Re}g(z)<0 for all zz in D\mathbb{D}. Writing g(z)=∑j=0∞bjzjg(z)=\sum_{j=0}^{\infty}b_jz^j and assuming that b0b_0 is real and negative, it is immediate that the normalized function qq given by q(z)=g(z)/b0q(z)=g(z)/b_0 belongs to P\mathcal{P}; in particular, ∣bn∣≤2∣b0∣|b_n|\le2|b_0| for all n≥1n\ge1 by Carathéodory’s lemma. The relation between the coefficients of ff and gg is expressed by the following well-known result; cf. [19], for example.

Lemma 21. Let ff and gg be analytic functions in D\mathbb{D} such that f=egf=e^g, given by their respective power series f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j and g(z)=∑j=0∞bjzjg(z)=\sum_{j=0}^{\infty}b_jz^j. Then

(a) a0=eb0a_0=e^{b_0} and

an=∑k=0n−1n−knakbn−k=∑j=1njnan−jbj,n≥1.a_n=\sum_{k=0}^{n-1}\frac{n-k}{n}a_kb_{n-k}=\sum_{j=1}^{n}\frac{j}{n}a_{n-j}b_j,\qquad n\ge1.

(b) For a fixed integer j≥1j\ge1, b1=b2=⋯=bj=0b_1=b_2=\cdots=b_j=0 holds if and only if a1=a2=⋯=aj=0a_1=a_2=\cdots=a_j=0.

The following formula was described structurally in [19 Proposition 1] but had actually been known earlier to Lewandowski and Szynal [15 Lemma 1], whose detailed survey may not have received all the attention it deserved. The formula was stated there in a more precise form as follows.

Lemma 22. Let ff and gg be analytic functions in D\mathbb{D} such that f=egf=e^{g}, given by their respective power series f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j and g(z)=∑j=0∞bjzjg(z)=\sum_{j=0}^{\infty}b_jz^j. Assuming that b0b_0 is real and negative and writing t=−b0>0t=-b_0>0 and qk=bk/b0q_k=b_k/b_0 for k≥1k\geq1, we have

an=e−t∑k1+2k2+⋯+nkn=n,0≤kj(−t)k1+k2+⋯+knq1k1q2k2⋯qnknk1!k2!⋯kn!.a_n=e^{-t}\sum_{\substack{k_1+2k_2+\cdots+nk_n=n,\\0\leq k_j}}\frac{(-t)^{k_1+k_2+\cdots+k_n}q_1^{k_1}q_2^{k_2}\cdots q_n^{k_n}}{k_1!k_2!\cdots k_n!}.

We now have all the ingredients needed to prove Theorem 2, formulated in § 1.

Proof of Theorem 2. One implication is trivial, hence it suffices to prove that (a) implies (b). Assume that (a) holds. Since f∈Bnf\in\mathcal{B}_n, it is a non-constant function in B∗\mathcal{B}_* with f(0)=a0>0f(0)=a_0>0, so it can be written as f=egf=e^g, where g(z)=∑j=0∞bjzjg(z)=\sum_{j=0}^{\infty}b_jz^j is analytic in D\mathbb{D}, Re⁡g<0\operatorname{Re}g<0, and b0=log⁡a0b_0=\log a_0 is real and negative. Writing again t=−b0>0t=-b_0>0 and qk=bk/b0q_k=b_k/b_0 for k≥1k\geq1 as in Lemma 22, we have g=−tg0g=-tg_0, where g0∈Pg_0\in\mathcal{P} and g0(z)=1+∑j≥1qjzjg_0(z)=1+\sum_{j\geq1}q_jz^j. By our assumption (a) and Lemma 21 (b), we have b1=⋯=bl=0b_1=\cdots=b_l=0, that is, q1=⋯=ql=0q_1=\cdots=q_l=0, so g0−1g_0-1 has a zero of order at least l+1l+1 at the origin. Hence, Lemma 22 yields

an=e−t∑(l+1)kl+1+⋯+nkn=n,0≤kj(−t)kl+1+⋯+knql+1kl+1⋯qnknkl+1!⋯kn!.(12)a_n=e^{-t}\sum_{\substack{(l+1)k_{l+1}+\cdots+nk_n=n,\\0\leq k_j}}\frac{(-t)^{k_{l+1}+\cdots+k_n}q_{l+1}^{k_{l+1}}\cdots q_n^{k_n}}{k_{l+1}!\cdots k_n!}. \tag*{(12)}

We will show that (l+1)kl+1+⋯+nkn=n(l+1)k_{l+1}+\cdots+nk_n=n for an (n−l)(n-l)-tuple of indices (kl+1,…,kn)(k_{l+1},\ldots,k_n) if and only if either (kl+1,…,kn)=(0,…,0,1)(k_{l+1},\ldots,k_n)=(0,\ldots,0,1) or there exists j∈{l+1,…,n−1−l}j\in\{l+1,\ldots,n-1-l\} such that kj=kn−j=1k_j=k_{n-j}=1 if j≠n−jj\ne n-j (respectively, kn/2=2k_{n/2}=2 if j=n/2j=n/2) and all the remaining indices are zero. These are exactly the solutions with ∑j=l+1nkj≤2\sum_{j=l+1}^{n}k_j\leq2; if ∑j=l+1nkj≥3\sum_{j=l+1}^{n}k_j\geq3, we would have that

n=∑j=l+1njkj≥(l+1)∑j=l+1nkj≥3(l+1),n=\sum_{j=l+1}^{n}jk_j\geq(l+1)\sum_{j=l+1}^{n}k_j\geq3(l+1),

which is impossible by the choice of ll. Then (12) becomes

an=−te−t(qn−t2∑j=l+1n−1−lqjqn−j).a_n=-te^{-t}\left(q_n-\frac{t}{2}\sum_{j=l+1}^{n-1-l}q_jq_{n-j}\right).

(note that every product qjqn−jq_jq_{n-j} with j≠n−jj\ne n-j appears twice in the sum, while for even nn the term qn/22q_{n/2}^2 appears once and comes with the factor 1/2!1/2! in (12)). We may apply Lemma 19 for 0<t≤20<t\leq2, with w=t/2w=t/2 (note that ∣1−2w∣=∣1−t∣≤1|1-2w|=|1-t|\leq1) and Lemma 20 for t≥2t\geq2, also with w=t/2w=t/2, to conclude that

an=∣an∣≤2te−tmax⁡{1,t−1}.a_n=|a_n|\leq2te^{-t}\max\{1,t-1\}.

A straightforward computation shows that

max⁡{te−t:0<t≤2}=e−1>(2+5)e−(3+5)/2=max⁡{t(t−1)e−t:t≥2}.\max\{te^{-t}:0<t\leq2\}=e^{-1}>(2+\sqrt{5})e^{-(3+\sqrt{5})/2}=\max\{t(t-1)e^{-t}:t\geq2\}.

The maximum of the expression on the left is only attained when t=1t=1, while the maximum on the right is attained only at t=(3+5)/2t=(3+\sqrt{5})/2. Thus, an≤2ea_n\leq\frac{2}{e}, and Gn(p)=2eG_n(p)=\frac{2}{e} readily follows from the chain of inequalities (the first one holds since fn∈Bnf_n \in\mathcal{B}_n and Φp(fn)=2/e\Phi_p(f_n)=2/e)

2e≤Gn(p)=Φp(f)≤an.\frac{2}{e} \le G_n(p)=\Phi_p(f)\le a_n.

Moreover, an=2/ea_n=2/e, so equality must hold in all the estimates above; in particular, t=1t=1 and equality holds in (10) with w=12w=\frac{1}{2}.

To finish the proof, it remains to show that f=fnf=f_n. Since f∈Bnf\in\mathcal{B}_n and t=1t=1, we can write

g0(z)=∑j=1Nrj1+αjz1−αjzg_0(z)=\sum_{j=1}^{N}r_j\frac{1+\alpha_jz}{1-\alpha_jz}

for certain N≤nN\le n, rj>0r_j>0 with ∑j=1Nrj=1\sum_{j=1}^{N}r_j=1, and distinct points αj∈T\alpha_j\in\mathbb{T}. Since equality holds in (10) for g0g_0 with w=12w=\frac{1}{2}, Lemma 19 shows that q1=q2=⋯=qn−1−l=0q_1=q_2=\cdots=q_{n-1-l}=0 (hence, b1=b2=⋯=bn−1−l=0b_1=b_2=\cdots=b_{n-1-l}=0) and that there is λ∈T\lambda\in\mathbb{T} with αjn=λ\alpha_j^n=\lambda for all 1≤j≤N1\le j\le N. The power series expansion of 1+αjz1−αjz\frac{1+\alpha_jz}{1-\alpha_jz} yields

bn−k=−2∑j=1Nrjαjn−k=λbk‾,1≤k≤n−1,b_{n-k}=-2\sum_{j=1}^{N}r_j\alpha_j^{n-k}=\lambda\overline{b_k},\qquad1\le k\le n-1,

hence bl+1=⋯=bn−1=0b_{l+1}=\cdots=b_{n-1}=0 as well. Now Lemma 21 (b) readily implies that a1=⋯=an−1=0a_1=\cdots=a_{n-1}=0. Once we know this, the fact that f(z)=ezn−1zn+1f(z)=e^{\frac{z^{n-1}}{z^n+1}} is proved by following verbatim the proof of the implication (b) ⟹\Longrightarrow (c) of [19 Theorem 1], based on a simple use of a system of linear equations and the properties of the Vandermonde determinant. □

As already mentioned, the case p≡0p\equiv0 of Theorem 2 is the following result of MacDonald [17]; our proof of it is independent of the one in [17].

Corollary 23. Let n≥6n\ge6 and let f(z)=∑j=0∞ajzjf(z)=\sum_{j=0}^{\infty}a_jz^j be a normalized extremal function for the Krzyż problem (1). If a1=⋯=a⌊n/3⌋=0a_1=\cdots=a_{\lfloor n/3\rfloor}=0, then Mn=2/eM_n=2/e and f(z)=ezn−1zn+1f(z)=e^{\frac{z^{n-1}}{z^n+1}}.

Proof. By Theorem A, f∈Bnf\in\mathcal{B}_n and ff is an extremal function for (6) with p≡0p\equiv0, for which Gn(0)=MnG_n(0)=M_n. Thus, the statement is the case p≡0p\equiv0 of Theorem 2. □

It should be noted that if the recent proof [26] is confirmed for n=6n=6, then this case in Theorem 2 will no longer represent a novelty.

AI disclosure

This project was started in 2024, well before the most recent advances in AI; AI tools were used only in the final stages of our work.

ChatGPT 5.6 Sol was used to simplify our original proof of Lemma 19. More precisely, it suggested the use of the Schur complement formula (Lemma 18) to simplify the notation. The original statement of Theorem 2 (a) included the extra assumption a0≥1/e2a_0\ge1/e^2. ChatGPT 6 Astra provided the observation to be found in Lemma 20, which was then used to remove this extra condition. Example 10 in Section 3 was also suggested by ChatGPT 5.6 Sol.

Claude Opus 5.5 (Anthropic) and ChatGPT 6 Astra were used to proofread the manuscript. The former suggested various useful simplifications, improvements in exposition, and observations such as Remark 14.

Acknowledgments

The three authors are partially supported by PID2025-168349NB-I00 from Ministerio de Innovación, Ciencia y Universidades, Spain.

This work was partially supported by the project Pure Mathematics in Norway, funded by Trond Mohn Foundation and Tromsø Research Foundation. This work started when the first author was visiting Universidad Autónoma de Madrid and the first author would like to thank the university for the hospitality as well as Norges tekniske høgskoles fond for covering parts of this travel.

The second author is partially supported by PID2024-160326NA-I00, also from MICIU, and by CIGE/2026/195 from Conselleria d’Educació, Cultura i Universitats, Generalitat Valenciana, Spain.

Department of Mathematical Sciences, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway Email address: [email protected]

Departamento de Matemáticas, Universidad Autónoma de Madrid, 28049 Madrid, Spain Email address: [email protected] URL: https://sites.google.com/view/adrian-llinares/home

Departamento de Matemáticas, Universidad Autónoma de Madrid, 28049 Madrid, Spain Email address: [email protected] URL: http://www.uam.es/dragan.vukotic

References

  1. [1]M. Abramowitz, I.A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables, Dover, New York, 1970.
  2. [2]J. Agler, J.E. McCarthy, The Krzyż conjecture and an entropy conjecture, J. Anal. Math. 144 (2021), 207–226.DOI
  3. [3]C. Bénéteau and D. Khavinson, A survey of certain extremal problems for non-vanishing analytic functions, In: Complex and harmonic analysis, 45–61, DEStech Publ., Inc., Lancaster, PA, 2007.
  4. [4]J.E. Brown, On a coefficient problem for nonvanishing H^p functions, Complex Var. Theory Appl. 4 (1985), no. 2–3, 253–265.DOI
  5. [5]J.E. Brown, Iteration of functions subordinate to schlicht functions, Complex Var. Theory Appl. 9 (1987), 143–152.DOI
  6. [6]P.L. Duren, Univalent Functions, Springer-Verlag, New York, 1983.
  7. [7]I. Efraimidis, A generalization of Livingston’s coefficient inequalities for functions with positive real part, J. Math. Anal. Appl. 435 (2016), 369–379.DOI
  8. [8]I. Efraimidis, A. Llinares, M.J. Martín, E.T. Sawyer, I. Uriarte-Tuero, D. Vukotić, On extremal functions for the Krzyż conjecture, preprint, 2026.
  9. [9]R.J.P.M. Ermers, Coefficient estimates for bounded nonvanishing functions, Doctoral Thesis, Katholieke Universiteit Nijmegen, Wibro Dissertatiedrukkerij, Helmond, 1990.
  10. [10]C. Horowitz, Coefficients of nonvanishing functions in H∞, Israel J. Math. 30 (1978), 285–291.
  11. [11]J.A. Hummel, S. Scheinberg, and L. Zalcmann, A coefficient problem for bounded nonvanishing functions, J. Anal. Math. 31 (1977), 169–190.DOI
  12. [12]S.Ya. Khavinson, Two papers on extremal problems, AMS Translations, Providence, Rhode Island, 1986.DOI
  13. [13]J. Krzyż, Coefficient problem for bounded non-vanishing functions, Ann. Polon. Math. 20 (1968), 314–316.
  14. [14]J. Lei, T. Zhang, Proof of the Agler–McCarthy entropy conjecture, preprint, 2026, https://arxiv.org/abs/2605.03949.
  15. [15]Z. Lewandowski, J. Szynal, On the Krzyż conjecture and related problems, XIVth Rolf Nevanlinna Colloquium (Joensuu, 1995), 257–268, de Gruyter, Berlin, 1996.
  16. [16]A.E. Livingston, A coefficient inequality for functions of positive real part with an application to multivalent functions, Pacific J. Math. 120 (1985), no. 1, 139–151.DOI
  17. [17]S.F. MacDonald, Structural aspects of extremal functions in the Krzyż conjecture, preprint, 2026, https://arxiv.org/abs/2605.08411.
  18. [18]D.E. Marshall, personal communication with the third author.
  19. [19]M.J. Martín, E.T. Sawyer, I. Uriarte-Tuero, D. Vukotić, The Krzyż conjecture revisited, Adv. Math. 273 (2015), 716–745.DOI
  20. [20]S.V. Romanova, Estimates for functionals of bounded nonvanishing functions [Russian], Sibirsk. Mat. Zh. 44 (2003), no. 5, 1113–1123; translation in Siberian Math. J. 44 (2003), no. 5, 869–876.DOI
  21. [21]N. Samarias, A proof of Krzyż’s conjecture for the fifth coefficient, Complex Var. Theory Appl. 48 (2003), 753–766.
  22. [22]J. Schur, Über Potenzreihen die im Inneren des Einheitskreises beschränkt sind, J. Reine Angew. Math. 147 (1917), 205–232.
  23. [23]W. Szapiel, A new approach to the Krzyz conjecture, Workshop on Complex Analysis (Lublin, 1994), Ann. Univ. Mariae Curie-Sklodowska Sect. A 48 (1994), 169–192.
  24. [24]D.L. Tan, Coefficient estimates for bounded nonvanishing functions (Chinese), Chinese Ann. Math. Ser. A 4 (1983), 97–104.
  25. [25]M. Tsuji, Potential Theory in Modern Function Theory, Maruzen, Tokyo, 1959.
  26. [26]F. Wikström, A computer-assisted proof of Krzyż’s conjecture for the sixth coefficient, preprint, 2026, https://arxiv.org/abs/2610.15206.

Paper details

vlnc:W_t2szbyze
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

GPT-5.6 Sol (bare), GPT-6 Astra (bare), Claude Opus 5.5 (bare)

View statement
Looking for related papers…
Contents