A remark on Kolmogorov’s theorem


Abstract

The aim of this note is to illustrate that the set of integrable functions on the torus \(\mathbb{T}\) for which the Fourier series diverges almost everywhere is of the second Baire category in \(L^1(\mathbb{T})\).

1

1 Introduction↩︎

1.1 A very brief history↩︎

In 1923, Kolmogorov [1], as a nineteen-year-old student, constructed an integrable function on the torus \(\mathbb{T}:=\mathbb{R}/\mathbb{Z}\) whose Fourier series diverges almost everywhere. Three years later, Kolmogorov [2] improved on this by creating an integrable function on \(\mathbb{T}\) whose Fourier series diverges everywhere on \(\mathbb{T}\). Kolmogorov’s seminal papers [1], [2] initiated new directions of research in Fourier series and engendered an extensive literature on diverging Fourier series. An important simplification and several refinements of Kolmogorov’s examples appear in Stein’s article [3], which also covers the case of the Walsh–Paley series. For a comprehensive historical background and a thorough bibliography on the subject, we refer to Ul’yanov’s survey [4] as well as to the books of Zygmund [5] and Grafakos [6]. In particular, Grafakos’ books [6], [7] offer a modern treatment to the subject.

Kolmogorov’s results arose from his attempt to disprove Luzin’s conjecture from 1913, which asserts that the partial sums of the Fourier series of a square-integrable function on \(\mathbb{T}\) converge almost everywhere. At that time, there was a strong belief that Luzin’s conjecture was likely false.

However, four decades after Kolmogorov’s papers, Carleson [8] proved Luzin’s conjecture in the affirmative for \(L^2(\mathbb{T})\) functions and Hunt [9] subsequently extended this result to all \(L^p(\mathbb{T})\) functions with \(p \in (1, \infty)\). We refer to Fefferman’s paper [10] and to Lacey and Thiele’s paper [11] for different proofs of Carleson’s theorem, as well as to Grafakos [7], where the Carleson–Hunt theorem is presented in great detail for \(L^p(\mathbb{T})\) functions with \(p \in (1, \infty)\).

Collectively, Kolmogorov’s papers [1], [2] and the Carleson–Hunt theorem [8], [9] reveal that Kolmogorov’s examples are obstructions which characterize \(L^1(\mathbb{T})\)-space phenomena for Fourier series. This advances a natural question: how large topologically (in the sense of Baire category) is the set of all integrable functions \(f \in L^1(\mathbb{T})\) whose Fourier series diverges almost everywhere on \(\mathbb{T}\)?

In this short note, we give an ad-hoc argument based on Kolmogorov’s methods [1], [2], (see Proposition 1 below), illustrating that the set of integrable functions whose Fourier series diverge almost everywhere on \(\mathbb{T}\) is topologically generic in \(L^1(\mathbb{T})\). In other words, this property holds on a residual set, i.e., a countable intersection of dense open sets. Loosely speaking, Kolmogorov’s pathological examples are topologically quite common in \(L^1(\mathbb{T})\).

Acknowledgments↩︎

We thank Leonidas Daskalakis, Jacek Dziubański, Dariusz Kosz and Jim Wright for helpful comments on an earlier draft of this paper.

1.2 Notation and statement of the main result↩︎

Throughout the paper the set of positive integers and the set of nonnegative integers will be denoted, respectively, by \(\mathbb{Z}_+ \mathrel{\vcenter{:}}= \{1, 2, \ldots\}\) and \(\mathbb{N} \mathrel{\vcenter{:}}= \{0,1,2,\ldots\}\). The sets \(\mathbb{Z}\), \(\mathbb{R}\), and \(\mathbb{C}\) have their standard meaning. We also denote \(\mathbb{R}_+ \mathrel{\vcenter{:}}= (0, \infty)\). The torus \(\mathbb{T} := \mathbb{R}/\mathbb{Z}\) is endowed with the Lebesgue measure \(d\mu_{\mathbb{T}}(x):=dx\) inherited from the real line \(\mathbb{R}\), and \(\mu_{\mathbb{T}}(E):=\int_Edx\) denotes the Lebesgue measure of a Lebesgue measurable set \(E \subseteq \mathbb{T}\).

All vector spaces here will be defined over the complex numbers \(\mathbb{C}\). The space of all measurable functions \(f:\mathbb{T}\to \mathbb{C}\) whose modulus is integrable with \(p\)-th power is denoted by \(L^p(\mathbb{T})\) for \(p\in[1, \infty)\), whereas \(L^{\infty}(\mathbb{T})\) denotes the space of all essentially bounded measurable functions on \(\mathbb{T}\).

Let \(e(x):=e^{2\pi i x}\) for \(x\in \mathbb{R}\). For \(f\in L^1(\mathbb{T})\) the Fourier coefficients are defined by \[\widehat{f}(n):=\int_{\mathbb{T}}f(x)e(-x n)dx \quad \text{ for } \quad n\in\mathbb{Z}.\] The partial sums of the Fourier series corresponding to \(f\in L^1(\mathbb{T})\) are defined by \[S_N(f)(x):=\sum_{n\in[-N, N]\cap \mathbb{Z}}\widehat{f}(n)e(x n)\quad \text{ for } \quad x\in\mathbb{T}, \;N\in\mathbb{N}.\] Using the Dirichlet kernel \(D_N(x):=\sum_{n\in[-N, N]\cap \mathbb{Z}}e(x n)\), we see that \[S_N(f)(x)=D_N*f(x)=\int_{\mathbb{T}}f(x-y)D_N(y)dy.\]

The main theorem of the paper reads as follows.

Theorem 1. The set \[\begin{align} \label{eq:1} \mathcal{D}:=\big\{f\in L^1(\mathbb{T}): \mu_{\mathbb{T}}(\{x\in\mathbb{T}: \sup_{N\in\mathbb{N}}|D_N*f(x)|=\infty\})=1\big\} \end{align}\tag{1}\] is of the second Baire category in \(L^1(\mathbb{T})\). In particular, the set of all integrable functions whose Fourier series diverges almost everywhere on \(\mathbb{T}\) is of the second Baire category in \(L^1(\mathbb{T})\).

From Theorem 1 it follows immediately that there exist integrable functions whose Fourier series diverge almost everywhere, in fact with unbounded partial sums; however, our proof should not be regarded as an alternative method for producing Kolmogorov’s examples [1], [2], but rather as a complement to his approach, providing a topological characterization of this phenomenon.

Interestingly, Marcinkiewicz [12] constructed an integrable function on \(\mathbb{T}\) whose Fourier series diverges almost everywhere, yet whose partial sums are bounded at every point of \(\mathbb{T}\).

We will prove Theorem 1 by showing that \(\mathcal{D}^c\) is of the first category in the sense of Baire category, in other words it can be written as a countable union of nowhere dense sets in \(L^1(\mathbb{T})\).

A key tool will be the following result of Kolmogorov [1], [2].

Proposition 1. For each \(M\in\mathbb{R}_+\) there exists a trigonometric polynomial \(g_M\) and a measurable set \(G_M \subseteq \mathbb{T}\) with measure \(\mu_{\mathbb{T}}(G_M) > 1 - 2^{-M}\) such that \(\|g_M\|_{L^1(\mathbb{T})}=1\), and satisfying \[\begin{align} \label{eq:5} \inf_{x\in G_M}\sup_{k\in\mathbb{Z}_+}|D_k*g_M(x)|>2^M. \end{align}\qquad{(1)}\]

Proof. For a detailed proof we refer to [6]. ◻

Proposition 1 (see also [5]) was a key mechanism behind the construction of Kolmogorov’s [1], [2] explicit examples; it will also be essential in our categorical argument.

For \(m, n\in\mathbb{Z}_+\), define \[\begin{align} \label{eq:4} F(m, n):=\Big\{f\in L^1(\mathbb{T}): \mu_{\mathbb{T}}(\{x\in\mathbb{T}: \sup_{N\in\mathbb{N}}|D_N*f(x)|>n\})\le \frac{m}{m+1}\Big\}. \end{align}\tag{2}\]

Lemma 1. The sets \(F(m, n)\) defined in 2 are closed in \(L^1(\mathbb{T})\) for every \(m, n\in\mathbb{Z}_+\).

Proof. Fix \(m, n\in\mathbb{Z}_+\), equivalently it suffices to show that \(F(m, n)^c\) is open in \(L^1(\mathbb{T})\). Pick \(f\in F(m, n)^c\) and we show that there exists \(\eta>0\) such that for any \(g\in L^1(\mathbb{T})\) when \(\|f-g\|_{L^1(\mathbb{T})} < \eta\), then \(g\in F(m, n)^c\). Define two sets \[\begin{align} E_N(f)&:=\big\{x \in\mathbb{T}: \sup_{1 \leq k \leq N}|D_k * f(x)|>n\big\},\\ E_\infty(f)&:=\big\{x \in\mathbb{T}: \sup_{k\in\mathbb{N}}|D_k * f(x)|>n\big\}. \end{align}\] Observe that \(E_N\) is an increasing sequence of sets, that is \(E_{N} \subseteq E_{N+1}\). Using continuity from below of the measure, we have \(\lim_{N\to \infty}\mu_{\mathbb{T}}(E_N(f)) = \mu_{\mathbb{T}}(E_\infty(f))\). Since \(f\in F(m, n)^c\), it follows that there exists some \(N \in \mathbb{N}\) such that \(\mu_{\mathbb{T}}(E_N(f)) > \frac{m}{m+1}\). Define another set \[\begin{align} E_{N, \delta}(f):=\big\{x \in\mathbb{T}: \sup_{1 \leq k \leq N}|D_k * f(x)| > n+\delta\big\}. \end{align}\] We note that \(E_{N, \delta_2}(f)\subseteq E_{N, \delta_1}(f)\) for \(\delta_2>\delta_1>0\). Hence, we can choose \(\delta>0\) such that \(\mu_{\mathbb{T}}(E_{N, \delta}(f)) > \frac{m}{m+1}+\delta\). Set \(\eta:= \min_{1 \leq k \leq N} \delta^22^{-k}\|D_k\|_{L^1(\mathbb{T})}^{-1}\) and observe that for any \(g\in L^1(\mathbb{T})\) satisfying \(\|f-g\|_{L^1(\mathbb{T})} < \eta\), we obtain, for any \(k\in\{1, \ldots, N\}\), by Chebyshev’s inequality followed by Young’s convolution inequality that \[\begin{align} \mu_{\mathbb{T}}\big(\big\{x \in\mathbb{T}: |D_k *(f-g)(x)| \geq \delta\big\}\big)\le \frac{1}{\delta}\|D_k\|_{L^1(\mathbb{T})}\|f-g\|_{L^1(\mathbb{T})}<\frac{\delta}{2^k}. \end{align}\] Taking \[\begin{align} B:=\bigcup_{k=1}^N\{x\in \mathbb{T}:\left|D_k *(f-g)(x)\right| \geq \delta\}, \end{align}\] we see that \(\mu_{\mathbb{T}}(B)\le \sum_{k=1}^N\frac{\delta}{2^k}<\delta\). Moreover, we have \[\begin{align} E_{N, \delta}(f)\subseteq \big\{x \in\mathbb{T}: \sup_{1 \leq k \leq N}|D_k * g(x)|>n\big\}\cup B. \end{align}\] Therefore, \(\mu_{\mathbb{T}}(E_{N}(g))\ge \mu_{\mathbb{T}}(E_{N, \delta}(f))-\mu_{\mathbb{T}}(B)> \frac{m}{m+1}+\delta-\delta=\frac{m}{m+1}\), which implies that \(g\in F(m, n)^c\). This in turn yields that \(F(m, n)^c\) is open and the proof of the lemma follows. ◻

Lemma 2. The sets \(F(m, n)\) defined in 2 have empty interiors for every \(m, n\in\mathbb{Z}_+\).

Proof. We fix \(m, n\in\mathbb{Z}_+\) and we will show that \(F(m, n)\) has empty interior. Equivalently, we prove that for all \(f \in F(m, n)\) and for all \(\varepsilon>0\), there exists \(h \in L^1(\mathbb{T})\) such that \(\|f-h\|_{L^1(\mathbb{T})} < \varepsilon\) and \(h \not \in F(m, n)\). By a simple density argument, there exists a trigonometric polynomial \(g\in L^1(\mathbb{T})\) such that \(\|f-g\|_{L^1(\mathbb{T})} < \varepsilon/2\). Then there exists a finite constant \(K\in\mathbb{R}_+\) such that \[\begin{align} \sup_{x\in\mathbb{T}}\sup _{N \in \mathbb{Z}_+}|D_N * g(x)| \leq K. \end{align}\] Choose \(M\in\mathbb{Z}_+\) so that \[\begin{align} 2^{-M/2} < \frac{\varepsilon}{2}, \quad \text{ and } \quad 2^{M/2} > n+K, \quad \text{ and } \quad 1-2^{-M}>\frac{m}{m+1}. \end{align}\] Define the perturbed function \(h:= g + 2^{-M/2}g_M\), where \(g_M\) is a polynomial as in Proposition 1. By our choice of \(M\in\mathbb{Z}_+\) we see that \(\|f-h\|_{L^1(\mathbb{T})} < \varepsilon\), since \(\|g_M\|_{L^1(\mathbb{T})}=1\) and \[\begin{align} \|f-h\|_{L^1(\mathbb{T})} \leq \|f-g\|_{L^1(\mathbb{T})} + \|g-h\|_{L^1(\mathbb{T})} = \|f-g\|_{L^1(\mathbb{T})}+ 2^{-M/2} \|g_M\|_{L^1(\mathbb{T})} < \frac{\varepsilon}{2}+ 2^{-M/2}. \end{align}\] It suffices to show that \(h \not \in F(m, n)\). Indeed, if \(x \in G_M\), then by the triangle inequality, we have \[\begin{align} \sup_{k\in\mathbb{Z}_+}|D_k * h(x)|&=\sup_{k\in\mathbb{Z}_+}|D_k * (g+2^{-M / 2} g_M)(x)| \\ &\geq 2^{-M / 2} \sup_{k\in\mathbb{Z}_+}|D_k * g_M(x)|-\sup_{k\in\mathbb{Z}_+}|D_k * g(x)| \\ & > 2^{-M/2} 2^M - K = 2^{M/2} -K. \end{align}\] By our choice of \(M\in\mathbb{Z}_+\), we conclude that \[\begin{align} \sup_{k\in\mathbb{Z}_+}|D_k * h(x)| > n, \end{align}\] which also guarantees that \(G_M \subseteq \{x\in\mathbb{T} : \sup_{k\in\mathbb{Z}_+}|D_k * h)(x)| >n\}\), and yields \[\begin{align} \mu_{\mathbb{T}}\big(\big\{x \in\mathbb{T}: \sup_{k\in\mathbb{Z}_+}|D_k * h(x)|>n\big\}\big)\ge \mu_{\mathbb{T}}(G_M)>1-2^{-M}>\frac{m}{m+1}. \end{align}\] Therefore, \(h \not \in F(m, n)\) and consequently \(F(m, n)\) has empty interior as desired. ◻

Proof of Theorem 1. Observe that \[\begin{align} \mathcal{D}^c= \big\{f\in L^1(\mathbb{T}): \mu_{\mathbb{T}}(\{x\in\mathbb{T}: \sup_{N\in\mathbb{N}}|D_N*f(x)|=\infty\})<1\big\}. \end{align}\] For \(f\in L^1(\mathbb{T})\), define \[\begin{align} A_n:=\{x \in\mathbb{T}: \sup _{N \in \mathbb{N}}|D_N * f(x)|>n\} \quad \text{ for } \quad n\in\mathbb{Z}_+. \end{align}\] Observe that \(A_n\) is a decreasing sequence of sets, that is \(A_{n+1} \subseteq A_n\). Further \(\mu_{\mathbb{T}}(A_1) \leq 1\). So, we may invoke continuity from above of the measure to see that \[\begin{align} \mu_{\mathbb{T}}(\{x\in\mathbb{T}: \sup_{N\in\mathbb{N}}|D_N*f(x)|=\infty\})=\lim_{n\to \infty}\mu_{\mathbb{T}}(A_n)\le 1. \end{align}\] Hence, we obtain \[\begin{align} \mathcal{D}^c\subseteq \bigcup_{n\in\mathbb{Z}_+} \big\{f\in L^1(\mathbb{T}): \mu_{\mathbb{T}}(\{x\in\mathbb{T}: \sup_{N\in\mathbb{N}}|D_N*f(x)|>n\})<1\big\}. \end{align}\] Consequently, using the sets \(F(m, n)\) from 2 , we can further write \[\begin{align} \label{eq:3} \mathcal{D}^c\subseteq \bigcup_{n\in\mathbb{Z}_+}\bigcup_{m\in\mathbb{Z}_+} F(m, n). \end{align}\tag{3}\] The set \(\bigcup_{n\in\mathbb{Z}_+}\bigcup_{m\in\mathbb{Z}_+} F(m, n)\) is of the first category, since each \(F(m, n)\) is closed and has empty interior by Lemma 1 and Lemma 2. This ensures that \(\mathcal{D}^c\) is of the first category as a subset of a first category set in 3 . The proof of Theorem 1 now follows. ◻

References↩︎

[1]
A.N. Kolmogorov.
[2]
A.N. Kolmogorov.
[3]
E.M. Stein.
[4]
P. L. Ul’yanov.
[5]
A. Zygmund.
[6]
L. Grafakos.
[7]
L. Grafakos.
[8]
L. Carleson.
[9]
R. Hunt.
[10]
C. Fefferman.
[11]
M. Lacey, C. Thiele.
[12]
J. Marcinkiewicz.

  1. The authors were partially supported by the NSF CAREER grant (DMS-2236493).↩︎