Fourier convergence at the L log L threshold

← All digests

Family 075 · Real and complex analysis
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-14

1. The question

A periodic signal can be decomposed into pure frequencies. If we add back successively more of those frequencies, do the resulting Fourier sums approach the original signal at almost every point?

For smooth signals, this is a familiar success of Fourier analysis. For rough functions, summing more frequencies can produce large oscillations rather than straightforward improvement. The twentieth-century theory established convergence under $L^p$ integrability for every $p>1$, but ordinary integrability alone is insufficient.

Family 075 claims the classical intermediate sufficiency conjecture: every function in $L\log L$ has ordinary symmetric Fourier partial sums converging almost everywhere along the full sequence. The result concerns the actual partial sums, without averaging them or selecting a favorable subsequence.

2. Background and essential definitions

Fourier coefficients and symmetric sums

Represent the circle by $\mathbb T=\mathbb R/(2\pi\mathbb Z)$, with normalized length measure $d\mu(x)=dx/(2\pi)$. For an integrable complex-valued function $f$, define

$$\widehat f(k)=\int_{\mathbb T}f(x)e^{-ikx}\,d\mu(x), \qquad S_Nf(x)=\sum_{k=-N}^{N}\widehat f(k)e^{ikx}.$$

The integer $k$ is a frequency. The sum is symmetric because it includes all frequencies from $-N$ through $N$. For a real-valued function, positive and negative frequencies combine into the familiar sine and cosine terms.

Almost-everywhere convergence means that there is a set of measure zero outside which $S_Nf(x)\to f(x)$. The exceptional set may depend on $f$. This does not mean convergence at every point or uniform convergence over the circle.

How much larger than Lp is L log L

The class used in the theorem is

$$L\log L(\mathbb T)= \left\{f:\int_{\mathbb T}|f(x)|\log(2+|f(x)|)\,d\mu(x)<\infty\right\}.$$

The added two simply keeps the logarithm well behaved for small values. At large amplitudes the extra cost beyond $L^1$ is one logarithm. On this finite-measure space,

$$L^p\subset L\log L\subset L^1\qquad(p>1).$$

These inclusions allow strictly different classes of functions. For example, near zero a function behaving like

$$f(x)=\frac1{x(\log(1/x))^3},\qquad 0<x<e^{-2},$$

and set to zero elsewhere belongs to $L\log L$, but to no $L^p$ with $p>1$. Substituting $u=\log(1/x)$ turns its $L\log L$ integral near zero into an integral comparable to $\int u^{-2}\,du$, which is finite. Its power singularity makes every higher $L^p$ integral diverge. This illustrates the additional roughness admitted by the claimed theorem.

Why averaging changes the problem

Fejér sums average the initial Fourier partial sums. Their kernels are nonnegative, allowing convergence arguments based on local averaging. Ordinary Fourier sums instead use an oscillatory kernel with positive and negative parts. A theorem for Fejér means therefore does not settle the question asked here.

3. Development history and previous partial results

Kolmogorov constructed an integrable function whose Fourier series diverges almost everywhere in 1923. That example shows why a theorem for every $L^1$ function is impossible.

Carleson’s 1966 theorem established almost-everywhere convergence for $L^2$, resolving a landmark problem. Hunt extended it to every $L^p$, $p>1$, in 1968. The resulting Carleson–Hunt theorem leaves open how close one can come to ordinary integrability while retaining convergence for an entire function class.

Progress near $L^1$ was measured by slowly growing logarithmic corrections. Sjölin obtained a class with an additional iterated logarithm, while Soria developed extrapolation methods giving larger convergence spaces. Antonov’s 1996 theorem reached the class customarily written $L\log L\log\log\log L$, with the iterated logarithms understood as regularized near zero.

Arias de Reyna introduced a larger quasi-Banach convergence space, denoted $QA$. Carro, Mastyło and Rodríguez-Piazza developed related spaces, and Lie provided a direct time-frequency approach to a maximal bound near the endpoint. Di Plinio and Fragkos later obtained weak-type estimates with sharp order as $p$ approaches one.

The source manuscript cites these original works and distinguishes them from the proposed $L\log L$ conclusion. Removing an extra iterated logarithm can look minor in notation, but functions can concentrate on increasingly rare, increasingly high spikes. Even a very slowly growing factor changes which such functions are admitted.

4. What the new manuscripts claim

The main theorem states that every complex-valued $f\in L\log L(\mathbb T)$ has a measurable exceptional set $N_f$ with $\mu(N_f)=0$ such that

$$\lim_{N\to\infty}S_Nf(x)=f(x) \qquad(x\notin N_f).$$

The limit uses every integer $N=0,1,2,\ldots$. No assumption of positivity, bounded variation or piecewise smoothness is imposed on $f$.

The paper’s central analytical estimate treats a nonnegative input $\rho$ of total mass one and height at most $e^d$, for large $d$. Let $M\rho$ be its spatial maximal average: the supremum of the averages of $\rho$ over circle arcs containing the given point. With $h=\log\log d$, the paper removes the set where $M\rho>e^{4h}$ and claims

$$\int_{\{M\rho\le e^{4h}\}} \sup_{N\ge0}|S_N\rho|\,d\mu\le Cd.$$

The excluded set has measure at most $Ce^{-4h}$. These estimates have different jobs: one bounds the Fourier oscillations on the retained set, while the other limits how much of the circle was discarded.

The manuscript also states a weak-$L^1$ maximal consequence in the Luxemburg norm of $L\log L$. This is a distributional bound on the maximal partial sum, not a strong $L^1$ estimate for its integral over the whole circle.

5. Why the advance matters

Importance: 5/5. If correct, the theorem settles the classical $L\log L$ sufficiency problem for Fourier convergence and extends one of harmonic analysis’s central convergence theorems to a substantially larger class of rough functions.

The hypothesis is easy to express directly in terms of the function’s amplitude. A reader need not know its Fourier coefficients in advance to recognize the class. The conclusion then controls the full sequence of reconstructions from those coefficients, despite their oscillatory nature.

The proposed improvement also sharpens the conceptual boundary between two facts: Fourier analysis reconstructs every sufficiently integrable function almost everywhere, yet it can fail dramatically for general integrable functions. The theorem supplies a distinguished class between these regimes without claiming a necessary condition for each individual function.

6. The main idea

Break a rough function into height layers

A general $L\log L$ function may be unbounded. Divide it into portions according to the size of its values. Each normalized portion is bounded, so the good-set estimate can be applied to it.

The challenge is to sum the resulting bounds. If the analytical cost grew too rapidly with the layer height, the $L\log L$ integral would not pay for it. If the exceptional sets were too large, infinitely many bad events could cover a positive portion of the circle.

The paper uses layers with doubly exponential heights. Its stated integral bounds sum using the logarithmic integrability hypothesis, while the measures of the discarded sets are also summable. Approximation by well-behaved functions then supplies the convergence conclusion.

Organize many frequencies without paying for each one

The hard estimate behind this summation is first formulated on a finite tree of intervals, where each interval splits into two children. Differences between averages on the children are Haar-type coefficients, a discrete model for oscillation at a specified scale.

Modulation multiplies an input by $e^{iux}$, shifting its frequency. To control all Fourier cutoffs, the argument must handle many possible modulations together. Paying a separate full cost for every frequency would make the bound useless.

The manuscript groups frequencies whose phases behave similarly on a common region. It uses noisy observations of an input point to identify such groups. Entropy records how much information those observations reveal; a compression argument says that a large collection of frequencies with controlled information cost contains a useful common phase structure.

One distribution samples the input point in proportion to its mass. Another samples the output location uniformly on the circle. These distributions are not interchangeable. The paper’s bookkeeping converts information bounds under the first into estimates for transforms under the second.

Return from the tree to Fourier sums

Averaging translated and rescaled dyadic kernels produces a multiple of the Hilbert-transform kernel, together with a controlled error. The periodic Hilbert transform is a singular integral closely related to projection onto positive and negative Fourier frequencies.

This passage connects the finite tree estimate to ordinary symmetric Fourier sums. It is essential that the control survives the averaging and limiting operations, because convergence for the discrete model alone would not answer the original Fourier question.

7. Scope, limitations, and open questions

The theorem asserts almost-everywhere convergence, allowing an exceptional null set that depends on the function. It does not supply a universal exceptional set, uniform convergence or a rate of convergence.

The result is a sufficiency theorem. It does not characterize all functions whose Fourier series converge almost everywhere, and many individual functions outside $L\log L$ can still have convergent Fourier series. Kolmogorov’s counterexample prevents replacing the hypothesis by unrestricted $L^1$ membership.

The paper concerns one-dimensional ordinary symmetric Fourier sums. Other summation orders, higher-dimensional frequency regions and non-Fourier orthogonal systems require their own analysis. Nor should the weak maximal consequence be rewritten as a strong integral bound over the entire circle.

8. Sources and further reading

The source manuscript’s references include Carleson’s On convergence and growth of partial sums of Fourier series (1966), Hunt’s On the convergence of Fourier series (1968), Antonov’s Convergence of Fourier series (1996), and Lie’s On the boundedness of the Carleson operator near L1 (2013). These works describe the historical progression from power integrability toward logarithmic endpoint classes.

Optional background on Fourier series, Carleson’s theorem, and Fejér’s theorem helps distinguish ordinary partial sums from averaged reconstruction.

Source manuscripts

• M1. OpenAI, Almost-everywhere Fourier convergence in L log L (September 23, 2026).

Sources consulted on 14 October 2026 at repository revision 1581d4c. Download the manuscript citations.