Convergence of Fourier Series

Take a function on the circle. For example, the sawtooth wave $f(x) = x$ on $(-\pi,\pi)$ , extended periodically. Form its Fourier series. Does the series actually converge back to $f$ at every point? If so, in what sense? This question sounds like a technicality. But it occupied analysts for over a century. It produced some of the deepest ideas in real analysis: the discovery that pointwise convergence can fail even for continuous functions, and the realization that local smoothness, not global continuity, controls convergence. The subtlety is that a Fourier series is not an arbitrary series of functions. Its partial sums are given by a single, explicit convolution — the Dirichlet kernel — against $f$ . This kernel is highly oscillatory. It is not an approximate identity in the naive sense. This is exactly why convergence questions are delicate. It is also why continuity alone does not suffice. Understanding the Dirichlet kernel's shape, its integral, and its concentration near a point is key. It unlocks every convergence theorem in this article. The payoff is a hierarchy of increasingly strong statements. The Riemann–Lebesgue Lemma shows Fourier coefficients vanish. The localization principle shows convergence at a point depends only on the function's behavior arbitrarily close to that point. The Dini test and its corollary, the Lipschitz/Dirichlet–Jordan-type criterion, give clean sufficient conditions for pointwise convergence. A uniform version, requiring only a bit more, upgrades this to convergence in the sup norm. This connects to the general theory of Uniform Limit of Continuous Functions. This article builds that hierarchy rigorously. It starts from the Dirichlet kernel's integral representation. It proves the Riemann–Lebesgue Lemma, and establishes localization. It culminates in the Dini and uniform convergence theorems. Full proofs and worked examples are included. An explicit counterexample shows why some regularity hypothesis is unavoidable.

Key takeaways

The $N$ -th partial sum of a Fourier series equals $f$ convolved with the Dirichlet kernel $D_N$ .
The Riemann–Lebesgue Lemma forces Fourier coefficients of integrable functions to $0$ , killing high-frequency "noise" terms in convergence proofs.
Localization: whether $S_N f(x_0) \to f(x_0)$ depends only on the values of $f$ in an arbitrarily small neighborhood of $x_0$ , not on $f$ globally.
The Dini test gives a clean, checkable pointwise convergence criterion; the Lipschitz condition is a convenient special case.
Continuity of $f$ alone is not enough for pointwise convergence — a genuine hypothesis on local oscillation is required.
Uniform convergence follows from a Lipschitz condition holding uniformly over the whole domain, combined with the Weierstrass M-Test philosophy of controlled tails.
Key idea: convergence of a Fourier series at a point is a local, quantitative question about how fast $f$ oscillates near that point, encoded through the concentration of the Dirichlet kernel.

Where this lives today

The Dirichlet kernel and its convergence theory underlie the Gibbs phenomenon in signal processing, where truncated Fourier series of discontinuous signals overshoot near jumps — directly explained by the non-uniform, oscillatory tail of $D_N$ analyzed above. In numerical analysis and the Fast Fourier Transform, understanding partial-sum behavior guides windowing techniques (Fejér, Hann, Hamming windows) designed specifically to replace the slowly-decaying Dirichlet kernel with faster-decaying, positive kernels for better convergence and less ringing. In PDE theory, pointwise and uniform Fourier convergence criteria justify term-by-term differentiation and solution representations for the heat and wave equations on intervals with periodic or Dirichlet boundary conditions. In approximation theory and statistics, Hölder/Lipschitz-type smoothness conditions of exactly the sort appearing in Dini's test quantify convergence rates of kernel density estimators and control error bounds in nonparametric regression. The broader theme — that local regularity controls Fourier convergence, refined enormously by Carleson's theorem (a.e. convergence for $L^2$ ) and Kolmogorov's example (an $L^1$ function with everywhere-divergent Fourier series) — remains a living research area connecting harmonic analysis, ergodic theory, and number theory.

Ask the tutor: How does the Fejér kernel (Cesàro means of Fourier series) fix the non-positivity problem of the Dirichlet kernel and give unconditional uniform convergence for all continuous functions?

f:\mathbb{R}\to\mathbb{C}

Intuition

The single most important fact about $D_N$ is that it is not a well-behaved approximate identity: it integrates to $2\pi$ (so $\frac1{2\pi}\int D_N = 1$ ), but it takes both large positive and large negative values, and it does not concentrate its mass near $t=0$ in the way, say, the Fejér kernel does. Near $t = 0$ , $D_N(t) \approx 2N+1$ , a tall spike of width $\sim 1/N$ — this is the "good" part responsible for localization. But away from $0$ , $D_N(t) = \sin((N+\tfrac12)t)/\sin(t/2)$ oscillates with an envelope decaying only like $1/|t|$ , not fast enough to be negligible in absolute value. This slow decay is the entire source of difficulty: it means $S_Nf(x)$ picks up a genuinely oscillatory, not just concentrated, contribution from all of $[-\pi,\pi]$ , and only cancellation (via Riemann–Lebesgue) tames the far-field part. The strategy in every proof below is the same three-step maneuver: (1) split the convolution integral into a piece near $t=0$ (where $D_N$ is large but the interval is small) and a piece bounded away from $0$ (where $D_N$ oscillates but $\sin(t/2)$ is bounded below); (2) rewrite the difference $S_Nf(x_0) - f(x_0)$ using $\frac{1}{2\pi}\int D_N = 1$ so that only the increment $f(x_0 - t) - f(x_0)$ matters; (3) show that increment, divided by $\sin(t/2)$ , is itself integrable — which is exactly where a Dini/Lipschitz-type hypothesis enters — so that Riemann–Lebesgue kills the whole thing as $N \to \infty$ .

:::svg[The Dirichlet kernel $D_5(t)$ on $[-\pi,\pi]$ : a tall central spike of height $2N{+}1=11$ and slowly-decaying oscillatory side-lobes]

f

f = \mathbf{1}_{[c,d]}

f,g

h = f - g

f

D_N

f

s = f(x_0)

Worked examples

Example 1 — the sawtooth wave, convergence at a discontinuity. Let $f(x) = x$ for $x\in(-\pi,\pi)$ , extended $2\pi$ -periodically (so $f(\pi)$ is redefined by periodicity, creating a jump at $x=\pi \equiv -\pi$ ). Its Fourier coefficients: $\hat f(0) = 0$ by oddness, and for $n \neq 0$ , integrating by parts,

\hat f(n) = \frac{1}{2\pi}\int_{-\pi}^\pi t\, e^{-int}\,dt = \frac{(-1)^{n+1}}{in}.

So the real Fourier series is $f(x) \sim \sum_{n=1}^\infty \frac{2(-1)^{n+1}}{n}\sin(nx)$ . At $x_0 = \pi/3$ (an ordinary interior point where $f$ is smooth, in fact $f(x)=x$ locally): we check the Lipschitz corollary with $\alpha=1$ , $M=1$ : for small $t$ , $f(x_0+t) = x_0+t$ , so $|f(x_0+t)-f(x_0)| = |t| = M|t|^1$ , and similarly on the left. Hence Dini's/Lipschitz criterion applies directly and

S_N f(\pi/3) \to f(\pi/3) = \pi/3.

Numerically: $\sum_{n=1}^\infty \frac{2(-1)^{n+1}}{n}\sin(n\pi/3)$ converges (conditionally) to $\pi/3 \approx 1.0472$ , consistent with the theorem — this is not something we would be entitled to assert without a convergence theorem, since the series is only conditionally convergent. At $x_0 = \pi$ (the point of periodic discontinuity, where $f(\pi^-) = \pi$ and $f(\pi^+) = f(-\pi^+) = -\pi$ ): apply the second part of the corollary. For $t>0$ small, $f(\pi - t) = \pi - t \to \pi = f(\pi^-)$ , and $f(\pi+t) = f(-\pi+t) = -\pi+t \to -\pi = f(\pi^+)$ ; both satisfy the Lipschitz bound with $\alpha=1,M=1$ . Hence

S_Nf(\pi) \to \frac{f(\pi^-)+f(\pi^+)}{2} = \frac{\pi + (-\pi)}{2} = 0.

Indeed, every term $\sin(n\pi) = 0$ , so $S_N f(\pi) = 0$ for every $N$ — an exact, trivial confirmation, and a good sanity check that the "average of one-sided limits" formula is not just asymptotic but matches term-by-term here.

Example 2 — a function satisfying Dini but not a Lipschitz condition. Let $x_0 = 0$ and define, near $t=0$ ,

f(t) = \frac{1}{\ln(1/|t|)} \quad (0<|t|<1/2), \qquad f(0)=0,

extended arbitrarily (say, smoothly and periodically) outside $|t|<1/2$ to a $2\pi$ -periodic Riemann integrable function, and set $f(-t)=f(t)$ (even). This $f$ is continuous at $0$ (since $f(t)\to 0 = f(0)$ as $t\to0$ ), but for no $\alpha>0$ does $|f(t)-f(0)| \le M|t|^\alpha$ near $0$ : since $\frac{1}{\ln(1/t)} / t^\alpha = \frac{1}{t^\alpha \ln(1/t)} \to \infty$ as $t\to 0^+$ for every $\alpha>0$ (exponential beats logarithm), no Hölder/Lipschitz bound of any positive order holds. So the Corollary does not apply. However, Dini's test directly applies. With $s = f(0) = 0$ ,

\left|\frac{f(t)+f(-t) - 0}{t}\right| = \frac{2}{t\,\ln(1/t)} \qquad (0<t<1/2).

Is this integrable near $0$ ? Substituting $u = \ln(1/t)$ , $du = -dt/t$ :

\int_{0^+}^{c} \frac{2}{t\ln(1/t)}\,dt = 2\int_{\ln(1/c)}^{\infty} \frac{du}{u},

which diverges (harmonic-type divergence of $\int du/u$ )! So this particular $f$ , despite being continuous, does not satisfy the Dini condition at $0$ with this borderline decay rate — a useful cautionary computation showing how close to the boundary the Dini test sits. Let us instead take $f(t) = \dfrac{1}{(\ln(1/|t|))^{2}}$ near $0$ (still continuous, $f(0)=0$ , even, no Hölder exponent works by the same argument as above). Now

\left|\frac{f(t)+f(-t)}{t}\right| = \frac{2}{t(\ln(1/t))^2}, \qquad \int_{0^+}^c \frac{2\,dt}{t(\ln(1/t))^2} = 2\int_{\ln(1/c)}^\infty \frac{du}{u^2} < \infty,

which converges ( $\int_a^\infty u^{-2}\,du = 1/a<\infty$ ). So this $f$ satisfies the Dini test at $x_0=0$ with $s=0$ , even though it satisfies no Lipschitz/Hölder condition of any positive order there. This shows Dini's test is genuinely more general than the Lipschitz corollary, and the theorem's proof (via absolute integrability of the difference quotient, not a pointwise Hölder rate) is exactly the right level of generality.

f

f

|S_Nf(x)-f(x)|

Scope, edge cases & common traps

Situation	What happens	Correct statement
$f$ continuous at $x_0$ but no rate of continuity assumed	Convergence can fail (du Bois-Reymond)	Need Dini/Hölder-type hypothesis
$f$ has a jump discontinuity at $x_0$ , one-sided Hölder on each side	$S_Nf(x_0) \to \tfrac12(f(x_0^+)+f(x_0^-))$	Corollary to Dini's test
$f$ globally uniformly Hölder ( $C^{0,\alpha}$ )	$S_N f \to f$ uniformly	Uniform convergence theorem
$f \in L^2$ (only square-integrable)	Convergence in $L^2$ -norm always holds (Bessel/Parseval machinery), pointwise convergence is a much harder separate question (Carleson's theorem, beyond this scope)	Different mode of convergence
Localization	Convergence at $x_0$ depends only on an arbitrarily small neighborhood	Do not confuse with convergence rate, which can depend globally

Common trapthe classic error

A very common mistake is to argue: " $f$ is continuous, and Fourier coefficients $\hat f(n) \to 0$ by Riemann–Lebesgue, so the series converges to $f$ ." This confuses decay of coefficients with convergence of the partial sums; Riemann–Lebesgue only says the terms of auxiliary integrals vanish in the limit, not that the whole sum stabilizes. The quantifier order matters: Dini's test requires $\int_0^\delta |f(x_0+t)+f(x_0-t)-2s|/t\,dt <\infty$ — an integrability statement about the increment divided by $t$ , not merely that the increment itself tends to $0$ . Continuity gives you $f(x_0+t)-f(x_0)\to 0$ pointwise as $t\to 0$ , but says nothing about $1/t$ times that quantity being integrable, and indeed du Bois-Reymond's example shows the implication is genuinely false in general.

The mathematician behind it

Origins · 1 minPeter Gustav Lejeune Dirichlet and how this was discovered

Peter Gustav Lejeune Dirichlet

1805–1859 · Prussian mathematician, University of Berlin and Göttingen

Dirichlet gave the first rigorous convergence proof for Fourier series (1829), under a hypothesis close to what is now called the Dirichlet–Jordan condition (functions with finitely many maxima/minima and discontinuities on a period). He introduced the precise partial-sum kernel that now bears his name, and his 1829 memoir is widely regarded as the starting point of rigorous Fourier analysis as opposed to the more formal manipulations of Fourier and Euler. Dirichlet is also known for his pigeonhole principle, his work on the prime number theorem for arithmetic progressions, and for one of the first modern definitions of "function" as an arbitrary correspondence, a viewpoint essential to studying Fourier series of general (not merely analytic) functions.

History

- 1807/1822 — Fourier claims arbitrary periodic functions can be expanded in trigonometric series, sparking controversy over rigor.

- 1829 — Dirichlet proves the first rigorous pointwise convergence theorem, introducing the Dirichlet kernel.

- 1873 — du Bois-Reymond constructs a continuous function whose Fourier series diverges at a point, showing continuity alone is insufficient.

- 1880s — Dini and Jordan formulate sufficient conditions (Dini's test; bounded variation) refining and generalizing Dirichlet's criterion.

- 1966 — Carleson proves almost-everywhere convergence for all

L^2

functions, resolving a problem left open since Fourier's era (far beyond the classical pointwise theory here).

Why isn't continuity of

f

enough for its Fourier series to converge pointwise?

Continuity only guarantees that $f(x_0+t) - f(x_0) \to 0$ as $t\to 0$ , with no control on how fast. Dini's test requires the stronger statement that $[f(x_0+t)+f(x_0-t)-2s]/t$ is absolutely integrable near $t=0$ — an integrability condition, not just a limiting-to-zero condition. Du Bois-Reymond's 1873 example exhibits an explicit continuous function whose Fourier partial sums are unbounded at a point, proving the gap is real, not just a limitation of the proof technique. This is a recurring theme in analysis: qualitative smoothness (continuity) is often too weak to control quantitative, oscillatory sums; a rate (Hölder exponent, bounded variation, or an integrability condition on the difference quotient) is needed.

What exactly is the Dirichlet kernel, and why does its shape matter?

$D_N(t) = \sum_{n=-N}^N e^{int} = \sin((N+\tfrac12)t)/\sin(t/2)$ is the function such that convolving it with $f$ (with the $\tfrac{1}{2\pi}$ normalization) reproduces exactly the $N$ -th Fourier partial sum. Its shape — a tall central spike of height $2N+1$ near $t=0$ , surrounded by oscillating side-lobes whose envelope decays only like $1/|t|$ — is the reason Fourier convergence theory is subtle. Unlike "good" approximate identities (e.g. the Fejér kernel, which is nonnegative and decays fast away from $0$ ), $D_N$ takes both signs and its slow decay means far-away values of $f$ genuinely contribute to $S_Nf(x_0)$ ; only oscillatory cancellation, formalized by the Riemann–Lebesgue Lemma, tames this contribution.

How does Dini's test relate to the Lipschitz/Hölder criterion?

Dini's test is strictly more general. It asks that the increment quotient $[f(x_0+t)+f(x_0-t)-2s]/t$ be absolutely integrable near $t=0$ — an integral condition. A Hölder condition $|f(x_0+t)-f(x_0)|\le Mt^\alpha$ is one convenient sufficient condition guaranteeing this (since $t^{\alpha-1}$ is integrable near $0$ for any $\alpha>0$ ), but Dini's test also covers functions with no Hölder exponent at all, provided the decay of the increment is just barely fast enough in an integral-average sense (see Worked Example 2, with $f(t)\sim 1/(\ln 1/t)^2$ ).

Does localization mean the rate of convergence is also local?

No — this is a subtle and important distinction. Localization only says that whether $S_Nf(x_0)$ converges, and to what value, is determined by $f$ near $x_0$ . It says nothing about how fast the convergence is, which can genuinely depend on global properties of $f$ (e.g. the size of jumps elsewhere feeding into the Gibbs phenomenon's oscillatory tails, or the overall smoothness class of $f$ controlling uniform error bounds as in the uniform convergence theorem).

Why does the proof of the uniform convergence theorem need more than just applying Dini's test at every point?

Applying Dini's test pointwise only gives, for each $x$ , a value $N_0(x)$ beyond which $|S_Nf(x)-f(x)|<\varepsilon$ ; but $N_0(x)$ could a priori depend on $x$ and blow up, which would only give pointwise, not uniform, convergence. The uniform theorem needs the Hölder constant $M$ and exponent $\alpha$ to be the same for every $x$ , so that every estimate in the proof — the bound on $I_1$ , the equicontinuity modulus for $I_2$ , and the choice of $N_0$ — can be made independent of $x$ , yielding a single $N_0$ working for all $x$ at once. This mirrors the general distinction between pointwise and Uniform Limit of Continuous Functions-type convergence.

Is there a Fourier series that diverges at every point?

Yes — Kolmogorov (1926) constructed an $L^1$ function whose Fourier series diverges everywhere, showing $L^1$ integrability alone is far too weak a hypothesis. This lies beyond the classical Dini/Lipschitz theory in this article (which requires Riemann integrability plus a local regularity condition) and belongs to deeper 20th-century harmonic analysis, alongside Carleson's positive result that $L^2$ functions' Fourier series converge almost everywhere.

Why does Bessel's Inequality not already answer the pointwise convergence question?

Orthogonality and Bessel's Inequality shows $\sum |\hat f(n)|^2 \le \frac{1}{2\pi}\int|f|^2$ , which controls convergence of $S_Nf$ to $f$ in the $L^2$ (mean-square) norm, a fundamentally different and generally weaker notion than pointwise convergence at a specific $x_0$ . Two functions differing on a single point have identical Fourier coefficients and identical $L^2$ behavior, but pointwise convergence questions are precisely sensitive to such single-point information, which is why an entirely separate battery of tools (Dirichlet kernel, Dini's test) is needed.