Convergence of Fourier Series Solutions

The three-tier convergence framework here — mean-square, pointwise, uniform — is the working toolkit anywhere eigenfunction expansions are used to solve linear PDEs: heat conduction, vibrating membranes, quantum mechanical wavefunctions expanded in energy eigenstates, and Sturm–Liouville problems generally, as previewed in Eigenfunction Expansions. The exponential-smoothing argument used for the heat kernel is the prototype for parabolic regularity theory, which shows that solutions of diffusion-type PDEs (including the Black–Scholes equation in mathematical finance, and diffusion models in population biology) are automatically smooth away from the initial time, regardless of how irregular the initial data is. In signal processing, the Dirichlet kernel's poor $L^1$ behavior (responsible for the Gibbs phenomenon) motivates the use of Fejér and other averaging kernels, which are nonnegative and give better-behaved (uniformly convergent) approximations for merely continuous signals — a design principle in digital filtering and in windowing methods for the discrete Fourier transform used throughout numerical analysis and cryptographic hash constructions relying on fast transforms. The question of convergence "how far can you weaken the hypotheses" continues into deep 20th-century analysis: Carleson's theorem (1966) shows $L^2$ data already gives almost-everywhere convergence, a landmark result in harmonic analysis with influence on ergodic theory. Finally, whether a formal series solution actually solves the PDE — the differentiability question resolved here for the heat equation — is the first test case for the broader theory of Well-Posedness of PDE Problems, where existence, uniqueness, and continuous dependence on data are established rigorously.

Ask the tutor: Why doesn't Gibbs phenomenon contradict Dirichlet's pointwise convergence theorem, and how does the Gibbs overshoot behave as we add more terms?

Think of a Fourier series as an infinite orchestra: each term $a_n\cos(n\pi x/L)$ or $b_n\sin(n\pi x/L)$ is an instrument playing at frequency $n$, and the coefficients tell you the volume assigned to that instrument. Mean-square convergence just says that the total energy left in the unplayed instruments (the tail of the series) shrinks to zero — a statement about aggregate loudness, not about what happens at any single instant $x$. This is why it needs almost nothing: Eigenfunction Expansions of any energy-finite signal automatically have vanishing tail energy, by completeness of the trigonometric system. Pointwise convergence is a much more local and delicate question: does the orchestra's output converge to a definite value at this exact moment $x_0$? Gibbs-type oscillations near a jump show that partial sums can overshoot substantially near a discontinuity even while converging in energy nearby. The classical fix is the Dirichlet–Jordan criterion: if $f$ is piecewise smooth (piecewise $C^1$, with one-sided derivatives existing at jumps), the partial sums converge at every point to the midpoint of the jump — a clean, computable rule. Uniform convergence is the strongest and rarest of the three: it says the maximum error over the whole interval goes to zero, so the partial sums approximate $f$ uniformly well everywhere at once, and consequently the limit must be continuous. A jump discontinuity is instantly fatal to uniform convergence (partial sums of continuous functions cannot converge uniformly to a discontinuous limit), which is why continuity of $f$ (with matching periodic boundary values) is essential.

Example 1 — Pointwise convergence at a jump, sawtooth-type data. Let $L=\pi$ and define on $(-\pi,\pi]$ the function

extended $2\pi$-periodically. This is piecewise smooth: on $(-\pi,0)$ and $(0,\pi)$ it is linear (hence $C^1$ with derivative $1$), and it has a jump discontinuity at $x=0$ with $f(0^-)=1$, $f(0^+)=-1$, and also (by periodicity) a jump at $x=\pm\pi$ since $f(\pi^-)=\pi-1$ but $f(-\pi^+) = -\pi+1$. Since $f(x) = x + h(x)$ where $h$ is the odd square-wave-like jump function ($h=1$ on $(-\pi,0)$, $h=-1$ on $(0,\pi)$), and $x$ is odd while $h$ is odd, $f$ is odd, so $a_n=0$ for all $n\ge0$. Compute

(using oddness of $f\sin(nx)$... more carefully: $f$ is odd and $\sin(nx)$ is odd, so $f(x)\sin(nx)$ is even, giving the factor $2\int_0^\pi$). Compute directly:

So

For odd $n$: $(-1)^{n+1}=1$, $(-1)^n-1=-2$, so $b_n = \dfrac{2}{n} - \dfrac{4}{n\pi}$. For even $n$: $(-1)^{n+1}=-1$, $(-1)^n-1=0$, so $b_n = -\dfrac{2}{n}$. By the Dirichlet–Jordan theorem, at $x=0$ the series converges to $\frac{f(0^-)+f(0^+)}{2} = \frac{1+(-1)}{2}=0$ — matching the fact that every term $\sin(n\cdot0)=0$, so the series trivially sums to $0$ there, consistent with the theorem. At a point of continuity such as $x=\pi/2$ (where $f(\pi/2)=\pi/2-1$, and $f$ is smooth in a neighborhood since $\pi/2\ne0,\pi$), the theorem guarantees $S_N(\pi/2)\to f(\pi/2)=\pi/2-1\approx0.5708$, and since $f$ is not globally continuous (jumps at $0,\pm\pi$), the theorem's uniform-convergence clause does not apply — indeed one expects Gibbs overshoot near $x=0$.

Example 2 — Smoothing by the heat kernel: rough data becomes analytic instantly. Consider the heat equation on a rod of length $L=1$ with insulated... take the classical Dirichlet problem

with initial data

a discontinuous step function — about as rough as an $L^2$ function can be while remaining bounded. By separation of variables the formal solution is

Note $|b_n|\le \frac{4}{n\pi}$, so the coefficients decay only like $1/n$ — nowhere near enough for uniform convergence of $\sum b_n \sin(n\pi x)$ itself (consistent with $f$ being discontinuous, so Theorem's uniform clause cannot hold for the initial data series). Now fix any $t_0>0$ and ask about the solution series. Differentiating $k$ times in $x$ introduces a factor of at most $(n\pi)^k$, and differentiating $j$ times in $t$ introduces a factor of at most $(n^2\pi^2\alpha^2)^j$; the general term of any such differentiated series is bounded by

Since exponential decay in $n^2$ dominates any polynomial growth $n^{k+2j}$, i.e.

the series $\sum_n |b_n|\,n^{k+2j} e^{-n^2\pi^2\alpha^2 t_0}$ converges absolutely for every $k,j\ge 0$ and every $t_0>0$ (compare to $\sum n^{k+2j}e^{-cn^2}$, which converges by the ratio test since consecutive ratios $\to0$). By the Weierstrass $M$-test, every differentiated series converges uniformly on $[0,1]\times[t_0,\infty)$ for each $t_0>0$, justifying term-by-term differentiation to any order. Hence $u(x,t)$ is $C^\infty$ (indeed real-analytic) in $(x,t)$ for $t>0$, despite $f$ being discontinuous. This is the smoothing phenomenon: the heat kernel's exponential factors erase all memory of the roughness of the initial data for any positive time. Quantitatively, at $t_0=0.05$, $\alpha=1$: the ratio of successive polynomial-times-exponential terms is $\left(\frac{n+1}{n}\right)^{k+2j} e^{-(2n+1)\pi^2 t_0} \to 0$ extremely fast — e.g. for $n=10$, $e^{-(21)\pi^2(0.05)} = e^{-10.36}\approx 3.16\times10^{-5}$, already tiny, confirming numerically the qualitative claim.