Fejér sum (${\sigma }_n(f)$)

The $n$th Fejér sum is the arithmetic mean of the first $n$ Fourier partial sums. Equivalently, it is the convolution of the function with the Fejér kernel.

For a periodic Hölder continuous function of Hölder exponent $0<\alpha <1$ and Hölder seminorm at most $M$, the degree-$n-1$ Fejér sum has the displayed supremum norm error bound. In the half-kernel normalization $F_{n-1}=(2n{)}^{-1}{\sin }^2(nt/2)/{\sin }^2(t/2)$, use $F_{n-1}(t)\le \min \{n/2,{\pi }^2/(2n{t}^2)\}$. Splitting the convolution error integral at $1/n$ gives the valid constant $C_{\alpha }=1/[\pi (\alpha +1)]+\pi /(1-\alpha )$. At the endpoint $\alpha =1$ this argument instead gives $O(M\log n/n)$.

The normalized Fejér kernel satisfies $0\le F_n(t)\le \min \{n/(2\pi ),\pi /(2n{t}^2)\}$ on $0<|t|\le \pi$. Splitting its first absolute moment at $1/n$ gives $\int |t|F_n(t)dt\le 1/(2\pi n)+(\pi /n)\log (\pi n)$. The chaining inequality for the modulus of continuity, $\omega (f,t)\le (1+t/\delta )\omega (f,\delta )$, then bounds the Fejér sum error by $\omega (f,\delta )[1+{\delta }^{-1}\int |t|F_n]$. Taking $\delta =\log n/n$ proves the stated bound for $n\ge 2$ and hence uniform convergence for continuous periodic functions. The logarithm vanishes at $n=1$, so the displayed estimate is not intended at that index.

The positive normalized Fejér kernel has a $1/(n{t}^2)$ tail. Split the error integral at $\delta =n^{-\alpha }$ and use the chaining inequality $\omega (f,t)\le (1+t/\delta )\omega (f,\delta )$. Its tail contributes at most a constant times $n^{\alpha -1}(1+\alpha \log n)\omega (f,\delta )$. The factor is uniformly bounded for each fixed $\alpha <1$, proving the displayed estimate. Its constant need not remain bounded as $\alpha$ tends to one.

The even positive Fejér kernel of total mass one gives ${\sigma }_{n}f-f=\frac{1}{2}\int F_n(t)[f(\cdot +t)-2f+f(\cdot -t)]dt$. The scaling property of the second modulus of smoothness bounds this by ${\omega }_2(f,\delta )(1+{\delta }^{-2}\int t^2{F}_n(t)dt)$. On $[-\pi ,\pi ]$, $\sin (|t|/2)\ge |t|/\pi$ gives $\int t^2{F}_n(t)dt\le {\pi }^2/n$. Choose $\delta =n^{-1/2}$. For twice continuously differentiable functions, the second-difference integral formula makes the error $O(n^{-1})$.

For fixed positive integer $k$ and $n>k$, the Fourier coefficient of the Fejér kernel at frequency $k$ multiplies the mode by $1-k/n$. The displayed error is exact. Consequently the order $n^{-1}$ cannot be replaced by $o(n^{-1})$ for all twice continuously differentiable periodic functions, even though each fixed Fourier mode is infinitely differentiable.

The Fejér kernel is nonnegative and has unit mass in its convolution normalization. The integral triangle inequality therefore bounds the supremum norm of a Fejér sum by that of the original function. In the half-kernel convention $F_n=(2n{)}^{-1}|{\sum }_{j=0}^{n-1}{e}^{ijt}{|}^2$, its integral is $\pi$ and the convolution prefactor is $1/\pi$.