1: \begin{abstract}

2:   The most popular algorithm for the

3:    \NU\ fast Fourier transform 

4:    (NUFFT) uses the dilation of a kernel $\phi$ to spread (or interpolate)

5:    between given \NU\ points and a uniform upsampled grid,

6:   combined with an FFT and diagonal scaling

7:   (deconvolution) in frequency space.

8:   %

9:   The high performance of the recent FINUFFT library

10:   is in part due to its use of

11:   a new ``exponential of semicircle'' kernel

12:   $\phi(z)=e^{\beta \sqrt{1-z^2}}$, for $z\in[-1,1]$, zero otherwise, whose

13:   Fourier transform $\hat\phi$ is unknown analytically.

14:   %

15:   We place this kernel on a rigorous footing

16:   by proving an aliasing error estimate

17:   which 

18:   bounds the error of the

19:   one-dimensional NUFFT of types 1 and 2

20:   in exact arithmetic.

21:   %

22:   %

23:   Asymptotically in the kernel width measured in upsampled grid points,

24:   the error is shown to decrease with an exponential rate

25:   arbitrarily close to

26:   that of the popular Kaiser--Bessel kernel.

27:   %

28:   This requires controlling a

29:   conditionally-convergent

30:   sum over the tails of $\hat\phi$,

31:   using steepest descent, other classical estimates on contour integrals,

32:   and a phased sinc sum.

33:   %

34:   We also draw new connections between the above kernel,

35:   Kaiser--Bessel, and prolate spheroidal wavefunctions of order zero,

36:   which all appear to share an optimal exponential convergence rate.

37: \end{abstract}

38: