11/4/24, 3:44 PM Mathematical definitions of transforms — finufft 2.3.0 documentation 11/4/24, 3:44 PM Mathematical definitions of transforms — finufft 2.3.

Mathematical definitions of transforms — finufft 2.3.0 documentation

Finally, the type 3 (NU to NU) transform does not have restrictions on the NU points, and there is
Mathematical definitions of transforms no periodicity. Let x j ∈ R d , j = 1, … , M , be NU locations, with strengths c j ∈ C, and let s k ,
k = 1, … , N be NU frequencies. Then the type 3 transform evaluates:
We use notation with a general space dimensionality d, which will be 1, 2, or 3, in our library. The
arbitrary (ie nonuniform) points in space are denoted x j ∈ R d , j = 1, … , M . We will see that for M (3)
type 1 and type 2, without loss of generality one could restrict to the periodic box [−π, π) d . For f k := ∑ c j e ±isk ⋅xj for k = 1, … , N
type 1 and type 3, each such NU point carries a given associated strength c j ∈ C. Type 1 and j=1
type 2 involve the Fourier “modes” (Fourier series coefficients) with integer frequency indices lying
in the Cartesian product set For all three transforms, the computational effort scales like the product of the space-bandwidth
products (real-space width times frequency-space width) in each dimension. For type 1 and type 2
K = K N1 ,…,Nd := K N1 × K N2 × ⋯ × K Nd , this means near-linear scaling in the total number of modes N := N 1 … N d . However, be
warned that for type 3 this means that, even if N and M are small, if the product of the tightest in‐
where tervals enclosing the coordinates of x j and s k is large, the algorithm will be inefficient. For such
NU points, a direct sum should be used instead.
{−N i /2, … , N i /2 − 1}, N i even,
K Ni := { We emphasise that the NUFFT tasks that this library performs should not be confused with either
the discrete Fourier transform (DFT), the (continuous) Fourier transform (although it may be used
{−(N i − 1)/2, … , (N i − 1)/2}, N i odd.
to approximate this via a quadrature rule), or the inverse NUFFT (the iterative solution of the linear
For instance, K 10 = {−5, −4, … , 4}, whereas K 11 = {−5, −4, … , 5}, and so
K 10,11 = {(−5, −5), (−4, −5), … , (4, −5), (−5, −4), (−4, −4), … , (3, 5), (4, 5)}. Note system arising from nonuniform Fourier sampling, as in, eg, MRI). It is also important to know that,
for NU points, the type 1 is not the inverse of the type 2. See the references for clarification.
that the ordering in the last case is with the first index “fast”, second “slow”; this matches the stor‐
age ordering in the library interface. Thus, in the 1D case K is an interval containing N 1 integer in‐
dices, in 2D it is a list of N 1 N 2 index pairs (which may be thought of as a rectangle of frequen‐
cies), and in 3D it is a list of N 1 N 2 N 3 index triplets (which may be thought of as a cuboid).

Then the type 1 (nonuniform to uniform, aka “adjoint”) NUFFT evaluates

M (1)
f k := ∑ c j e ±ik⋅xj for k ∈ K
j=1

This can be viewed as evaluating a set of Fourier series coefficients due to sources with strengths
c j at the arbitrary locations x j . Either sign of the imaginary unit in the exponential can be chosen
in the interface. Note that our normalization differs from that of references [DR,GL].

The type 2 (U to NU, aka “forward”) NUFFT evaluates

c j := ∑ f k e ±ik⋅xj for j = 1, … , M (2)

This is the adjoint of the type 1, ie the evaluation of a given Fourier series at a set of arbitrary
points. Both type 1 and type 2 transforms are invariant under translations of the NU points by multi‐
ples of 2π, thus one could require that all NU points live in the origin-centered box [−π, π) d . In
fact, as a compromise between library speed, and flexibility for the user (for instance, to avoid
boundary points being flagged as outside of this box due to round-off error), our library only re‐
quires that the NU points lie in the three-times-bigger box x j ∈ [−3π, 3π] d . This allows the user
latest latest
to choose a convenient periodic domain that does not touch this three-times-bigger box. However,
there may be a slight speed increase if most points fall in [−π, π) d .

https://finufft.readthedocs.io/en/latest/math.html 1/2 https://finufft.readthedocs.io/en/latest/math.html 2/2