FFTW 3.3.5: Real even/odd DFTs (cosine/sine transforms)

The Fourier transform of a real-even function f(-x) = f(x) is cannam@127: real-even, and i times the Fourier transform of a real-odd cannam@127: function f(-x) = -f(x) is real-odd. Similar results hold for a cannam@127: discrete Fourier transform, and thus for these symmetries the need for cannam@127: complex inputs/outputs is entirely eliminated. Moreover, one gains a cannam@127: factor of two in speed/space from the fact that the data are real, and cannam@127: an additional factor of two from the even/odd symmetry: only the cannam@127: non-redundant (first) half of the array need be stored. The result is cannam@127: the real-even DFT (REDFT) and the real-odd DFT (RODFT), also cannam@127: known as the discrete cosine and sine transforms (DCT and cannam@127: DST), respectively. cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127: cannam@127:

(In this section, we describe the 1d transforms; multi-dimensional cannam@127: transforms are just a separable product of these transforms operating cannam@127: along each dimension.) cannam@127:

Because of the discrete sampling, one has an additional choice: is the cannam@127: data even/odd around a sampling point, or around the point halfway cannam@127: between two samples? The latter corresponds to shifting the cannam@127: samples by half an interval, and gives rise to several transform cannam@127: variants denoted by REDFTab and RODFTab: a and cannam@127: b are 0 or 1, and indicate whether the input cannam@127: (a) and/or output (b) are shifted by half a sample cannam@127: (1 means it is shifted). These are also known as types I-IV of cannam@127: the DCT and DST, and all four types are supported by FFTW’s r2r cannam@127: interface.³ cannam@127:

The r2r kinds for the various REDFT and RODFT types supported by FFTW, cannam@127: along with the boundary conditions at both ends of the input cannam@127: array (n real numbers in[j=0..n-1]), are: cannam@127:

Note that these symmetries apply to the “logical” array being cannam@127: transformed; there are no constraints on your physical input cannam@127: data. So, for example, if you specify a size-5 REDFT00 (DCT-I) of the cannam@127: data abcde, it corresponds to the DFT of the logical even array cannam@127: abcdedcb of size 8. A size-4 REDFT10 (DCT-II) of the data cannam@127: abcd corresponds to the size-8 logical DFT of the even array cannam@127: abcddcba, shifted by half a sample. cannam@127:

All of these transforms are invertible. The inverse of R*DFT00 is cannam@127: R*DFT00; of R*DFT10 is R*DFT01 and vice versa (these are often called cannam@127: simply “the” DCT and IDCT, respectively); and of R*DFT11 is R*DFT11. cannam@127: However, the transforms computed by FFTW are unnormalized, exactly cannam@127: like the corresponding real and complex DFTs, so computing a transform cannam@127: followed by its inverse yields the original array scaled by N, cannam@127: where N is the logical DFT size. For REDFT00, cannam@127: N=2(n-1); for RODFT00, N=2(n+1); otherwise, N=2n. cannam@127: cannam@127: cannam@127:

Note that the boundary conditions of the transform output array are cannam@127: given by the input boundary conditions of the inverse transform. cannam@127: Thus, the above transforms are all inequivalent in terms of cannam@127: input/output boundary conditions, even neglecting the 0.5 shift cannam@127: difference. cannam@127:

FFTW is most efficient when N is a product of small factors; note cannam@127: that this differs from the factorization of the physical size cannam@127: n for REDFT00 and RODFT00! There is another oddity: n=1 cannam@127: REDFT00 transforms correspond to N=0, and so are not cannam@127: defined (the planner will return NULL). Otherwise, any positive cannam@127: n is supported. cannam@127:

For the precise mathematical definitions of these transforms as used by cannam@127: FFTW, see What FFTW Really Computes. (For people accustomed to cannam@127: the DCT/DST, FFTW’s definitions have a coefficient of 2 in front cannam@127: of the cos/sin functions so that they correspond precisely to an cannam@127: even/odd DFT of size N. Some authors also include additional cannam@127: multiplicative factors of cannam@127: √2for selected inputs and outputs; this makes cannam@127: the transform orthogonal, but sacrifices the direct equivalence to a cannam@127: symmetric DFT.) cannam@127:

Which type do you need?

Since the required flavor of even/odd DFT depends upon your problem, cannam@127: you are the best judge of this choice, but we can make a few comments cannam@127: on relative efficiency to help you in your selection. In particular, cannam@127: R*DFT01 and R*DFT10 tend to be slightly faster than R*DFT11 cannam@127: (especially for odd sizes), while the R*DFT00 transforms are sometimes cannam@127: significantly slower (especially for even sizes).⁴ cannam@127:

Thus, if only the boundary conditions on the transform inputs are cannam@127: specified, we generally recommend R*DFT10 over R*DFT00 and R*DFT01 over cannam@127: R*DFT11 (unless the half-sample shift or the self-inverse property is cannam@127: significant for your problem). cannam@127:

If performance is important to you and you are using only small sizes cannam@127: (say n<200), e.g. for multi-dimensional transforms, then you cannam@127: might consider generating hard-coded transforms of those sizes and types cannam@127: that you are interested in (see Generating your own code). cannam@127:

We are interested in hearing what types of symmetric transforms you find cannam@127: most useful. cannam@127:

cannam@127:

Footnotes

cannam@127: cannam@127:

(3)

cannam@127:

There are also type V-VIII transforms, which cannam@127: correspond to a logical DFT of odd size N, independent of cannam@127: whether the physical size n is odd, but we do not support these cannam@127: variants.

cannam@127:

(4)

cannam@127:

R*DFT00 is cannam@127: sometimes slower in FFTW because we discovered that the standard cannam@127: algorithm for computing this by a pre/post-processed real DFT—the cannam@127: algorithm used in FFTPACK, Numerical Recipes, and other sources for cannam@127: decades now—has serious numerical problems: it already loses several cannam@127: decimal places of accuracy for 16k sizes. There seem to be only two cannam@127: alternatives in the literature that do not suffer similarly: a cannam@127: recursive decomposition into smaller DCTs, which would require a large cannam@127: set of codelets for efficiency and generality, or sacrificing a factor of cannam@127: 2 cannam@127: in speed to use a real DFT of twice the size. We currently cannam@127: employ the latter technique for general n, as well as a limited cannam@127: form of the former method: a split-radix decomposition when n cannam@127: is odd (N a multiple of 4). For N containing many cannam@127: factors of 2, the split-radix method seems to recover most of the cannam@127: speed of the standard algorithm without the accuracy tradeoff.

cannam@127:

2.5.2 Real even/odd DFTs (cosine/sine transforms)

Which type do you need?

Footnotes

(3)

(4)