Package {zufft}

Title: Fast Fourier, Cosine and Sine Transforms
Version: 0.1.0
Description: Computes discrete Fourier transforms (complex, real-to-complex and complex-to-real) and discrete cosine and sine transforms of types one to four over any axes of an array, with the fast Fourier transform (FFT) code of 'ducc0' (https://github.com/mreineck/ducc) bundled in the BSD-licensed subset that 'SciPy' uses. Transforms of any length run in O(n log n) time, real input uses true real transforms, and several threads can share one transform.
License: MIT + file LICENSE
URL: https://github.com/pedrobtz/zufft, https://pedrobtz.github.io/zufft/
BugReports: https://github.com/pedrobtz/zufft/issues
Depends: R (≥ 4.1)
Suggests: fftwtools, knitr, parallel, rmarkdown, testthat (≥ 3.0.0), withr
VignetteBuilder: knitr
Config/roxygen2/version: 8.1.0
Config/testthat/edition: 3
Copyright: See inst/COPYRIGHTS.
Encoding: UTF-8
Language: en-GB
SystemRequirements: C++17
NeedsCompilation: yes
Packaged: 2026-10-13 21:42:17 UTC; pbtz
Author: Pedro Baltazar [aut, cre, cph], Martin Reinecke [ctb] (ducc0, bundled in src/ducc0), Peter Bell [ctb, cph] (ducc0, bundled in src/ducc0), Max-Planck-Society [cph] (ducc0, bundled in src/ducc0), Matteo Frigo [cph] (odd-sized DCT-IV in src/ducc0/fft/fft1d_impl.h), Massachusetts Institute of Technology [cph] (odd-sized DCT-IV in src/ducc0/fft/fft1d_impl.h), The TensorFlow Authors [cph] (helpers in src/ducc0/infra/threading.cc), SciPy Developers [cph] (fast-length search in src/zff_lengths.cpp, from SciPy)
Maintainer: Pedro Baltazar <pedrobtz@gmail.com>
Repository: CRAN
Date/Publication: 2026-10-22 11:10:02 UTC

zufft: Fast Fourier, Cosine and Sine Transforms

Description

Computes discrete Fourier transforms (complex, real-to-complex and complex-to-real) and discrete cosine and sine transforms of types one to four over any axes of an array, with the fast Fourier transform (FFT) code of 'ducc0' (https://github.com/mreineck/ducc) bundled in the BSD-licensed subset that 'SciPy' uses. Transforms of any length run in O(n log n) time, real input uses true real transforms, and several threads can share one transform.

Author(s)

Maintainer: Pedro Baltazar pedrobtz@gmail.com [copyright holder]

Authors:

Other contributors:

See Also

Useful links:

Discrete Fourier transforms

Description

fft_c2c() computes the complex discrete Fourier transform of x or its inverse, fft_r2c() the transform of real input keeping only the non-redundant half of the last transformed axis, and fft_c2r() the inverse of fft_r2c(), returning real values. All three work over any subset of the dimensions of an array, in O(n log n) time for every length.

Usage

fft_c2c(
  x,
  inverse = FALSE,
  n = NULL,
  axes = NULL,
  norm = c("backward", "ortho", "forward", "none"),
  workers = NULL
)

fft_r2c(
  x,
  n = NULL,
  axes = NULL,
  norm = c("backward", "ortho", "forward", "none"),
  workers = NULL
)

fft_c2r(
  x,
  n = NULL,
  axes = NULL,
  norm = c("backward", "ortho", "forward", "none"),
  workers = NULL
)

Arguments

x

A double, integer, logical or complex vector or array. Integer and logical values are converted to double. fft_r2c() needs real input; fft_c2r() treats real input as complex.

inverse

TRUE for the inverse transform. 0 and 1 are accepted too, as in fftwtools.

n

Transform lengths, one per transformed axis in the order of axes (no recycling); with axes = NULL, one per dimension, in dimension order. Each axis of x is cropped or zero-padded to its length first. NULL keeps the lengths of x. For fft_c2r(), n gives the lengths of the real output; the halved axis of x must then hold n %/% 2 + 1 values and is cropped or padded to that. Its default is 2 * (m - 1) for an input extent m, which is always even: give n to recover an odd length.

axes

The dimensions to transform, 1-based, without repeats. NULL means all of them, in order; for fft_r2c() and fft_c2r() it means all of them in reverse order, so the first dimension is halved. With explicit axes, the real transforms halve the last axis listed, as SciPy does: axes = seq_along(dim(x)) halves the last dimension. Use axes = 1 for a column-wise transform of a matrix, as stats::mvfft() does.

norm

Where the scaling goes. "backward" (the default) leaves the forward transform unscaled and divides the inverse by N, the product of the transformed lengths; "forward" does the opposite; "ortho" divides both by sqrt(N); "none" scales neither direction, as FFTW, fftwtools and stats::fft() do.

workers

Threads to use. NULL uses the session default (1). Negative values count back from the number of available cores: -1 means all of them. Requests are capped at the available cores, and at 2 when ⁠_R_CHECK_LIMIT_CORES_⁠ is set. Threads split the independent 1-D transforms of a multi-dimensional array, so they help only arrays with at least 32 768 elements and more than one line. Results agree to rounding whatever the number of threads; they are bit-identical on builds without fused multiply-add (CRAN's x86-64 builds).

Details

The conventions are those of scipy.fft (fft, fftn, rfftn, irfftn): n crops or zero-pads each transformed axis before the transform, and norm chooses where the scaling goes. They differ from SciPy in three ways: axes are 1-based; axes = NULL means every dimension rather than the last one; and with axes = NULL the real transforms halve the first dimension, not the last. R stores arrays column by column, so the first dimension is the contiguous one, and halving it is the fast choice (about 1.8 times faster than halving the last for a 2048 x 2048 matrix). It is also what fftwtools' 2-D and 3-D real transforms do.

Coming from fftwtools: fft_c2c(x) is fftw(x) and fftw_r2c(x), fft_r2c(x) is fftw_r2c(x, HermConj = 0), and an fftwtools inverse is the zufft inverse with norm = "none", because fftwtools, like stats::fft(), never scales.

Value

For fft_c2c() and fft_r2c() a complex vector or array, for fft_c2r() a double one. Arrays keep their dim, with the transformed extents, and the dimnames of axes that were not transformed; names are dropped.

See Also

stats::fft() for R's own complex transform.

Other transforms: fft_dct()

Examples

x <- c(1, 2, 3, 4, 3, 2)
fft_c2c(x)
fft_r2c(x)
fft_c2r(fft_r2c(x), n = length(x))

# an inverse that matches fftwtools and stats::fft(inverse = TRUE)
fft_c2c(fft_c2c(x), inverse = TRUE, norm = "none") / length(x)

# columns of a matrix, and both axes
m <- matrix(rnorm(12), 4, 3)
fft_c2c(m, axes = 1)
dim(fft_r2c(m))                 # halves the first dimension: 3 x 3
dim(fft_r2c(m, axes = 1:2))     # SciPy's choice, the last: 4 x 2
fft_c2r(fft_r2c(m), n = dim(m)) # back to m

Discrete cosine and sine transforms

Description

fft_dct() and fft_dst() compute discrete cosine and sine transforms of types 1 to 4, or their inverses, over any subset of the dimensions of an array. The definitions, scaling and orthogonalize behaviour are those of scipy.fft.dct(), dctn(), dst() and dstn().

Usage

fft_dct(
  x,
  type = 2L,
  inverse = FALSE,
  n = NULL,
  axes = NULL,
  norm = c("backward", "ortho", "forward", "none"),
  orthogonalize = NULL,
  workers = NULL
)

fft_dst(
  x,
  type = 2L,
  inverse = FALSE,
  n = NULL,
  axes = NULL,
  norm = c("backward", "ortho", "forward", "none"),
  orthogonalize = NULL,
  workers = NULL
)

Arguments

x

A double, integer, logical or complex vector or array. Complex input is transformed as its real and imaginary parts separately.

type

The transform type: 1, 2, 3 or 4.

inverse

TRUE for the inverse transform.

n

Transform lengths, one per transformed axis in the order of axes (no recycling); with axes = NULL, one per dimension, in dimension order. Each axis of x is cropped or zero-padded to its length first. NULL keeps the lengths of x. For fft_c2r(), n gives the lengths of the real output; the halved axis of x must then hold n %/% 2 + 1 values and is cropped or padded to that. Its default is 2 * (m - 1) for an input extent m, which is always even: give n to recover an odd length.

axes

The dimensions to transform, 1-based, without repeats. NULL means all of them, in order; for fft_r2c() and fft_c2r() it means all of them in reverse order, so the first dimension is halved. With explicit axes, the real transforms halve the last axis listed, as SciPy does: axes = seq_along(dim(x)) halves the last dimension. Use axes = 1 for a column-wise transform of a matrix, as stats::mvfft() does.

norm

Where the scaling goes, as in fft_c2c(). "ortho" makes the transform orthonormal when orthogonalize is TRUE.

orthogonalize

Whether to scale the first and last terms so that the "ortho" transform matrix is orthogonal. NULL (the default) means norm == "ortho", as in SciPy. For complex input an explicit value is ignored and norm == "ortho" is used, as SciPy does.

workers

Threads to use. NULL uses the session default (1). Negative values count back from the number of available cores: -1 means all of them. Requests are capped at the available cores, and at 2 when ⁠_R_CHECK_LIMIT_CORES_⁠ is set. Threads split the independent 1-D transforms of a multi-dimensional array, so they help only arrays with at least 32 768 elements and more than one line. Results agree to rounding whatever the number of threads; they are bit-identical on builds without fused multiply-add (CRAN's x86-64 builds).

Details

Unnormalised (norm = "backward" or "none" in the forward direction), for a length N, fft_dct() computes

and fft_dst() the matching sine sums (type 1 over N + 1, types 2 to 4 as above with sines and indices shifted by one). Type 1 needs lengths of at least 2 for the cosine transform.

inverse = TRUE computes the inverse: types 2 and 3 swap, and the scaling moves to the other direction. With norm = "backward" the inverse is scaled by 1/N, where N is 2(n - 1) for DCT-I, 2(n + 1) for DST-I and ⁠2n⁠ otherwise, per transformed axis, so that fft_dct(fft_dct(x), inverse = TRUE) returns x.

Value

A double vector or array (complex for complex input) of the same shape as x after cropping or padding to n.

See Also

Other transforms: fft_c2c()

Examples

x <- c(1, 2, 3, 4, 3, 2)
fft_dct(x)
fft_dct(fft_dct(x), inverse = TRUE)
fft_dst(x, type = 4, norm = "ortho")

# an orthonormal DCT-II along the columns of a matrix
m <- matrix(rnorm(12), 4, 3)
fft_dct(m, axes = 1, norm = "ortho")

Sample frequencies

Description

fft_freq() returns the frequency of each entry of a length-n transform from fft_c2c(), in cycles per unit of the sample spacing d: ⁠0, 1, ..., ceiling(n / 2) - 1, -floor(n / 2), ..., -1⁠, divided by d * n. fft_rfreq() returns the frequencies of the n %/% 2 + 1 entries of fft_r2c(): ⁠0, 1, ..., n %/% 2⁠, divided by d * n. They are SciPy's fftfreq() and rfftfreq().

Usage

fft_freq(n, d = 1)

fft_rfreq(n, d = 1)

Arguments

n

The transform length, a positive whole number.

d

The sample spacing, a positive number.

Value

A double vector.

See Also

Other helpers: fft_next_len(), fft_shift()

Examples

fft_freq(8)
fft_freq(7, d = 0.1)
fft_rfreq(8)

Fast transform lengths

Description

fft_next_len() returns the smallest length at least n that zufft transforms fastest, and fft_prev_len() the largest at most n. Padding or cropping an input to such a length with the n argument of the transforms can speed them up considerably. They are SciPy's next_fast_len() and prev_fast_len().

Usage

fft_next_len(n, real = FALSE)

fft_prev_len(n, real = FALSE)

Arguments

n

Positive whole numbers, at most 2^52.

real

TRUE for the lengths of real transforms (fft_r2c(), fft_c2r(), fft_dct(), fft_dst()).

Details

Fast lengths have no prime factor above 11 for complex transforms, or above 5 for real ones (real = TRUE), as in ducc0's good_size.

Value

A double vector of the same length as n.

See Also

stats::nextn(), R's own version for stats::fft().

Other helpers: fft_freq(), fft_shift()

Examples

fft_next_len(c(1009, 999983))
fft_next_len(1009, real = TRUE)
fft_prev_len(1009)

Shift the zero frequency to the centre

Description

fft_shift() rotates each chosen dimension of x by floor(n / 2), so that the zero-frequency entry of a transform sits in the middle; fft_ishift() undoes it. They are SciPy's fftshift() and ifftshift(). dim, dimnames and names move with the values.

Usage

fft_shift(x, axes = NULL)

fft_ishift(x, axes = NULL)

Arguments

x

A vector or array.

axes

The dimensions to shift, 1-based. NULL means all of them.

Value

x, rotated.

See Also

Other helpers: fft_freq(), fft_next_len()

Examples

fft_shift(fft_freq(8))
fft_ishift(fft_shift(1:7))

Drop-in replacements for stats::fft() and stats::mvfft()

Description

fft_stats() and mvfft_stats() take the same arguments, check them the same way and return the same structure as stats::fft() and stats::mvfft(): the attributes of z are kept, the inverse is not scaled, and inputs of length 0 or 1 come back unchanged (as complex). The values agree to rounding. The differences: every length runs in O(n log n) (R's code is quadratic in a large prime factor), real input uses a real transform, and threads follow fft_workers().

Usage

fft_stats(z, inverse = FALSE)

mvfft_stats(z, inverse = FALSE)

Arguments

z

A real or complex vector or array (fft_stats()), or matrix (mvfft_stats()).

inverse

TRUE for the unnormalised inverse transform. NA means forward, as in stats::fft().

Details

A package moves by replacing ⁠fft(⁠ with ⁠zufft::fft_stats(⁠ and adding zufft to Imports.

Value

A complex vector, array or matrix with the attributes of z.

Examples

x <- 1:4
fft_stats(x)
fft_stats(fft_stats(x), inverse = TRUE) / length(x)
all.equal(fft_stats(x), stats::fft(x))

m <- matrix(rnorm(12), 4)
all.equal(mvfft_stats(m), stats::mvfft(m))

Threads per transform

Description

fft_workers() gets or sets the number of threads a transform uses when its workers argument is NULL. with_fft_workers() sets it while code runs and local_fft_workers() until the calling function returns, in the style of the withr package; both restore the old value. Together they are SciPy's get_workers() and set_workers().

Usage

fft_workers(n = NULL)

with_fft_workers(n, code)

local_fft_workers(n, .local_envir = parent.frame())

Arguments

n

A non-zero whole number. NULL (in fft_workers()) only reads the current value.

code

Code to run with n threads.

.local_envir

The environment whose exit restores the old value.

Details

A positive value asks for that many threads; a negative one counts back from the number of available cores, so -1 means all of them. Requests are capped at the available cores, and at 2 when ⁠_R_CHECK_LIMIT_CORES_⁠ is set (as during ⁠R CMD check⁠). The initial value is 1, as in SciPy.

Threads help when a transform has many independent lines (the columns of a matrix, a multi-dimensional array) and at least 32 768 elements; a single 1-D real transform always runs on one thread. Results agree to rounding whatever the number of threads, and are bit-identical on builds without fused multiply-add, such as CRAN's x86-64 binaries.

Value

fft_workers() returns the current value when n is NULL, and otherwise the previous value, invisibly. with_fft_workers() returns the value of code; local_fft_workers() returns the previous value, invisibly.

Examples

fft_workers()
with_fft_workers(2, fft_c2c(matrix(rnorm(1e4), 100), axes = 1))
f <- function() {
  local_fft_workers(2)
  fft_workers()
}
f()
fft_workers()

Information about the zufft build

Description

Reports which ducc0 source zufft was built from and how: the vendored commit and the SciPy commit that pins it, the SIMD backend and vector width the compiler selected, whether the target has fused multiply-add instructions, the threading mode, and the thread pool.

Usage

zufft_info()

Value

A named list:

ducc0_commit

ducc0 commit of the vendored FFT subset.

scipy_commit

SciPy commit whose vendoring script pins it.

simd_backend

"ducc0", "std::experimental::simd" or "none".

simd_width

Doubles per SIMD vector, 1 without SIMD.

fma

TRUE if the target has fused multiply-add.

fp_contract

"off" when the transforms never fuse multiply-adds, so results are bit-identical whatever workers is; "fast" for a GCC build targeting FMA hardware, where they can differ in the last bit. CRAN's builds are "off".

threading

"zufft" when zufft supplies the thread pool, "none" when threads are compiled out.

pool_size

Threads the pool can use now, the caller included.

hardware_threads

Threads the hardware offers.

complex_path

"zero-copy" when complex vectors are viewed in place, "marshal" when they are copied (design ยง12.2).

Examples

zufft_info()$simd_width