nifty-ls

A fast Lomb-Scargle periodogram. It's nifty, and uses a NUFFT!

Overview

The Lomb-Scargle periodogram, used for identifying periodicity in irregularly-spaced observations, is useful but computationally expensive. However, it can be phrased mathematically as a pair of non-uniform FFTs (NUFFTs). This allows us to leverage Flatiron Institute's finufft package, which is really fast! It also enables GPU (CUDA) support and is several orders of magnitude more accurate than Astropy's Lomb Scargle with default settings.

Background

The Press & Rybicki (1989) method for Lomb-Scargle poses the computation as four weighted trigonometric sums that are solved with a pair of FFTs by "extirpolation" to an equi-spaced grid. Specifically, the sums are of the form:

$$\begin{align} S_k &= \sum_{j=1}^M h_j \sin(2 \pi f_k t_j), \\\ C_k &= \sum_{j=1}^M h_j \cos(2 \pi f_k t_j), \end{align}$$

where the $k$ subscript runs from 0 to $N$, the number of frequency bins, $f_k$ is the cyclic frequency of bin $k$, $t_j$ are the observation times (of which there are $M$), and $h_j$ are the weights.

The key observation for our purposes is that this is exactly what a non-uniform FFT computes! Specifically, a "type-1" (non-uniform to uniform) complex NUFFT in the finufft convention computes:

$$g_k = \sum_{j=1}^M h_j e^{i k t_j}.$$

The complex and real parts of this transform are Press & Rybicki's $S_k$ and $C_k$, with some adjustment for cyclic/angular frequencies, domain of $k$, real vs. complex transform, etc. finufft has a particularly fast and accurate spreading kernel ("exponential of semicircle") that it uses instead of Press & Rybicki's extirpolation.

There is some pre- and post-processing of $S_k$ and $C_k$ to compute the periodogram, which can become the bottleneck because finufft is so fast. This package also optimizes and parallelizes those computations.

Installation

From PyPI

For CPU support:

$ pip install nifty-ls

For GPU (CUDA) support:

$ pip install nifty-ls[cuda]

The default is to install with CUDA 12 support; one can use nifty-ls[cuda11] instead for CUDA 11 support (installs cupy-cuda11x).

From source

First, clone the repo and cd to the repo root:

$ git clone https://www.github.com/flatironinstitute/nifty-ls
$ cd nifty-ls

Then, to install with CPU support:

$ pip install .

To install with GPU (CUDA) support:

$ pip install .[cuda]

or .[cuda11] for CUDA 11.

For development (with automatic rebuilds enabled by default in pyproject.toml):

$ pip install nanobind scikit-build-core
$ pip install -e .[test] --no-build-isolation

Developers may also be interested in setting these keys in pyproject.toml:

[tool.scikit-build]
cmake.build-type = "Debug"
cmake.verbose = true
install.strip = false

For best performance

You may wish to compile and install finufft and cufinufft yourself so they will be built with optimizations for your hardware. To do so, first install nifty-ls, then follow the Python installation instructions for finufft and cufinufft, configuring the libraries as desired.

nifty-ls can likewise be built from source following the instructions above for best performance, but most of the heavy computations are offloaded to (cu)finufft, so the performance benefit is minimal.

Usage

From Astropy

Importing nifty_ls makes nifty-ls available via method="fastnifty" in Astropy's LombScargle module. The name is prefixed with "fast" as it's part of the fast family of methods that assume a regularly-spaced frequency grid.

import nifty_ls
from astropy.timeseries import LombScargle
frequency, power = LombScargle(t, y).autopower(method="fastnifty")

Full example

import matplotlib.pyplot as plt
import nifty_ls
import numpy as np
from astropy.timeseries import LombScargle

rng = np.random.default_rng(seed=123)
N = 1000
t = rng.uniform(0, 100, size=N)
y = np.sin(50 * t) + 1 + rng.poisson(size=N)

frequency, power = LombScargle(t, y).autopower(method='fastnifty')
plt.plot(frequency, power)
plt.xlabel('Frequency (cycles per unit time)')
plt.ylabel('Power')

To use the CUDA (cufinufft) backend, pass the appropriate argument via method_kws:

frequency, power = LombScargle(t, y).autopower(method="fastnifty", method_kws=dict(backend="cufinufft"))

In many cases, accelerating your periodogram is as simple as setting the method in your Astropy Lomb Scargle code! More advanced usage, such as computing multiple periodograms in parallel, should go directly through the nifty-ls interface.

From nifty-ls (native interface)

nifty-ls has its own interface that offers more flexibility than the Astropy interface for batched periodograms.

Single periodograms

A single periodogram can be computed through nifty-ls as:

import nifty_ls
# with automatic frequency grid:
nifty_res = nifty_ls.lombscargle(t, y, dy)

# with user-specified frequency grid:
nifty_res = nifty_ls.lombscargle(t, y, dy, fmin=0.1, fmax=10, Nf=10**6)

Full example

import nifty_ls
import numpy as np

rng = np.random.default_rng(seed=123)
N = 1000
t = np.sort(rng.uniform(0, 100, size=N))
y = np.sin(50 * t) + 1 + rng.poisson(size=N)

# with automatic frequency grid:
nifty_res = nifty_ls.lombscargle(t, y)

# with user-specified frequency grid:
nifty_res = nifty_ls.lombscargle(t, y, fmin=0.1, fmax=10, Nf=10**6)

plt.plot(nifty_res.freq(), nifty_res.power)
plt.xlabel('Frequency (cycles per unit time)')
plt.ylabel('Power')

Batched Periodograms

Batched periodograms (multiple objects with the same observation times) can be computed as:

import nifty_ls
import numpy as np

N_t = 100
N_obj = 10
Nf = 200

rng = np.random.default_rng()
t = np.sort(rng.random(N_t))
obj_freqs = rng.random(N_obj).reshape(-1,1)
y_batch = np.sin(obj_freqs * t)
dy_batch = rng.random(y_batch.shape)

batched = nifty_ls.lombscargle(t, y_batch, dy_batch, Nf=Nf)
print(batched.power.shape)  # (10, 200)

Note that this computes multiple periodograms simultaneously on a set of time series with the same observation times. This approach is particularly efficient for short time series, and/or when using the GPU.

Support for batching multiple time series with distinct observation times is not currently implemented, but is planned.

Limitations

The code only supports frequency grids with fixed spacing; however, finufft does support type 3 NUFFTs (non-uniform to non-uniform), which would enable arbitrary frequency grids. It's not clear how useful this is, so it hasn't been implemented, but please open a GitHub issue if this is of interest to you.

Performance

Using 16 cores of an Intel Icelake CPU and a NVIDIA A100 GPU, we obtain the following performance. First, we'll look at results from a single periodogram (i.e. unbatched):

In this case, finufft is 5x faster (11x with threads) than Astropy for large transforms, and 2x faster for (very) small transforms. Small transforms improve futher relative to Astropy with more frequency bins. (Dynamic multi-threaded dispatch of transforms is planned as a future feature which will especially benefit small $N$.)

cufinufft is 200x faster than Astropy for large $N$! The performance plateaus towards small $N$, mostly due to the overhead of sending data to the GPU and fetching the result. (Concurrent job execution on the GPU is another planned feature, which will especially help small $N$.)

The following demonstrates "batch mode", in which 10 periodograms are computed from 10 different time series with the same observation times:

Here, the finufft single-threaded advantage is consistently 6x across problem sizes, while the multi-threaded advantage is up to 30x for large transforms.

The 200x advantage of the GPU extends to even smaller $N$ in this case, since we're sending and receiving more data at once.

We see that both multi-threaded finufft and cufinufft particularly benefit from batched transforms, as this exposes more parallelism and amortizes fixed latencies.

We use FFTW_MEASURE for finufft in these benchmarks, which improves performance a few tens of percents.

Multi-threading hurts the performance of small problem sizes; the default behavior of nifty-ls is to use fewer threads in such cases. The "multi-threaded" line uses between 1 and 16 threads.

On the CPU, nifty-ls gets its performance not only through its use of finufft, but also by offloading the pre- and post-processing steps to compiled extensions. The extensions enable us to do much more processing element-wise, rather than array-wise. In other words, they enable "kernel fusion" (to borrow a term from GPU computing), increasing the compute density.

Accuracy

While we compared performance with Astropy's fast method, this isn't quite fair. nifty-ls is much more accurate than Astropy fast! Astropy fast uses Press & Rybicki's extirpolation approximation, trading accuracy for speed, but thanks to finufft, nifty-ls can have both.

In the figure below, we plot the median periodogram error in circles and the 99th percentile error in triangles for astropy, finufft, and cufinufft for a range of $N$ (and default $N_F \approx 12N$).

The astropy result is presented for two cases: a nominal case and a "worst case". Internally, astropy uses an FFT grid whose size is the next power of 2 above the target oversampling rate. Each jump to a new power of 2 typically yields an increase in accuracy. The "worst case", therefore, is the highest frequency that does not yield such a jump.

Errors of $\mathcal{O}(10\%)$ or greater are common with worst-case evaluations. Errors of $\mathcal{O}(1\%)$ or greater are common in typical evaluations. nifty-ls is conservatively 6 orders of magnitude more accurate.

The reference result in the above figure comes from the "phase winding" method, which uses trigonometric identities to avoid expensive sin and cos evaluations. One can also use astropy's fast method as a reference with exact evaluation enabled via use_fft=False. One finds the same result, but the phase winding is a few orders of magnitude faster (but still not competitive with finufft).

In summary, nifty-ls is highly accurate while also giving high performance.

float32 vs float64

While 32-bit floats provide a substantial speedup for finufft and cufinufft, we generally don't recommend their use for Lomb-Scargle. The reason is the challenging condition number of the problem. The condition number is the response in the output to a small perturbation in the input—in other words, the derivative. It can easily be shown that the derivative of a NUFFT with respect to the non-uniform points is proportional to $N$, the transform length (i.e. the number of modes). In other words, errors in the observation times are amplified by $\mathcal{O}(N)$. Since float32 has a relative error of $\mathcal{O}(10^{-7})$, transforms of length $10^5$ already suffer $\mathcal{O}(1\%)$ error. Therefore, we focus on float64 in nifty-ls, but float32 is also natively supported by all backends for adventurous users.

The condition number is also a likely contributor to the mild upward trend in error versus $N$ in the above figure, at least for finufft/cufinufft. With a relative error of $\mathcal{O}(10^{-16})$ for float64 and a transform length of $\mathcal{O}(10^{6})$, the minimum error is $\mathcal{O}(10^{-10})$.

Testing

First, install from source (pip install .[test]). Then, from the repo root, run:

$ pytest

The tests are defined in the tests/ directory, and include a mini-benchmark of nifty-ls and Astropy, shown below:

$ pytest
======================================================== test session starts =========================================================
platform linux -- Python 3.10.13, pytest-8.1.1, pluggy-1.4.0
benchmark: 4.0.0 (defaults: timer=time.perf_counter disable_gc=True min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /mnt/home/lgarrison/nifty-ls
configfile: pyproject.toml
plugins: benchmark-4.0.0, asdf-2.15.0, anyio-3.6.2, hypothesis-6.23.1
collected 36 items                                                                                                                   

tests/test_ls.py ......................                                                                                        [ 61%]
tests/test_perf.py ..............                                                                                              [100%]


----------------------------------------- benchmark 'Nf=1000': 5 tests ----------------------------------------
Name (time in ms)                       Min                Mean            StdDev            Rounds  Iterations
---------------------------------------------------------------------------------------------------------------
test_batched[finufft-1000]           6.8418 (1.0)        7.1821 (1.0)      0.1831 (1.32)         43           1
test_batched[cufinufft-1000]         7.7027 (1.13)       8.6634 (1.21)     0.9555 (6.89)         74           1
test_unbatched[finufft-1000]       110.7541 (16.19)    111.0603 (15.46)    0.1387 (1.0)          10           1
test_unbatched[astropy-1000]       441.2313 (64.49)    441.9655 (61.54)    1.0732 (7.74)          5           1
test_unbatched[cufinufft-1000]     488.2630 (71.36)    496.0788 (69.07)    6.1908 (44.63)         5           1
---------------------------------------------------------------------------------------------------------------

--------------------------------- benchmark 'Nf=10000': 3 tests ----------------------------------
Name (time in ms)            Min              Mean            StdDev            Rounds  Iterations
--------------------------------------------------------------------------------------------------
test[finufft-10000]       1.8481 (1.0)      1.8709 (1.0)      0.0347 (1.75)        507           1
test[cufinufft-10000]     5.1269 (2.77)     5.2052 (2.78)     0.3313 (16.72)       117           1
test[astropy-10000]       8.1725 (4.42)     8.2176 (4.39)     0.0198 (1.0)         113           1
--------------------------------------------------------------------------------------------------

----------------------------------- benchmark 'Nf=100000': 3 tests ----------------------------------
Name (time in ms)              Min               Mean            StdDev            Rounds  Iterations
-----------------------------------------------------------------------------------------------------
test[cufinufft-100000]      5.8566 (1.0)       6.0411 (1.0)      0.7407 (10.61)       159           1
test[finufft-100000]        6.9766 (1.19)      7.1816 (1.19)     0.0748 (1.07)        132           1
test[astropy-100000]       47.9246 (8.18)     48.0828 (7.96)     0.0698 (1.0)          19           1
-----------------------------------------------------------------------------------------------------

------------------------------------- benchmark 'Nf=1000000': 3 tests --------------------------------------
Name (time in ms)                  Min                  Mean            StdDev            Rounds  Iterations
------------------------------------------------------------------------------------------------------------
test[cufinufft-1000000]         8.0038 (1.0)          8.5193 (1.0)      1.3245 (1.62)         84           1
test[finufft-1000000]          74.9239 (9.36)        76.5690 (8.99)     0.8196 (1.0)          10           1
test[astropy-1000000]       1,430.4282 (178.72)   1,434.7986 (168.42)   5.5234 (6.74)          5           1
------------------------------------------------------------------------------------------------------------

Legend:
  Outliers: 1 Standard Deviation from Mean; 1.5 IQR (InterQuartile Range) from 1st Quartile and 3rd Quartile.
  OPS: Operations Per Second, computed as 1 / Mean
======================================================== 36 passed in 30.81s =========================================================

The results were obtained using 16 cores of an Intel Icelake CPU and 1 NVIDIA A100 GPU. The ratio of the runtime relative to the fastest are shown in parentheses. You may obtain very different performance on your platform! The slowest Astropy results in particular may depend on the Numpy distribution you have installed and its trig function performance.

Authors

nifty-ls was originally implemented by Lehman Garrison based on work done by Dan Foreman-Mackey in the dfm/nufft-ls repo, with consulting from Alex Barnett.

Citation

If you use nifty-ls in an academic work, please cite our RNAAS research note:

@article{Garrison_2024,
  doi = {10.3847/2515-5172/ad82cd},
  url = {https://dx.doi.org/10.3847/2515-5172/ad82cd},
  year = {2024},
  month = {oct},
  publisher = {The American Astronomical Society},
  volume = {8},
  number = {10},
  pages = {250},
  author = {Lehman H. Garrison and Dan Foreman-Mackey and Yu-hsuan Shih and Alex Barnett},
  title = {nifty-ls: Fast and Accurate Lomb–Scargle Periodograms Using a Non-uniform FFT},
  journal = {Research Notes of the AAS},
  abstract = {We present nifty-ls, a software package for fast and accurate evaluation of the Lomb–Scargle periodogram. nifty-ls leverages the fact that Lomb–Scargle can be computed using a non-uniform fast Fourier transform (NUFFT), which we evaluate with the Flatiron Institute NUFFT package (finufft). This approach achieves a many-fold speedup over the Press & Rybicki method as implemented in Astropy and is simultaneously many orders of magnitude more accurate. nifty-ls also supports fast evaluation on GPUs via CUDA and integrates with the Astropy Lomb–Scargle interface. nifty-ls is publicly available at https://github.com/flatironinstitute/nifty-ls/.}
}

A pre-print of the article is available on arXiv: https://arxiv.org/abs/2409.08090

Acknowledgements

nifty-ls builds directly on top of the excellent finufft package by Alex Barnett and others (see the finufft Acknowledgements).

Many parts of this package are an adaptation of Astropy LombScargle, in particular the Press & Rybicki (1989) method.