cuPeriod

Optimized, GPU-accelerated periodograms for astronomy.

cuPeriod computes period-search statistics for variable stars and transiting systems with fast CPU backends and GPU-accelerated paths that scale from a single light curve to millions. The NVIDIA CUDA fast paths are joined by a portable PyTorch backend that also runs on AMD, Intel, and Apple GPUs (and a CPU-only path), so the accelerated code is no longer NVIDIA-only. One API, one CLI, and an optional desktop GUI cover every method, with frictionless column handling, joint multi-band search, alias diagnostics, raw-spectrum output, and an N-best-periods utility.

📖 Documentation:https://cuperiod.readthedocs.io — a 5-minute quickstart, a full user guide, and the complete API reference, including installation, backends, the GUI, and benchmarks pages.

Every implementation is validated against the standard reference (astropy LombScargle / BoxLeastSquares) to floating-point round-off.

Why cuPeriod

• Validated, not just fast. GLS and BLS match astropy to floating-point round-off, and every other method is checked against an independent reference implementation (PyAstronomy, the supersmoother package and gatspy, or a direct implementation of the published algorithm) on identical grids. Across the seven benchmark-suite methods, CPU, CUDA, and the portable PyTorch backend agree to round-off and pick the identical best period on 126 real ASAS-SN light curves across six variability classes plus 12 confirmed Kepler transits — 88–96% harmonic-aware period recovery on this deliberately heterogeneous sample (a synthetic injection–recovery sweep further characterizes sensitivity vs. signal-to-noise) — see the full benchmark report and the benchmarks docs page. The multi-band methods are further validated blind on 100 real SDSS Stripe 82 RR Lyrae with literature periods — the pooled fold statistics recover up to 93% of periods strictly (97% harmonic-aware) on real five-band data.
• A fast CPU tier, no GPU required. The [fast] extra's multicore numba kernels make backend="cpu" 20x faster than astropy's BoxLeastSquares and 2177x faster than PyAstronomy's PDM on a representative light curve, while recovering the same periods.
• GPU acceleration beyond NVIDIA. The portable PyTorch backend runs every method on AMD (ROCm), Intel (XPU), and Apple (MPS) GPUs, in addition to the NVIDIA CUDA fast paths — so the accelerated code isn't locked to one vendor.
• Built for catalogue scale.batch_periodograms sustains up to 574 light curves/s (>2 million/hour) on a single GPU, with a resumable batch sink for runs spanning millions of curves.
• Eight methods, one API. GLS, BLS, PDM, CE, String-Length, MHAOV, TLS, and SuperSmoother share one entry point, one CLI, and an optional desktop GUI, with frictionless column handling.
• Multi-band search that earns its keep. Seven of the eight methods fit several filters of one star jointly. The native multi-band GLS offers three models — shared-phase offsets (the default), independent per-band sinusoids, and a regularized per-band harmonic model — runs on every backend (a six-band star over 200k frequencies: 55 s through astropy, 0.13 s native), and comes with bootstrap false-alarm probabilities, which astropy's LombScargleMultiband does not provide at all. On a simulated Rubin-like cadence with 30 epochs spread over six bands, single-band GLS recovers 0% of faint RRab proxies and the shared-phase model recovers 82%.
• Survey catalogues, in place.cuperiod.interop runs a search directly over LINCC nested-pandas / lsdb light curves — one row per object, epochs in a nested column — with no flattening or groupby. A partition tier reuses one GPU engine across every object in a dask partition, with verified column presets for ZTF and Rubin DP1.
• Alias-checked periods.alias_diagnostics measures the sampling's spectral window, predicts the alias family it implies, scores each competitor against the peak you got, and says plainly when the period is ambiguous.
• Pulsators get a frequency solution, not just a period.prewhiten automates the whole Period04-style loop — GPU/NUFFT amplitude spectrum, iterative sinusoid extraction, simultaneous re-fitting, principled stopping criteria, propagated uncertainties, combination-frequency identification, and g-mode period-spacing tools — for one star or a million.

Status

Implemented now, each with optimized CPU and GPU backends:

Method	What it's for	GPU	Multi-band
GLS	general variability (Lomb-Scargle)	yes	yes
BLS	eclipses / box-like transits	yes	yes
MHAOV	sharply non-sinusoidal signals (multiharmonic AOV)	yes	yes
TLS	limb-darkened transit matched filter	yes	—
PDM	non-sinusoidal folds (Stellingwerf)	yes	yes
CE	sparse survey data (conditional entropy)	yes	yes
String-Length	eclipsing / eccentric shapes	yes	yes
SuperSmoother	any repeating shape, non-parametric (Friedman)	yes	yes

All eight methods have CPU and GPU backends, plus the full single/batch/CLI machinery.

Install

pip install cuperiod # CPU (numpy, scipy, astropy, finufft)
pip install "cuperiod[gpu]"# + CUDA 12 GPU backends (cupy, cufinufft)
pip install "cuperiod[torch]"# + portable PyTorch backend (AMD/Intel/Apple GPUs + CPU)
pip install "cuperiod[fast]"# + numba multicore CPU kernels (7 methods, no GLS, 20-300x)
pip install "cuperiod[gui]"# + interactive desktop GUI (cuperiod-gui)
pip install "cuperiod[pandas]"# + pandas DataFrame ingestion
pip install "cuperiod[nested]"# + nested-pandas light curves (cuperiod.interop)
pip install "cuperiod[lsdb]"# + lsdb HATS catalogs (lazy, dask-partitioned)

The [gpu] extra needs an NVIDIA GPU with the CUDA 12 runtime; it pulls in cupy-cuda12x, cufinufft, and the CUDA runtime wheels. The [torch] extra adds a portable PyTorch backend that reaches AMD (ROCm), Intel (XPU), and Apple-Silicon (MPS) GPUs — and a CPU path everywhere — so the accelerated code runs beyond NVIDIA (install the wheel matching your accelerator from pytorch.org; the default is CPU-only). The [fast] extra adds multicore numba CPU kernels that become the default "cpu"/"auto" backend for every method but GLS — BLS, PDM, CE, String-Length, MHAOV, TLS, and SuperSmoother — one to two orders of magnitude faster than the fallback CPU paths and matching them to floating point.

Not sure what will run where? cuperiod doctor reports every installed backend, the torch devices it sees and the precision each uses, and what backend="auto" resolves to.

Quick start (Python)

The recommended import alias is cup:

importcuperiodascup# A single light curve, straight from arrays:pg=cup.periodogram((time, mag, mag_err), "GLS")
print(pg.best_period())
forpeakinpg.best_periods(10):
print(peak.period, peak.power, peak.extra.get("fap"))
# From a table with arbitrary column names (auto-detected, or pinned):pg=cup.periodogram(df, "BLS",
columns=cup.ColumnMap(time="HJD", value="flux", error="flux_err"))
# Several methods at once:res=cup.periodogram(lc, ["GLS", "BLS"]) # -> MultiResultres["BLS"].best_periods(5, alias_diverse=True)
# Raw spectrum for your own analysis:frequency, power=pg.frequency, pg.power

Method names are case-insensitive ("gls" == "GLS").

📓 New here? The examples/cuperiod_tour.ipynb notebook works through real light curves — a Cepheid, an RR Lyrae, an eclipsing binary, a Mira, and a Kepler exoplanet — showing each periodogram and phase-folded result.

Multi-band (one star, several filters)

GLS, BLS, MHAOV, PDM, CE, String-Length, and SuperSmoother jointly model two or more bands of the same star — one period, but each band keeps its own mean, amplitude, and normalization:

mb=cup.MultiBandLightCurve.from_light_curves({"g": lc_g, "r": lc_r})
pg=cup.periodogram(mb, "GLS") # VanderPlas & Ivezić shared-phase model# ...or straight from a long-format file / the CLI / batch:mb=cup.MultiBandLightCurve.from_file("star_ugrizy.parquet", band_column="band")
# One of three joint models (offsets / perband / flex):settings=cup.GLSSettings(mb_model="flex")
pg=cup.periodogram(mb, "GLS", settings=settings)
# Honest false-alarm probabilities, from a within-band bootstrap:calib=cup.multiband_fap(mb, settings, n_bootstrap=1000)
print(calib.fap(pg.best_periods(1)[0].power), calib.level(0.01))

mb_model is "offsets" (shared phase + per-band offsets, the default), "perband" (independent per-band sinusoids, chi2_0-weighted), or "flex" (regularized per-band harmonics, matching astropy to ~2e-10). All three run natively on finufft, cufinufft, and torch — astropy is kept only as the reference they're tested against.

Period recovery on a simulated Rubin-like six-band cadence (benchmarks/multiband_recovery.py; faint RRab proxies, 0.20 mag per-point noise, 300 stars per cell, top period within 1% and no harmonic credit), by total epochs across all six bands:

strategy	30 epochs	60 epochs	120 epochs
best single band (r)	0.0%	37.3%	97.3%
any single band	0.0%	49.3%	99.0%
multi-band offsets (1,0)	81.7%	99.7%	100.0%

perband and flex land in between (17.7% / 20.0% at 30 epochs, 97.0% at 60): sharing the phase is what buys sparse-cadence recovery, which is why "offsets" is the default. These are simulations on a deliberately simplified cadence (random nights, no rolling cadence) — read the ordering, not the absolute numbers. See the multi-band guide.

Survey catalogs (LINCC: nested-pandas / lsdb)

fromcuperiod.interopimportnested_periodogram, partition_periodogramout=nested_periodogram(frame, "lc", preset="ztf_dr22") # row-wiseres=partition_periodogram(cat, preset="rubin_dp1_object", # one GPU enginemethod="GLS", backend="gpu") # per partition

Runs on the nested layout directly — no flattening, no groupby. Column presets for ZTF DR22, ZTF alerts, and Rubin DP1 object/DIA photometry; a failed object yields NaN rather than aborting the run. See the interop guide.

Backends

backend="auto" (default) uses the GPU when available and falls back to CPU. Force a path with backend="cpu", backend="gpu", or a concrete name ("finufft", "cufinufft", "numpy", "astropy", "cupy"). The portable PyTorch backend runs any method on any torch device: backend="torch" (best device present) or "torch:cpu" / "torch:cuda" / "torch:mps" / "torch:xpu". On Apple MPS (no float64) it uses float32; precision="auto" keeps float64 everywhere else.

Pre-whitening a pulsator

solution=cup.prewhiten((time, mag, mag_err))
print(solution.summary()) # ranked frequencies with 1-sigma uncertainties and S/Nforcinsolution.components:
print(c.label, c.frequency, c.frequency_error, c.snr, c.combination)
series=cup.find_period_spacing( # gamma Dor / SPB g-mode pattern
[c.periodforcinsolution.independent()],
[c.amplitudeforcinsolution.independent()],
)

Every judgement call an interactive session leaves to the operator is an explicit setting, and the result records why the extraction stopped. See the pre-whitening guide.

Batch processing (millions of light curves)

# CPU pool across cores, written to Parquet:cup.batch_periodograms("lightcurves/*.parquet", ["GLS", "BLS"],
device="cpu", workers=8, sink="results/")
# GPU, with an auto-sized worker count:cup.batch_periodograms(df_groups, "GLS", device="gpu",
workers=cup.suggest_gpu_workers("GLS"), sink="out.parquet")

Inputs can be an iterable of light curves, a glob, a directory, or a (DataFrame, group_column) pair. A directory sink is resumable — re-running skips chunks already written. suggest_gpu_workers sizes the GPU pool from probed device memory.

Command line

cuperiod run star.csv --method GLS,BLS --n-best 10
cuperiod batch "lcs/*.csv" --method GLS --device gpu --out results/
cuperiod prewhiten star.csv --snr 4.6 -n 30 --spacing
cuperiod batch-prewhiten "lcs/*.csv" --out modes.parquet
cuperiod methods # list methods and backends
cuperiod gpu-info # device + suggested worker counts
cuperiod doctor # backends, torch devices, precision, auto-resolution
cuperiod grid-info star.fits -m GLS

run accepts --time/--value/--error column overrides and --domain magnitude|flux, and can write JSON (--out) and the raw spectrum (--save-periodogram). --band splits a long-format file on its filter column and runs a joint multi-band fit; the Python API's batch_periodograms(..., band_column=...) takes the same (the batch command has no band option yet).

Desktop GUI

An interactive explorer ships with the [gui] extra — run any method with all of its options, explore the full-resolution spectrum, and watch the phased light curve update live as you drag across peaks. Single curves or a whole folder in batch mode, multi-band overlays, and a dark/light theme remembered across launches. An Analysis picker switches the same window to pre-whitening: the amplitude spectrum with the residual overlaid, a frequency table with uncertainties, and a g-mode period-spacing explorer.

pip install "cuperiod[gui]"# add [gpu] or [torch] for accelerated backends
cuperiod-gui # or: python -m cuperiod.gui

The toolbar's Load demo menu always offers a synthetic multi-band curve, so there's something to explore on first launch. In a source checkout (or with CUPERIOD_EXAMPLE_DATA pointed at examples/data) it also lists a Kepler transit and six ASAS-SN variables.

Light-curve inputs

Time may be JD/HJD/BJD/MJD; values may be magnitude or flux; errors are optional. Column names are auto-detected (case-insensitive) and can be pinned with ColumnMap. Box/transit methods (BLS, TLS) work in flux — magnitudes are converted automatically.

Citing cuPeriod

If you use cuPeriod in your research, please cite the Research Note (ADS, doi:10.3847/2515-5172/ae9b06). It is also the preferred-citation in CITATION.cff, so GitHub's "Cite this repository" button and cffconvert both hand it back.

@ARTICLE{2026RNAAS..10..244J,
author = {{Jayasinghe}, Tharindu},
title = "{cuPeriod: Seven Validated Periodograms for CPUs and Any GPU}",
journal = {Research Notes of the American Astronomical Society},
keywords = {Stellar astronomy, Time domain astronomy, Variable stars, Astronomy software, 1583, 2109, 1761, 1855},
year = 2026,
month = aug,
volume = {10},
number = {8},
eid = {244},
pages = {244},
doi = {10.3847/2515-5172/ae9b06},
adsurl = {https://ui.adsabs.harvard.edu/abs/2026RNAAS..10..244J},
adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}

The Note describes the v1.1 release (the seven methods of its title); multi-band, SuperSmoother and pre-whitening arrived in v1.2. Cite the version you ran alongside it if that matters for reproducibility — cuperiod.__version__ reports it.

License

GPL-3.0-or-later.