A redesigned pulse-train estimator could make a demanding period-search calculation more practical on parallel hardware, according to a new preprint. In its simulated benchmark, the compiled correlation-and-gather version was reported as more than five times faster when a GPU was available, while its memory footprint fell from 1 GB to 167 MB after just-in-time compilation.
The calculation estimates the time between repeated pulses in a signal. The paper examines whether changes to the algorithm and its implementation can lower both runtime and memory enough for time-domain maximum-likelihood estimation to be feasible in real-time use. It is a computational test of the estimator, with the reported comparisons based on synthetic signals and benchmark runs.
A costly grid search
The estimator searches over period and initial delay, which are treated as nonlinear parameters, while the pulse amplitude is handled as a linear parameter. That split permits a separable least-squares calculation: one part searches the grid, while the linear part can be solved separately under the modeled noise assumptions. In the paper's signal model, the observations are periodic, non-overlapping pulses that are short compared with the period, scaled by an unknown factor and corrupted by additive white Gaussian noise, commonly shortened to AWGN.
The first optimization uses the fact that most samples in a short pulse train carry zeros or no pulse information. Sparse matrix operations reduce the stated cost of projecting the model onto the data, and precomputing model terms is reported to reduce inference work, though it consumes more memory. This trade-off is central to the redesign: a faster calculation is useful only if the hardware can hold the intermediate results.
The second optimization separates the calculation into two stages. The signal is first correlated with the pulse, then a sparse gather assembles the periodic candidates needed for the grid search. To handle shifts smaller than one sample, the implementation uses pulse values and time derivatives in a Taylor approximation. The paper's reported runtime model reflects that split, with one term tied to the number of measurements and another tied to the grid and pulse count.
The test setup
For the benchmark, each signal contained 20,000 samples taken at 1-second intervals. The pulse was a 50-second Hann waveform, and the nominal period was 1,000 seconds. The estimator searched periods from 975 to 1,025 seconds in 0.1-second steps, with a 3-second delay resolution. The comparison covered 1,000 runs and used a global batch size of 16.
Where the savings came from
The asymptotic comparison helps explain why the implementations diverged as the search grew. Dense matrix multiplication scales with the number of grid points and measurements. Sparse multiplication adds the number of pulses and the nonzero samples in each pulse to that calculation. Correlation and gather instead combine a measurement-and-pulse term with a grid-and-pulse term, separating measurement-dependent work from grid-dependent work.
In the measured GPU comparisons, batching the matrix calculations was faster than processing grid points separately. Sparse versions were faster still, and the compiled correlation-and-gather implementation was reported as more than five times faster when a GPU was available. The paper does not report confidence intervals or repeated-measurement uncertainty for this comparison, so the exact size and stability of the speed advantage remain unclear.
Memory use showed a similarly large gap. The uncompiled baseline required over 37 GB, while its JIT version was reported at 89% lower memory use. Dense batched matrix multiplication used 13.4 GB and was unaffected by compilation. The correlation-and-gather approach used 1 GB, dropping to 167 MB after compilation, the smallest footprint reported.
Speed without a reported accuracy trade-off
The lower resource use did not come with a reported difference among the implementations in simulated mean squared error, a measure of average squared estimation error. Their performance was described as very similar. At low signal-to-noise ratio, the time-domain models outperformed multiharmonic estimators, while methods that did not use the short-pulse structure performed worse. At high signal-to-noise ratio, the methods reached the Cramér-Rao lower bound, a theoretical benchmark for estimator accuracy.
What the benchmark leaves open
The results remain bounded by the test conditions. The signals were simulated, the benchmark used a fixed Hann pulse and a limited period and delay grid, and the hardware test was run on a single high-end server with an AMD Threadripper PRO 9975WX processor with 32 cores and an NVIDIA RTX 6000 Ada GPU with 48 GB of VRAM. The evidence therefore does not answer whether the same runtime or memory gains would persist on lower-end, embedded or non-NVIDIA systems, or with other pulse shapes and parameter ranges.
Because the shift handling is an approximation, its accuracy beyond the tested conditions remains an open question. The MSE discussion gives qualitative comparisons without numerical estimates, error bars or statistical tests. The supplied evidence also does not include a live industrial deployment or an explicit operational latency target.
The document is an arXiv version 1 preprint dated 28 August 2026. It acknowledges financial support from the Austrian Federal Ministry of Economy, Energy and Tourism, the National Foundation for Research, Technology and Development, and the Christian Doppler Research Association.
Paper data and sources
Original title: Fast Time-Domain MLE for Period Estimation of Pulse Trains
Authors: Sebastian Schertler, Daniel Guger, Stefan Schuster et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text