acceptodds
Under review as a conference paper at ICLR 2027

LagFormer: regressing periods from sparse, irregularly sampled series

Abstract

Period estimation for irregularly sampled series is usually a search over trial frequencies, on a grid whose density grows with record length and whose highest peak can be a harmonic of the true period. Several learned estimators take such frequency-resolved inputs and inherit the same grid. We introduce LagFormer, which identifies the period from self-similarity on a logarithmic lag axis, where a fixed relative resolution suffices. An exact correlogram bins every pair of observations by time difference, handling arbitrary sampling without interpolation, and compares observed values alongside channels learned by a transformer. Since each period bin spans a whole number of lag bins, every candidate period’s multiples lie at fixed lag offsets, which a harmonic comb gathers for a network outputting a distribution over log-period. A local refinement sharpens the estimate within the selected harmonic. Trained only on synthetic series that replay real sampling and noise, LagFormer finds the period within 10% on 90.2% of held-out synthetic series, against 69.3% for a Lomb–Scargle periodogram and 85.0% for PDM, without a fine global frequency search. On real light curves it agrees with the catalogue period on 80.6%, and its refined period folds the data more tightly than the catalogue period on 52.8% of a common same-harmonic subset, nearly the 53.7% from refining the catalogue period itself.

Primary area:
learning on time series and dynamical systems
Keywords:
Irregularly Sampled Time Series, Period Estimation, Self-Similarity, Lag-Domain Representation, Autocorrelation, Transformers, Time Series Representation Learning, Harmonic Structure, Distributional Regression

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EmYtOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper considers the problem of period estimation from irregularly sampled time series. The model involves latent representations obtained from transformers that augment the input data and to which correlograms and a shared combed network are applied, following the standard autocorrelation function approach to this problem.

Critical strengths

None identified.

Critical weaknesses

Numerical results (Tables 1 and 3) show only minor improvements from the input-only autocorrelation function setup (C=1) to the high-complexity LagFormer, and the results also show poor performance for other ML-based methods. Furthermore, the results are comparable to those obtained by legacy methods, so the advantages of the proposed approach are not clear.

There is little motivation for the role of the ML components of the data pipeline (e.g., why do transformers provide useful features? Why do the time series need to be augmented to compute the correlograms?) Hyperparameter selections are not well motivated.

While this is a sound signal processing problem, its application appears to be fairly focused to the astrophysics literature (as most papers considering the problem are published in venues in that area), and it is not clear why ICLR is a good match for the paper.

Questions for authors

What is the computational cost of the proposed method and how does it compare to the baselines, including the C=1 case? What is the role of the representations learned by the transformers?

4 comments
AuthorsOpenReview ↗

Response to Reviewer EmYt (1/4)

We thank the reviewer for the feedback. We address the three concerns in W1 in order, then summarise the advantages of the resulting formulation.

Critical Weakness 1. Numerical results (Tables 1 and 3) show only minor improvements from the input-only autocorrelation function setup (C=1) to the high-complexity LagFormer, and the results also show poor performance for other ML-based methods. Furthermore, the results are comparable to those obtained by legacy methods, so the advantages of the proposed approach are not clear.

1. The gain over C = 1. C = 1 uses raw values instead of learned features, but retains the lag-domain correlogram, harmonic comb, per-period network, output head and second-stage refinement, so it is not a classical autocorrelation baseline. The learned features are most useful when observations are corrupted.

ConditionC = 24C = 1Gain
Clean88.7288.23+0.49
5% outliers70.8564.74+6.11
10% outliers56.2547.01+9.24
10% sessions with systematic offsets64.3360.27+4.05

Across 129 stress conditions on five test sets, C = 24 has a positive 95% confidence interval in 125. Against periods from an independent survey (Catalina), C = 24 is correct where C = 1 is wrong on 1,962 series, versus 1,180 in the reverse direction. The second-stage transformer adds independently. Replacing it with a Gauss–Newton polish alone lowers the share of held-out real series with tight folds (θ < 0.4) from 67.7% to 4.7% (see Tables R4 and R7 in our reply to Reviewer ezF1, Q1).

2. The other ML methods. Generic learned representations infer periods from spectrum images or pooled embeddings, and reach only 16–26% Acc10 on irregularly sampled data. The most direct evidence comes from LagFormer's own encoder. Replacing its log-lag correlogram, harmonic comb and decoder with a generic MLP head reduces Acc10 from 90.20% to 16.43%. A transformer encoder alone therefore does not suffice for period estimation. The key is the structured lag-domain representation, which explicitly captures self-similarity across integer multiples of a period. The poor performance of the other ML methods reflects the difficulty of learning this structure from generic representations, rather than an inherent limitation of learning-based period estimation.

3. Legacy methods. Comparable accuracy at the cost of an exhaustive search. Legacy methods find the period by scanning about 1.1 million trial frequencies per series, which makes them 11–36 times slower than LagFormer on the same GPU (see Q1). Their answer is also confined to that discrete grid, so its precision is fixed by the grid spacing. LagFormer avoids both. It estimates the period directly from the observations and refines it in a continuous space, with no trial grid.

Despite this exhaustive search, legacy methods are less accurate. On 60,227 synthetic test series with known periods, LagFormer achieves 90.20% Acc10 against 85.04% for PDM, which is 34% fewer errors.

On real data, the apparent similarity comes from the reference. The ZTF catalogue's periods are selected with AoV, which shares its optimum with PDM, so agreement with it favours PDM by construction (see our reply to Reviewer ezF1, W2/Q3 (a)). Against independent labels, LagFormer leads all nine methods on both surveys (Catalina and Gaia DR3), and a label-free F-test on the real series finds PDM's period provably too short nearly twice as often as LagFormer's (see our reply to Reviewer ezF1, W2/Q3 (b, c)).

AuthorsOpenReview ↗

Response to Reviewer EmYt (2/4)

4. The advantages of the approach. The setting is among the hardest for period estimation. The time series are sparse (the median series has about 1,066 observations spread over 8,700 cycles, roughly one observation per eight cycles), irregular (a few observations per session, about 500 sessions over six years, with long seasonal gaps), and heteroscedastic (every observation carries its own uncertainty). In this setting, LagFormer offers the following advantages together.

  • No exhaustive search and no fixed grid. It estimates the period directly from the observations instead of scanning a trial-frequency grid, so its cost does not grow with the observing baseline, and it refines the estimate continuously rather than on a fixed grid. Its first stage is about 24 times faster than PDM on the same GPU.
  • The fundamental, not a harmonic. It resolves the harmonic ambiguity that is the dominant failure of classical searches, which often report half the true period, and makes fewer than half as many such errors as PDM.
  • Precision for multi-year folding. Its refinement keeps the fold coherent over thousands of cycles, with 97.4% of its periods staying within 0.2 cycle over the full baseline.
  • Better fits to real data, and generalisation. Trained only on synthetic time series and applied unchanged to real data, its periods fold the observations more tightly than those of the classical searches on over 80% of real series, and against independent labels it is the most accurate of the nine methods we compared.
  • Robust to corrupted measurements. Its learned features keep outliers and sessions with systematic offsets from dominating the estimate, a regime in which raw-value statistics degrade sharply.
  • One model for all signal types. The same network handles smooth periodic signals and signals with sharp, dip-like features, without per-class tuning or a choice of periodogram.
  • A differentiable estimate. Unlike a grid search, the period estimate is the output of a trainable network, so it can be fine-tuned on new data or used inside larger learned models.
AuthorsOpenReview ↗

Response to Reviewer EmYt (3/4)

Critical Weakness 2. There is little motivation for the role of the ML components of the data pipeline (e.g., why do transformers provide useful features? Why do the time series need to be augmented to compute the correlograms?) Hyperparameter selections are not well motivated.

W2. The role of each learned component, and how the hyperparameters were set. Each learned component addresses a specific difficulty of period estimation on sparse, irregular, heteroscedastic data. We describe them in the order in which data flow through the model.

  1. Transformer encoder. What each observation means. A raw value alone does not reveal whether an observation lies in a sharp dip, on a rising segment, or is unreliable. Its temporal neighbours and reported uncertainty provide this context. The encoder reads each observation in this context and maps it to 24 bounded features. The model then compares observations by what they represent rather than only by their raw values, and a single bad measurement cannot dominate every comparison.

  2. Log-lag correlogram. Why the series is transformed first. The correlogram measures the similarity between observations at every lag, which is where periodicity is detected. On raw values, every pair contributes a product of raw values, so an outlier contaminates every lag it touches while a narrow dip contributes weakly because few pairs contain it. A network applied afterwards cannot undo this. Computing the correlogram on learned features changes what is compared while keeping the exact lag structure. The raw-value correlogram is the special case C = 1, so this extends classical autocorrelation rather than replacing it. Each lag bin also records its number of contributing pairs, exposing where the observing cadence leaves limited support.

  3. Per-period network. Reading every multiple. Because period bins are aligned with lag bins, the k-th multiple of every period sits at a fixed offset, so the correlogram at all 218 multiples of each period is gathered in one operation. A convolutional network reads this harmonic map with dilated convolutions along the multiple index and a learned embedding of that index, which lets it compare neighbouring multiples and weigh them by position. At half the true period, the odd multiples fall where successive cycles do not repeat, which contradicts the shorter period.

  4. Harmonic decoder. Choosing among competing multiples. A convolutional network along the period axis turns these features into one score per period. It contains two harmonic convolution layers whose taps lie at fixed period ratios (from 1/5 to 5, including 1/2, 2 and 3). On a logarithmic axis these ratios are fixed offsets for every series, so a learned filter compares each period directly with its half, double and triple, which generic convolutions cannot do. Harmonic selection therefore comes from the architecture (see Table R10 in our reply to Reviewer QN72, Q3).

  5. HL-Gauss output head. Keeping competing answers apart. The output is a distribution over 512 log-period bins, trained with Gaussian-smoothed classification targets (HL-Gauss) rather than a regression loss. Periodic signals often give multimodal posteriors, because both P and 2P can explain the observations, and a point regression or the mean of a bimodal distribution can fall between the two modes and yield a wrong period. HL-Gauss keeps the hypotheses as separate modes, the decoder selects between them, and the estimate is read locally around the selected mode for sub-bin precision.

  6. Second-stage transformer. Precision without a search. The first stage identifies the period to a fraction of a bin, but multi-year series need much finer precision to keep their folds coherent over thousands of cycles. A second transformer reads lag tokens around the first-stage estimate and refines the period within the selected harmonic, followed by a Gauss–Newton polish. It evaluates 8 trial frequencies, where a classical local zoom needs about 6,147.

Hyperparameters. The structural settings follow from the representation rather than from tuning. The period-bin width is an integer multiple of the lag-bin width, which places every multiple at a fixed offset. K = 218 is the largest number of multiples the lag grid resolves, because each logarithmic lag bin spans a fixed fraction δ ≈ 0.0046 of its lag, and multiples beyond ⌈1/δ⌉ = 218 no longer occupy distinct bins. The lag-bin width has the greatest effect, because coarser bins blur the multiples that separate a period from its half and double, reducing Acc10 by 1.4–3.1 points. Additional multiples improve precision at low signal-to-noise and agreement on real data, at little additional cost. Model width, depth, learning rate and schedule were selected on validation data (Appendix A). Sensitivity results are in Table R9 (see our reply to Reviewer QN72, Q2).

AuthorsOpenReview ↗

Response to Reviewer EmYt (4/4)

Critical Weakness 3. While this is a sound signal processing problem, its application appears to be fairly focused to the astrophysics literature (as most papers considering the problem are published in venues in that area), and it is not clear why ICLR is a good match for the paper.

W3. Why ICLR. The contribution is a learning method. It is a representation that lets a network estimate the period of an irregularly sampled time series directly, without an exhaustive search over trial frequencies. Period estimation has appeared mainly in astronomy venues because its established solutions are search statistics developed there. A learned estimator that outperforms them is a representation-learning result.

  1. A core ICLR topic. Irregularly sampled time series are an established ICLR line, including multi-time attention networks (Shukla & Marlin, ICLR 2021) and graph-guided networks (Zhang et al., ICLR 2022). Learned periodicity identification has appeared at ICLR for regularly sampled series with missing values (AmortizedPeriod, Yu et al., ICLR 2024). We address truly irregular sampling, where FFT- and convolution-based learners do not apply.
  2. General contributions. None of them depends on the data domain. We compute an exact correlogram directly from irregular timestamps, on a logarithmic lag axis and in linear time, so it can serve as a layer in any sequence model. We compare learned features across lags, generalising autocorrelation beyond raw values. We align the lag and period grids, so that aggregating over all multiples of every period becomes a single learned gather. Finally, a learned refinement replaces the exhaustive search. The method needs only timestamps, values and their uncertainties, so it applies unchanged to irregularly sampled periodic series from any domain.
  3. Lessons for evaluating learned models. Both apply whenever a learned model is scored against a classical algorithm's labels or metrics. First, such labels favour the algorithm that produced them. On identical predictions, LagFormer beats PDM against true periods but trails it against labels from a phase-binned search, rewarding similarity to the labeller, not correctness (Table R1, reply to ezF1). Second, a fit-quality score can favour a wrong answer. Phase-fold tightness systematically prefers half the true period, so it is a fair comparison only between periods at the same multiple (reply to ezF1, Q3).
  4. Why astronomical data. We chose it to test the method at the extremes of the problem. The series are extremely sparse (the median has 1,066 observations over 8,702 cycles) and irregular, with about 500 short sessions over six years and long gaps, so there is no uniform grid. Each observation has its own uncertainty, outliers occur, and sharp signals create harmonic ambiguity. We also wish to introduce these surveys to the ML community as a benchmark for irregular time series, offering over a billion real series and independently measured periods for the same series.

Questions. What is the computational cost of the proposed method and how does it compare to the baselines, including the C=1 case? What is the role of the representations learned by the transformers?

Q1. Computational cost. LagFormer is an order of magnitude faster than classical searches. We timed every method alone on the same 2,000 real series and one GPU (RTX 6000 Ada). Times are sustained wall-clock, including start-up and loading.

MethodPer seriesvs LagFormer
LagFormer, first stageᵃ3.9 ms–
C = 1ᵃ3.7 ms–
PDM (fused GPU, ~1.1M freqs)93.5 ms24×
AoV (same kernel)94.3 ms24×
Conditional entropy (same kernel)42.2 ms11×
Lomb–Scargle (same GPU)140 ms36×

ᵃ Median of 5 runs.

The first stage evaluates no trial frequencies, so its cost does not grow with the time span, unlike grid searches. The second stage evaluates 8 trial frequencies per series, against about 6,147 for a classical local zoom. For over a billion series, the first stage needs about 45 GPU-days, against 1,080 for PDM.

C = 1 versus C = 24. The 24 learned features add about 0.2 ms per series (about 5%). The correlogram, harmonic comb, per-period network, decoder and second stage are kept in both settings. This overhead buys the robustness and accuracy gains in W1, point 1.

Q2. The role of the learned representations. They determine what the correlogram compares across lags. A raw value alone cannot separate an observation in a sharp signal feature from an unreliable one, and its neighbours and reported uncertainty supply this context. The transformer maps each observation, in this context, to bounded features, so the correlogram compares what observations represent and no single corrupted observation dominates. This is why the learned representations matter more as the data degrade, with a gain of 9.24 Acc10 points under 10% outliers (W1, point 1).

Reviewer_QN72OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates period estimation for sparse and irregularly sampled time series. Traditional methods typically rely on dense searches over candidate frequencies, which are computationally expensive and can confuse the true period with its harmonics or subharmonics. To address these challenges, the authors propose LagFormer, which reformulates period estimation as self-similarity inference in the logarithmic lag domain. Specifically, a Transformer learns observation-level features, a correlogram captures self-similarity across different time lags, and a harmonic comb aggregates information at candidate periods and their integer multiples. A subsequent local refinement stage further improves estimation precision. Overall, the paper presents a technically innovative approach supported by relatively comprehensive experiments. However, its competitiveness on real-world data and generalization across different sampling patterns remain somewhat limited.

Critical strengths
  1. Clear motivation and practical relevance. Period estimation for irregularly sampled time series is a meaningful problem with practical applications, as periodicity is an important characteristic of temporal data.
  2. Innovative design of the logarithmic lag space and harmonic comb. By exploiting the properties of the logarithmic space, the authors map integer multiples of different candidate periods to fixed relative offsets, enabling the harmonic comb to aggregate self-similarity information at these locations. Compared with directly regressing periods using a Transformer, this design introduces a more targeted inductive bias.
  3. Efficient handling of irregularly sampled data. LagFormer constructs a correlogram based on pairwise time differences, without requiring interpolation onto a regular time grid. Moreover, the correlogram statistics can be accumulated with a computational complexity of O(n,nlagC)O(n,n_{\mathrm{lag}}C), contributing to computational efficiency.
  4. Significant performance improvements on synthetic data.
Critical weaknesses
  1. Inconsistent performance advantages on real-world data. Although LagFormer substantially outperforms traditional methods on synthetic data, its advantages are less consistent on real-world datasets. Therefore, I suggest that the authors evaluate the method on additional real-world datasets with more reliable ground-truth periods or further investigate which estimates better explain the observed data when LagFormer and PDM disagree.
  2. Heavy reliance on local refinement for high-precision estimation. LagFormer adopts a two-stage framework, where the first stage identifies the candidate period and its harmonic structure, while the second stage refines the estimate within the selected harmonic. As shown in Tab. 2, the improvements in phase-folding quality largely stem from the second-stage local refinement rather than the proposed representation learning mechanism in the first stage. This makes the independent contribution of the core method to high-precision period estimation less convincing.
Questions for authors
  1. Tab. 3 shows that the raw-value-only model achieves comparable, or even slightly better, performance on real-world data. Under what conditions do the Transformer-learned features provide clear advantages?
  2. How sensitive is LagFormer to the number of harmonics K and the resolution of the logarithmic lag grid? Could a smaller number of harmonics achieve comparable performance with lower computational cost?
  3. Since the local refinement stage cannot correct an incorrectly selected harmonic, how does LagFormer distinguish the fundamental period from its integer multiples, particularly for signals with strong harmonic components?
7 comments
AuthorsOpenReview ↗

Response to Reviewer QN72 (1/5)

We thank the reviewer for the careful reading and the specific questions, and we answer each below.

Critical Weakness 1. Inconsistent performance advantages on real-world data. Although LagFormer substantially outperforms traditional methods on synthetic data, its advantages are less consistent on real-world datasets. Therefore, I suggest that the authors evaluate the method on additional real-world datasets with more reliable ground-truth periods or further investigate which estimates better explain the observed data when LagFormer and PDM disagree.

W1. Real-world performance.

We followed both suggestions, and both favour LagFormer. The apparent inconsistency on real data comes from the reference used, not from the method. We first explain why agreement with the ZTF catalogue understates LagFormer, and then give the two analyses, (1) periods from other surveys, and (2) a test on the light curves themselves that decides which period the data support.

Why agreement with the ZTF catalogue understates LagFormer. The ZTF catalogue period is itself a search result. SCoPe selects the candidate with the highest AoV significance among the top 50 Lomb–Scargle and conditional-entropy peaks (Healy et al. 2024, Sec. 2.4). At a fixed number of phase bins, AoV and PDM's θ share their optimum, so agreement with this catalogue favours PDM by construction. This effect is evident even on synthetic data. Against true periods, LagFormer outperforms PDM (90.2% vs 85.0%), while against phase-binned search labels the ranking reverses (82.7% vs 90.3%, see our reply to Reviewer ezF1, W2/Q3 (a)). The real-data gap (2.1 points) is smaller than this label effect.

(1) Against periods from other surveys, LagFormer leads every method (Table R6). We evaluated LagFormer zero-shot against independently published periods from Catalina (Drake et al. 2014, 44,890 stars) and Gaia DR3 (6,776 stars), without retraining or fine-tuning. LagFormer achieves the highest Acc10 among all nine methods on both surveys, 41.49% against 28.36% for PDM on Catalina, and 69.91% against 64.45% on Gaia. LagFormer is right where PDM is wrong on 6,977 Catalina stars, and the reverse holds for 1,080. On Gaia, the counts are 559 and 189.

Catalina's absolute agreement is lowered because 66% of its sample are contact binaries, where two nearly equal eclipses make the light curve repeat at half the labelled orbital period. Recovering the full period is a harmonic-selection problem, and LagFormer succeeds on 26.5% of these stars, compared with 9.7% for PDM. Among the 1,128 stars listed in both Catalina and the ZTF catalogue, LagFormer agrees with Catalina on 42.2%, whereas the ZTF catalogue agrees on only 31.8%, so LagFormer is not reproducing the ZTF catalogue (see our reply to Reviewer ezF1, W2/Q3 (c)).

(2) Where LagFormer and PDM disagree, an F-test on the data sides with LagFormer (Table R2). For a period P, an F-test checks whether the two halves of a 2P cycle differ significantly, as with the deep and shallow eclipses of a binary. If they do, P is provably too short. The threshold is calibrated on synthetic curves to a 1% false-positive rate and applied identically to every method on all 17,129 held-out curves.

PDM's period is provably too short on 13.0% of the curves, nearly twice the rate for LagFormer (6.7%) and above the catalogue (9.1%). When LagFormer predicts twice PDM's period, the data require the longer period in about half of the cases (51%), whereas when PDM predicts twice LagFormer's period, they require it in only 27%. After excluding provably too-short periods, LagFormer also keeps the most well-folding curves (θ < 0.4), 61.3% against 55.9% for PDM and 60.9% for the catalogue.

On the same observations, LagFormer also folds more tightly than PDM. Across the same 17,129 curves, LagFormer produces tighter folds than PDM on 13,990 curves, and PDM on 3,131, despite PDM explicitly minimising θ over its trial-frequency grid (see our reply to Reviewer ezF1, Q3 (second part)).

Against independent periods and against the observations themselves, LagFormer is the more accurate estimator on real data.

AuthorsOpenReview ↗

Response to Reviewer QN72 (2/5)

Critical Weakness 2. Heavy reliance on local refinement for high-precision estimation. LagFormer adopts a two-stage framework, where the first stage identifies the candidate period and its harmonic structure, while the second stage refines the estimate within the selected harmonic. As shown in Tab. 2, the improvements in phase-folding quality largely stem from the second-stage local refinement rather than the proposed representation learning mechanism in the first stage. This makes the independent contribution of the core method to high-precision period estimation less convincing.

W2. Why the method is split into two stages.

The two stages solve different problems. The first stage decides which period is correct, and every accuracy gain in the paper comes from it. The second stage measures that period to the precision a multi-year fold requires.

Multi-year folds demand extreme precision (Table R8, new). A small period error accumulates over thousands of cycles. The median held-out curve spans 6 years and about 8,700 cycles, so keeping the phase drift below 0.2 cycles requires a relative precision of about 2 × 10⁻⁵.

First stageAfter second stage
Median phase drift over the baseline10.3 cycles0.007 cycles
Curves within 0.2-cycle drift2.2%97.4%

Measured against the catalogue period, on curves where LagFormer and the catalogue select the same harmonic.

No single-pass estimator is built for this precision. The first stage predicts over 512 logarithmic period bins, each about 800 times wider than this requirement, and reaching it in one pass would need about 410,000 bins. Classical methods face the same requirement. PDM meets it through a dense search over about 1.1 million trial frequencies, and a classical local zoom needs about 6,147 evaluations per series. Our second stage needs 8.

The second stage is itself learned. A transformer on lag tokens, followed by a Gauss–Newton polish, reduces the median log-period error on synthetic series from 7.1 × 10⁻⁴ to 1.2 × 10⁻⁶ dex. Replacing the transformer with the polish alone leaves far fewer well-folding real curves (θ < 0.4), 4.7% instead of 67.7% (Table R7, see our reply to Reviewer ezF1, Q1).

Period identification is determined by the first stage. The second stage searches only within ±1.25% of the first-stage estimate, so it cannot change the selected harmonic, and Acc10 is identical before and after refinement. LagFormer's accuracy advantage over PDM therefore comes entirely from the first stage, mainly from selecting the correct harmonic (see our reply to Reviewer ezF1, Q4). On real observations, the F-test in W1 likewise supports LagFormer's choice of harmonic, which is again a first-stage decision.

AuthorsOpenReview ↗

Response to Reviewer QN72 (3/5)

Q1. Tab. 3 shows that the raw-value-only model achieves comparable, or even slightly better, performance on real-world data. Under what conditions do the Transformer-learned features provide clear advantages?

Q1. When the Transformer-learned features help.

The 24 learned channels help most when observations are corrupted by outliers or by systematic offsets between observing sessions, and they also help at low signal-to-noise and on non-sinusoidal signals. On clean, mostly sinusoidal series, the raw values already carry much of the periodic signal, which is why the two arms perform comparably in Table 3.

The advantage grows under corruption, reaching +9.2 points (Table R4). We evaluate both arms on 6,601 unused synthetic test series with the class mix of the real held-out series, under controlled perturbations, averaged over three seeds.

ConditionC = 24C = 1Gain
None88.7288.23+0.49
5% outliers70.8564.74+6.11
10% outliers56.2547.01+9.24
10% outliers, amplitudes drawn from real data72.5067.91+4.60
10% of sessions offset64.3360.27+4.05
Half the sessions removed, 10% outliers37.1531.35+5.80

Across a broader grid of 129 conditions and their combinations, C = 24 is ahead, with a 95% interval above zero, in 125. The only exceptions combine all stresses at their most extreme levels, where both arms fall below 30% Acc10. This follows from the correlogram construction. A raw-value channel propagates an outlier into every lag product involving that observation, whereas the learned channels are bounded.

The learned channels also improve precision on uncorrupted series. On the series where both arms estimate the period within 10%, the median relative error is 7% lower with C = 24 on synthetic data (against the true period) and 9% lower on real data (against the catalogue period).

On real data, independent labels favour C = 24. Against periods from an independent survey (Catalina), C = 24 is right where C = 1 is wrong on 1,962 series, against 1,180 in the reverse direction. On the real series themselves, the F-test from W1 finds C = 24's period too short on 14% fewer series than C = 1's.

Reviewer_ezF1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LagFormer, a neural network designed to identify repeating cycles (periodicity) in patchy, unevenly spaced time-series data—such as telescope observations tracking changes in a star's brightness over time. Instead of laboriously scanning through millions of possible frequencies one by one, the model checks how similar measurements are across various time delays ("lags"), maps them on a logarithmic scale, and uses a harmonic filter to lock onto repeating rhythms. A subsequent refinement step then sharpens the precision of the estimated cycle.

On synthetic test data, LagFormer identifies the correct period within a 10% margin of error in 90.2% of cases, outperforming traditional astronomical tools like the Lomb–Scargle periodogram (69.3%) and Phase Dispersion Minimization (85.0%). On real telescope data, it matches established astronomical catalog periods 80.6% of the time. The paper's core contribution is marrying lag-based pattern matching with data-driven harmonic filtering, delivering fast, efficient period estimates without needing an exhaustive, brute-force frequency search.

Critical strengths

• Smart, resource-efficient design: Bypasses computationally expensive, fine-grained frequency searches by pairing time-lag correlations with logarithmic harmonic mapping.

• Effective harmonic disambiguation: Backed by sound mathematical principles and outperforms classical baselines on synthetic data, notably when distinguishing true fundamental cycles from their harmonic echoes (such as double or half-speed rhythms).

Critical weaknesses
  1. Incremental novelty: The underlying building blocks—measuring similarity across time lags, logarithmic representations, and combining harmonic multiples—are well-established techniques. The innovation lies primarily in stitching these existing ideas together.

  2. Unconvincing real-world performance: On actual telescope measurements, LagFormer trails behind traditional Phase Dispersion Minimization (80.6% vs. 82.8% catalog agreement). Consequently, its practical superiority on real-world data remains unproven.

Questions for authors
  1. Value of the transformer architecture: A simpler baseline using raw autocorrelation achieves virtually the same synthetic accuracy (89.5% vs. 90.2%) and slightly higher catalog agreement on real data. What specific advantages does the transformer network bring to the table? Does the real breakthrough stem purely from the lag-based representation and harmonic filter rather than the neural network itself?

  2. Distinction from prior work: Both lag-based correlation for uneven time series and logarithmic harmonic pooling have been explored in earlier literature. Could you clarify what aspects of your formulation are fundamentally new, beyond combining these pre-existing concepts?

  3. Demonstrating real-world accuracy: LagFormer aligns with catalog periods less often than standard Phase Dispersion Minimization (PDM). While catalog entries are not always infallible ground truth, what concrete evidence proves that your tool recovers real-world periods more accurately? Additionally, why were the curve-folding comparisons evaluated on differing subsets rather than head-to-head on the exact same observations?

  4. Cycle detection versus harmonic disambiguation: When predictions off by a factor of two (half or double periods) are counted as acceptable, PDM slightly edges out LagFormer on synthetic data. Does this indicate that LagFormer's primary advantage is simply choosing the correct harmonic multiple rather than detecting periodicity more effectively? Could you quantify how much of the performance gain is driven purely by harmonic selection?

10 comments
AuthorsOpenReview ↗

Response to Reviewer ezF1 (1/7)

We thank the reviewer for the careful reading and the specific questions, and we answer each below.

Critical Weakness 1. Incremental novelty: The underlying building blocks (measuring similarity across time lags, logarithmic representations, and combining harmonic multiples) are well-established techniques. The innovation lies primarily in stitching these existing ideas together.

W1 / Q2: Novelty beyond combining known techniques.

The novelty lies in how we formulate period estimation, rather than in any individual building block. We represent periodic structure through self-similarity on a logarithmic lag axis, where harmonic multiples of every candidate period sit at fixed offsets. Because the representation is built entirely from pairwise time differences, it is invariant to shifts in the time origin and naturally accommodates irregularly sampled observations. We highlight four aspects that, to our knowledge, are not present in prior work.

  1. An exact log-lag correlogram for irregular sampling at linear cost. Existing approaches either rely on uniform sampling (RepNet, TransRAC, Autoformer's FFT autocorrelation) or enumerate observation pairs and bin them on a linear lag axis at O(n²) cost (DCF, Edelson & Krolik 1988, Rehfeld et al. 2011). We instead accumulate every ordered pair exactly into logarithmically spaced lag bins using prefix sums, at O(n · n_lag · C) cost and without interpolation (Sec. 3.4, Appendix B). For the median synthetic series (about 4,900 cycles), a linear grid with the same resolution at short lags would need about 10⁸–10⁹ bins, while the logarithmic grid covers the same range with about 3 × 10³.

  2. A fixed-index representation of harmonic structure. We choose the period-bin width to be an integer multiple of the lag-bin width. Consequently, the k-th multiple of any period always maps to the same index offset, regardless of the period. This lets us gather 218 multiples across 512 periods with a single fixed-index operation. A shared network then learns how to combine this harmonic information. Its equal-weight special case aggregates value dispersion over the multiples, akin to classical fold statistics (Sec. 3.5).

  3. Self-similarity across learned channels. Rather than correlating only the raw observed values, our correlogram compares C = 24 learned per-observation channels across time lags. This extends the autocorrelation/variogram pair to learned representations (Sec. 3.4), allowing the model to capture periodic structure beyond similarity in the original values.

  4. A learned period regressor for irregularly sampled series. Existing learned period regressors (Claytor 2022, Schochet 2025) take a periodogram or wavelet spectrum as input, so their inputs still require a trial-frequency grid. LagFormer instead operates directly on irregularly sampled observations, without a frequency-resolved representation.

The existing pieces, used as they are, do not solve the problem. With our encoder and an MLP head, a learned regressor reaches 16.43% Acc10, and a regressor on a periodogram image reaches 25.72%. Replacing that MLP head with the log-lag correlogram, comb and decoder raises Acc10 to 90.20%, above every other period-finding method we compared, so the gain comes from the formulation and not the encoder.

AuthorsOpenReview ↗

Response to Reviewer ezF1 (2/7)

Critical Weaknesses. Unconvincing real-world performance: On actual telescope measurements, LagFormer trails behind traditional Phase Dispersion Minimization (80.6% vs. 82.8% catalog agreement). Consequently, its practical superiority on real-world data remains unproven.

Q3. Demonstrating real-world accuracy: LagFormer aligns with catalog periods less often than standard Phase Dispersion Minimization (PDM). While catalog entries are not always infallible ground truth, what concrete evidence proves that your tool recovers real-world periods more accurately?

W2 / Q3. Real-data performance and PDM's higher catalogue agreement.

Why PDM agrees more often with the catalogue (82.8 vs 80.6). The catalogue selection procedure is closely aligned with PDM's objective. SCoPe selects the candidate with the highest AoV significance among the top 50 LS and CE peaks (Healy et al. 2024, Sec. 2.4). At a fixed number of phase bins, AoV and PDM's θ are monotone transforms of each other (θ = SS_w / SS_tot and AoV ∝ SS_b / SS_w), so they share their optimum. The real set was also selected using θ ≤ 0.7 at the catalogue period. PDM's higher catalogue agreement is therefore expected in part from this shared optimisation criterion, and it is not, by itself, evidence of greater period accuracy.

(a) Synthetic ground-truth comparison (Table R1). We evaluate identical predictions against both the known true periods and catalogue-style search labels.

Estimatorvs true Pvs LS labelvs CE labelvs PDM label
Lomb–Scargle69.27(100)78.2078.25
PDM85.0478.2590.29(100)
CE82.1678.18(100)90.29
LagFormer (3 seeds)90.1870.8682.6685.41

PDM label = AoV label (same argmax at equal bins).

Against ground truth, LagFormer outperforms PDM by 5.1 points. The ranking reverses against CE labels, where PDM leads by 7.6 points. The real-data gap (2.1 points) is smaller than this label effect. PDM's apparent advantage therefore depends on the reference. It leads against CE labels, which are derived from phase-binned search statistics closely related to its own objective, while against the actual periods LagFormer performs better. Agreement with a search-derived catalogue cannot, on its own, establish which method recovers the period more accurately.

(b) A label-free test on real ZTF data (Table R2). To test real-world errors without relying on catalogue labels, we apply a label-free diagnostic to all 17,129 held-out ZTF light curves. For each estimated period P, we fit a Fourier series at frequency 1/(2P). The even harmonics describe a signal repeating every P, while the odd harmonics capture differences between the two halves of a 2P cycle. If adding the odd harmonics significantly improves the fit, the estimated period P is too short, because the light curve distinguishes the two halves and instead repeats over 2P.

We calibrate the F-test threshold on synthetic curves to a 1% false-positive rate, with 87% power, and apply the same test to every estimator. 2,228 PDM periods are identified as too short, compared with 1,154 for LagFormer and 1,553 for the catalogue.

Period sourcePeriods identified as too shortRemaining curves with θ < 0.4
PDM2,2289,583
LagFormer1,15410,502
Catalogue1,55310,435

After excluding periods flagged by this test, LagFormer retains 10,502 well-folding curves (θ < 0.4), compared with 9,583 for PDM and 10,435 for the catalogue. The advantage is therefore not merely closer agreement with a reference label. On the same real light curves, an independent harmonic test finds substantially fewer half-period errors in LagFormer's predictions, while more of its remaining predictions retain good phase-fold quality. In direct factor-2 disagreements, the data require LagFormer's longer period on 1,291 of 2,553 curves, and PDM's on 31 of 113.

AuthorsOpenReview ↗

Response to Reviewer ezF1 (3/7)

W2 / Q3 (continued)

(c) Validation against independent surveys (Table R6). Finally, we evaluate LagFormer zero-shot against independently published periods from two external surveys, Catalina (Drake et al., 2014) and Gaia DR3, on the same ZTF light curves. Without retraining or fine-tuning, LagFormer agrees with their published periods more often than any classical search, which shows that its advantage does not depend on the ZTF catalogue labels.

On Acc10, LagFormer leads all nine methods. On the Catalina sample of 44,890 stars, LagFormer agrees with the published periods in 41.49% of cases, compared with 28.36% for PDM. On the Gaia DR3 sample of 6,776 stars, agreement is 69.91% for LagFormer and 64.45% for PDM. The pairwise comparison further highlights this difference. On Catalina, LagFormer agrees with the independent period where PDM does not for 6,977 stars, while the reverse occurs for 1,080. On Gaia DR3, the corresponding counts are 559 and 189.

The aggregate Catalina agreement requires one qualification. 66% of its sample are contact binaries, where two nearly equal eclipses make the light curve repeat at half the labelled orbital period. Recovering the full orbital period is therefore a harmonic-selection problem, and LagFormer succeeds on 26.5% of these stars, compared with 9.7% for PDM and 0.9% for Lomb–Scargle. On the 8,408 pulsators, where this ambiguity does not arise, LagFormer and the classical search methods all reach 84–90%.

LagFormer is also not reproducing the ZTF catalogue. Where the two catalogues overlap (1,128 stars), LagFormer matches the Catalina period on 42.2% of them, while the ZTF catalogue period itself matches it on only 31.8%.

Together, these results show that LagFormer's advantage holds against period labels that are independent of the ZTF catalogue, where it agrees with the published periods more often than PDM and every other method we compared.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.