- 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
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.
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.
-
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.
-
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.
-
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.
-
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).
-
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.
-
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).
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.
- 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.
- 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.
- 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).
- 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.
| Method | Per series | vs LagFormer |
|---|---|---|
| LagFormer, first stageᵃ | 3.9 ms | – |
| C = 1ᵃ | 3.7 ms | – |
| PDM (fused GPU, ~1.1M freqs) | 93.5 ms | 24× |
| AoV (same kernel) | 94.3 ms | 24× |
| Conditional entropy (same kernel) | 42.2 ms | 11× |
| Lomb–Scargle (same GPU) | 140 ms | 36× |
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).
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.
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.
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)).