{"id":"5ff25d9e-0f0e-4587-9891-211ac3f9987e","arxiv_id":"2412.00772","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A wavelet-based tokenizer called WQ4TS embeds time series from different domains into a shared spectral latent space, enabling a transformer to transfer across forecasting, imputation, and classification tasks.","lead":"This paper presents WQ4TS, a pretraining method that converts time series from many domains into a shared wavelet-based token space before a transformer processes them, aiming at cross-domain transfer for forecasting, imputation, and classification. The authors report best performance on 87.5% of tested tasks and average gains up to 34.7% over strong baselines in few-shot and zero-shot settings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Finite wavebook projection is not a bijection: Eqs. 11-14 define a lossy linear map to R^λ, so the claimed common spectral space and semantic tokens are unsupported.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the finite λ-dimensional wavebook projection is not a bijection. My stress-test sharpens this by pointing out that the implemented filters are not mutually orthogonal and have differing lengths, so the map is not even a projection onto a well-defined orthonormal coordinate system. The paper's central contribution—an interpretable common spectral latent space that supports zero/few-shot cross-domain transfer—depends on the bijection claim; if the map is lossy, the mechanism by which domains are aligned is unsupported. The experimental section is extensive, and if the results are reproducible they would be a useful empirical finding, but the theoretical claim is internally inconsistent as written. The reader's REJECT verdict is appropriate; no adjustment is needed.","tokens_in":57159,"tokens_out":5400,"duration_ms":66842,"concrete_test":"Implement Eqs. (11)-(14) with a concrete orthogonal wavelet (e.g., Haar or Daubechies-2), the paper's λ=100, and any reasonable m and f_c. Form the matrix F whose columns are the effective filters (differenced, scaled, centered A_i) restricted to a common window of length L=max_i m·S_i. Compute rank(F) and the condition number of F^T F. If rank(F)<λ or the condition number exceeds ~10^6, the wavebook is not a linearly independent basis, so the map cannot be bijective. Then construct two distinct length-L sub-series x and x' with identical token vectors under the exact Eq. 14 operation (e.g., by taking a nonzero null-space vector of the finite map). If such a pair exists, the tokenization is not injective and the bijection claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the tokenization in Eqs. (11)-(14) bijectively maps each local sub-series to a λ-dimensional coordinate in a common spectral latent space, giving each token semantic information and enabling cross-domain transfer (Sections 3.1-3.3). Proposition 2 proves bijectivity only for an infinite orthonormal wavelet basis of L2(R); the implemented method uses a finite dictionary of λ scaled, sampled wavelet filters A_i ∈ R^{m·S_i}, with λ=100 in all experiments. Each token entry p_{i,j} is a convolution of the series with one such filter (plus differencing), so the map from a length-L segment to R^λ is linear. A finite linear map can be bijective only if L=λ and the matrix is invertible. Here the filters have different lengths, are not mutually orthogonal (scaled copies of a wavelet at different scales are generically not orthonormal), and L is not constrained to equal λ. Consequently the map is generically non-injective, tokens are lossy projections rather than coordinates in an orthonormal basis, and the 'bijection' and 'interpretable semantic information' claims are unsupported. This is an internal inconsistency in the paper's central mechanism, not merely a disagreement with external consensus. The empirical numbers could still hold, but the stated theoretical foundation for the common-space transfer does not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WQ4TS, a cross-domain time-series pretraining method built on a wavelet-based tokenization. A 'wavebook' of λ=100 sampled, scaled wavelet filters is used to map each local segment of a series to a λ-dimensional token of 'pattern similarities'; the tokens are fed to an encoder-only Transformer that is pretrained across domains and then fine-tuned or applied zero-shot to forecasting, imputation, and classification. The paper claims that the tokenization is bijective, that tokens carry semantic information in a common spectral latent space, and that WQ4TS achieves the best performance on 87.5% of tasks with an average improvement up to 34.7%.","tokens_in":57472,"tokens_out":10084,"duration_ms":96320,"significance":"The empirical scope is broad and the central idea—a generic, model-agnostic wavelet tokenization that reduces cross-domain heterogeneity—is interesting. The paper evaluates three tasks, three data-scarcity settings, multiple forecasting benchmarks, and 35 UCR classification datasets, and it includes an ablation and a plug-in study over existing backbones. However, the theoretical foundation for the common spectral space is not established for the implemented finite wavebook, and the reproducibility of the experiments is limited by missing implementation details, absent error bars, and no released code. If the theoretical claims were corrected or substantially weakened, the empirical study could still be a useful contribution; as written, the load-bearing justification for the method's cross-domain transfer mechanism is unsupported.","major_comments":[{"comment":"Proposition 2 proves coefficient uniqueness only for an infinite orthonormal wavelet basis of L2(R). The implemented map in Eqs. (11)–(14) uses λ=100 filters A_i of different lengths that are resampled from a single amplitude sequence; this is a finite linear map from a local segment to R^λ, with no proof that the segment dimension equals λ or that the filter matrix is invertible. In fact, filters of different lengths and scales are generically not mutually orthogonal. The claims that 'the bijection relation is satisfied between the sub-series and the coordinates' and that each token has 'semantic information' therefore do not follow. This is load-bearing because the common spectral latent space and the cross-domain transfer story rest on this bijection.","section":"Sections 3.2–3.3, Proposition 2 and Eqs. (11)–(14)"},{"comment":"The wavebook construction is underspecified. The manuscript never states which orthogonal wavelet F(t) is used, which amplitude sequence A∈R^{2m} is chosen, what precision m is, what central frequency fc is, or how F(t) is 'designed to adapt to the characteristic of the data domain' as Algorithm 1 instructs. Since these choices completely determine the basis functions in Eqs. (11)–(12), the method is not reproducible as described, and the sensitivity of the results to these choices is unknown.","section":"Section 3.2 and Algorithm 1"},{"comment":"Although Section 5.1 states that all experiments are repeated three times, no error bars, standard deviations, or per-seed results are reported anywhere, and no code is released. This matters particularly because several reported gaps are very large, e.g., zero-shot imputation ETTm1→Exchange in Table 9 reports MSE 0.003 for WQ4TS versus 3.107 for FEDformer. Without uncertainty estimates or code, the robustness and fairness of these comparisons cannot be verified.","section":"Section 5.1 and Tables 4–10"},{"comment":"The definition p_{i,j}=d_{i,j+n/2-1} requires n=m·S_i to be even, but no evenness constraint is stated for the scale factors S_i. For odd n, the index j+n/2-1 is not an integer and the expression is undefined. In addition, the effective 'local segment' length is never defined, and it varies with i because each filter A_i has a different length; this further undermines the interpretation of token_j as a coordinate of a single fixed sub-series.","section":"Section 3.3, Eq. (14)"}],"minor_comments":[{"comment":"The baseline called OneFitsAll in Table 4 appears as GPT4TS in Table 16 with identical numbers; the naming should be consistent.","section":"Table 4 vs Table 16"},{"comment":"The construction of the 'Resample dataset' is not described; the caption mentions rate=10 but the resampling procedure and the relation to the original sampling rate are unclear.","section":"Table 1"},{"comment":"The text says the tokenization has 'no specific requirements for hyperparameters', yet λ, m, fc, and the number of encoder layers L are free hyperparameters listed in Table 3; the claim should be qualified as 'no dataset-specific tuning'.","section":"Section 3, introduction"},{"comment":"The notation |F_{m,n}(t)|^2 in the denominator should be the squared L2 norm of F_{m,n}, not a pointwise absolute value.","section":"Proof of Proposition 2, Eq. (1)"},{"comment":"The abstract states 'best performance on 87.5% of all tasks', but Table 16's '1st Count' shows 78 wins out of 80 metric entries (97.5%); the discrepancy between 'tasks' and 'metric entries' should be clarified.","section":"Abstract and Table 16"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is extensive, but the central theoretical claim is not supported for the finite wavebook actually used. If the authors can provide a finite-dictionary injectivity or frame argument, the theory could be repaired; otherwise the bijection and semantic-information claims should be removed or explicitly downgraded to heuristic motivations. I would also expect error bars and code before this can be certified as a reproducible empirical contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper has a genuinely interesting idea—wavelet-based tokenization for cross-domain time series—and a very large set of experiments. The central theoretical claim, however, does not survive contact with the method's own equations. The finite wavebook projection in Eqs. (11)–(14) is a linear map into R^λ, and with λ=100 and non-orthogonal filters of different lengths, it cannot be the bijection that Proposition 2 promises. That proposition is about the infinite orthonormal wavelet basis of L2(R), which is not what the implementation uses. So the 'shared spectral latent space' and 'interpretable semantic tokens' claims are unsupported. The stress-test note is correct, and this is the load-bearing flaw, not a nit.\n\nWhat the paper does well: the empirical evaluation is broad and honest in scope—forecasting, imputation, and classification across many datasets, with full-data, few-shot, and zero-shot settings. If the numbers are reproducible, the plug-in tokenization gives consistent improvements over strong baselines, which would be a useful subfield contribution. The design is simple and potentially valuable.\n\nThe soft spots beyond the theory: no code released, no error bars, and a few baseline numbers look off (e.g., imputation MSE 2–3 for some FEDformer cases). Hyperparameters like the central frequency and wavelet choice are not specified clearly. The paper also says the wavebook is 'orthogonal' but the sampled filters are not mutually orthogonal.\n\nI do not think this deserves to be published as is. But it deserves a serious referee, because the empirical phenomenon—if real—is interesting enough to warrant revision. A revised version should either prove the finite case or drop the bijection language, release code, add error bars, and fix the baseline table.\n\nMy verdict: I would not cite this until the math is fixed and code is out. It is a good reading-group paper to dissect the gap between infinite-basis theory and finite dictionaries.","headline":"The empirical sweep is real, but the load-bearing bijection theorem is proved for an infinite basis, not the finite wavebook the method actually uses.","tokens_in":57953,"tokens_out":2836,"would_cite":false,"duration_ms":26907,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","42C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that time series from many domains can be mapped into one shared spectral space by wavelet quantization, so a single pretrained transformer handles forecasting, imputation, and classification across domains with few or…","keywords":["time series","cross-domain transfer","wave quantization","tokenization","spectral latent space","pretraining","zero-shot learning","WQ4TS"],"falsifier":"Run the tokenization of Equations (13)--(14) on all test windows of two datasets with very different dynamics, such as ETTh1 and Weather, and attempt to invert the finite $\\lambda$-dimensional projection, or simply count how often distinct segments produce near-identical token vectors. If the reconstruction error is substantial, or token collisions are frequent as $\\lambda$ shrinks, the claimed bijection and the semantic-content guarantee would be refuted.","tokens_in":56984,"feed_emoji":"🌊","tokens_out":10125,"duration_ms":76337,"temperature":0.7,"pith_summary":"The paper argues that the main obstacle to cross-domain time series learning is that raw series from different domains look nothing alike, so a model trained on one source fails on another. It proposes WQ4TS, a pretraining method that projects every series into a common spectral latent space built from a fixed set of wavelet basis functions, the 'wavebook', turning each local window of the series into a token of fluctuation-pattern similarities. A single Transformer encoder pretrained in that shared space is then applied to forecasting, imputation, and classification across datasets with different sampling rates, periodicities, and scales, under full-data, few-shot, or zero-shot supervision. If the claim holds, one pretrained model could replace per-dataset training, which matters because labeled time series data is costly while unlabeled series are abundant.","feed_headline":"Wavelet tokens cut cross-domain time series error up to 34.7%","feed_subtitle":"A common wavelet space lets one pretrained model serve many time series domains with little or no target data.","key_machinery":"The wavebook: a set of $\\lambda$ orthogonal wavelet basis functions $\\{A_1, \\ldots, A_{\\lambda}\\}$ obtained by rescaling one orthogonal wavelet at scale factors $S_i = 2 f_c \\lambda / i$, together spanning a $\\lambda$-dimensional common spectral latent space $V^{\\lambda}$. The tokenization computes, for each timestep $j$, the fluctuation pattern similarity between the local window around $j$ and each basis function, producing a token in $\\mathbb{R}^{\\lambda}$; the tokens are the coordinates of the series in $V^{\\lambda}$. This machinery does the work of the argument: it is what claims to make cross-domain transfer possible, since all domains are projected through the same fixed basis set, and it is what the paper claims to be interpretable and bijective, with each token carrying semantic pattern information and sub-series and coordinates corresponding one-to-one.","core_discovery":"WQ4TS introduces the 'wave as token' principle: a set of orthogonal wavelet basis functions (the wavebook) forms a $\\lambda$-dimensional common space $V^{\\lambda}$, and each local segment of an input series is mapped to coordinates in that space by an inner-product convolution, so $\\text{token}_j = (p_{1,j}, \\ldots, p_{\\lambda,j})$ records the pattern similarity between the segment at timestep $j$ and each basis function. Because every domain is projected through the same fixed wavebook, the paper claims the resulting tokens live in one shared, interpretable space with semantic content, and that the sub-series-to-coordinates map is bijective. A vanilla Transformer encoder is pretrained on tokens from multiple domains with a learnable multi-task weighting, then full-parameter fine-tuned on the target domain. In experiments across seven forecasting/imputation datasets and 35 UCR classification datasets, WQ4TS reports the best performance on 87.5% of all tasks, average metric improvement up to 34.7%, and 25.8% and 44.1% improvements in few-shot and zero-shot settings; multi-domain pretraining beats single-domain pretraining, which the paper reads as evidence that the common space mitigates negative migration.","pith_inferences":["Inference: if the tokens are genuinely shared coordinates, the same framework should extend to anomaly detection and regression on series with very different physical meanings, such as ECG to power load, a setting the paper does not test.","Inference: the bijection claim, if it held, would make the token stream a lossless encoding and suggest generative decoding from tokens is possible; the paper never reconstructs series from tokens, so reconstruction error is a natural untested check.","Inference: because the finite wavebook is a projection, token collisions should increase as $\\lambda$ shrinks; measuring collision rates across the paper's datasets would quantify how much semantic information is actually preserved.","Inference: the analogy to subword tokenization implies attention operates on roughly equal-information units, so token-importance analysis could identify which waveform patterns drive cross-domain transfer, an interpretability study the paper does not perform."],"forward_implications":["A single pretrained model can be deployed on a new time series dataset with no retraining on that domain's data, bridging differences in sampling rate, channel count, and periodicity.","Few-shot performance with 5% of target training data approaches what current specialized models achieve with full data, lowering the annotation cost for forecasting, imputation, and classification.","Adding the wave quantize module improves existing models such as DLinear, PatchTST, FEDformer, Autoformer, OneFitsAll, and TimesNet, so the transfer benefit is not tied to the WQ4TS architecture itself.","Multi-domain pretraining outperforms single-domain pretraining, indicating that the common space reduces negative migration rather than injecting noise from unrelated domains."],"supporting_citations":[{"why":"VQ-VAE motivates the design of a quantization-based tokenization strategy for projecting raw inputs into a latent space.","marker":"van den Oord et al. [2017]"},{"why":"Informer supplies the ETT benchmark datasets used in forecasting and imputation and serves as a Transformer baseline.","marker":"Zhou et al. [2021]"},{"why":"PatchTST is the patch-as-token approach that the wave-as-token strategy is contrasted with, and a central baseline.","marker":"Nie et al. [2023]"},{"why":"Autoformer uses FFT-based spectral analysis, the line of work the common spectral space extends, and is a baseline.","marker":"Wu et al. [2021]"},{"why":"FEDformer computes dependencies in the frequency domain and is a baseline whose performance improves when the wave quantize module is plugged in.","marker":"Zhou et al. [2022a]"},{"why":"OneFitsAll is the LLM-empowered cross-domain baseline that WQ4TS must beat in forecasting, imputation, and classification.","marker":"Zhou et al. [2023]"},{"why":"Provides the uncertainty-based learnable loss weighting used to balance gradients across domains in multi-domain pretraining.","marker":"Kendall et al. [2018]"},{"why":"The UCR archive is the source of the 35 classification datasets used for full-data and few-shot cross-domain evaluation.","marker":"Dau et al. [2018]"},{"why":"TimesNet is a CNN-based general time series baseline used across all three tasks.","marker":"Wu et al. [2023]"}],"fun_headline_variants":["Wavelet tokens unify time series domains, cut error 34.7%","One wavelet space, many time series: 34.7% lower error","Wave quantization: cross-domain time series, best on 87.5% tasks","Wavelet tokens transfer across time series, zero-shot gain 44.1%","WQ4TS: a spectral latent space for cross-domain time series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method assumes that projecting a series onto a fixed finite set of wavelet shapes (the wavebook) preserves each local pattern without loss, so that tokens carry the same information in every domain, but the paper proves lossless bijection only for an infinite wavelet family, not for the finite wavebook actually used.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet tokens unify time series domains, cut error 34.7%","One wavelet space, many time series: 34.7% lower error","Wave quantization: cross-domain time series, best on 87.5% tasks","Wavelet tokens transfer across time series, zero-shot gain 44.1%","WQ4TS: a spectral latent space for cross-domain time series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000292,"raw_usage":{"total_tokens":1791,"prompt_tokens":1119,"completion_tokens":672,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":735,"completion_tokens_details":{"reasoning_tokens":569}},"tokens_in":735,"tokens_out":672,"duration_ms":5482,"temperature":1.0,"reasoning_tokens":569,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:00:32.449712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the tokenization of Equations (13)--(14) on all test windows of two datasets with very different dynamics, such as ETTh1 and Weather, and attempt to invert the finite $\\lambda$-dimensional projection, or simply count how often distinct segments produce near-identical token vectors. If the reconstruction error is substantial, or token collisions are frequent as $\\lambda$ shrinks, the claimed bijection and the semantic-content guarantee would be refuted.","supporting_citations":[],"review_version":1}