{"id":"3d57c24b-415b-48ee-b0c6-a09d22136203","arxiv_id":"2411.13315","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"NMF with wind-based classification attributes most NO2 and PM2.5 in central Taiwan to domestic sources, most SO2 to transboundary sources, and PM10 about equally.","lead":"This paper applies a standard data-reduction technique, non-negative matrix factorization, to hourly air quality readings from central Taiwan and sorts the extracted patterns into local versus imported pollution using wind data. It reports that nitrogen dioxide and fine particles are mostly local, sulfur dioxide mostly imported, and coarse particles split about evenly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 is internally inconsistent and unvalidated: the SO2 main result (45.83% domestic) differs from the paper's own 2010 validation (26.92%) by ~19 percentage points despite a claimed ±6% agreement, and the NO2 row only sums to 66.36% if a component explicitly labeled transboundary is counted…","rationale":"The reader's REJECT verdict is correct, but the strongest reason is not merely the subjective wind-speed threshold. The paper's own validation section is not a validation of Table 3: the SO2 numbers differ by about 19 percentage points, and the claimed ±6% agreement applies only to a different k, a different year, and a different analysis. Additionally, the NO2 row of Table 3 can only be obtained if a component that the text explicitly classifies as transboundary is silently treated as domestic. Both are objective internal inconsistencies that directly undermine the central percentages, independent of how one judges the physical plausibility of wind-speed-based classification. The reader identified the validation mismatch as a supporting point but did not make it the load-bearing concern; the NO2 arithmetic inconsistency was not mentioned. Since my concern reinforces the existing rejection rather than changing it, the verdict remains unchanged.","tokens_in":8693,"tokens_out":7801,"duration_ms":81260,"concrete_test":"Recompute every row of Table 3 by summing the component shares listed in Section 5 under the classifications stated there. For NO2, verify whether 16.79+16.58+16.43+16.56 equals 66.36 even though NMF6 is labeled transboundary; if NMF6 must be moved to domestic to match, the table contains an error. Then rerun the SO2 analysis for 2008–2017 using the same k and classification protocol as the Section 6 validation (k=10, 2010-style wind-speed rule) to see whether the 45.83% domestic figure collapses to roughly 27%. If it does not, the validation mismatch reflects a real methodological inconsistency; if it does, Table 3's SO2 row is not reproducible under the paper's own validation protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table 3, but Table 3 is not supported by the paper's own validation and cannot be reproduced from the text's classifications. For SO2, Table 3 reports 45.83% domestic for 2008–2017 with k=5, while Section 6's 2010 validation run with k=10 gives 26.92% domestic, close to Chen et al.'s 27%. The 18.9-point gap is far outside the '±6% margin' claimed in the conclusion, meaning Table 4 validates the 2010 k=10 analysis, not the headline result. The paper never discusses this discrepancy. There is also an internal arithmetic problem in the NO2 row: Section 5 labels NMF1, NMF2, and NMF4 domestic and NMF3, NMF5, and NMF6 transboundary; summing the reported shares gives 49.8% domestic, not 66.36%. Table 3's 66.36% is recovered only by counting NMF6 (explicitly called transboundary) as domestic. Thus the printed percentages do not cohere with the paper's own classifications or its validation, independent of any judgment about wind-speed thresholds.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies non-negative matrix factorization (NMF) to hourly air-quality data for SO2, NO2, PM10, and PM2.5 from 14 monitoring stations in central Taiwan over 2008–2017 (87,672 hourly records). For each pollutant the data matrix is decomposed into k components; each component is then classified as 'domestic' or 'transboundary' based on wind-rose patterns, and the component shares are aggregated into Table 3, which reports domestic/transboundary proportions of 66.36%/33.64% for NO2, 45.83%/54.17% for SO2, 50.95%/49.05% for PM10, and 64.65%/35.35% for PM2.5. The paper concludes that NO2 and PM2.5 are domestically dominated, SO2 is substantially transboundary, and PM10 is nearly balanced. Section 6 re-runs the analysis on 2010 data with different k values (k=8 for NO2, k=10 for O3 and SO2) and compares the resulting domestic ratios against Chen et al. (2010), reporting agreement within a claimed ±6% margin.","tokens_in":8916,"tokens_out":7393,"duration_ms":69692,"significance":"If the reported proportions were robust, the paper would provide useful quantitative input for central Taiwan's air-quality policy, separating domestic from transboundary contributions. The study has several strengths: NMF is a reasonable tool for nonnegative compositional data; the validation against independent 2010 estimates from Chen et al. is a good idea; and including O3 as a sanity check (even though O3 is not part of the headline results) shows an attempt at external grounding. However, the paper ships no code, no data-availability statement, no error bars, and no sensitivity analysis for the classification rule. More importantly, the headline percentages in Table 3 are not backed by the manuscript's own component classifications, and the validation in Section 6 uses a re-configured version of the method rather than the main analysis. If the internal inconsistency and validation gap were fixed, the qualitative finding—especially for SO2 and PM10—could still be defensible, but the quantitative claims in Table 3 are not currently supported.","major_comments":[{"comment":"The NO2 row of Table 3 is internally inconsistent with the component classifications given in Section 5. The text labels NMF1, NMF2, and NMF4 as domestic (16.79%, 16.58%, 16.43%) and NMF3, NMF5, and NMF6 as transboundary (16.58%, 16.68%, 16.56%). Summing the domestic shares gives 49.80%, not 66.36%; the table's 66.36% is recovered only by counting NMF6—explicitly called transboundary in the text—as domestic. This is a load-bearing arithmetic error because the abstract's statement that NO2 is 'primarily influenced by local sources' hinges on this row.","section":"Section 5, Table 3"},{"comment":"The validation in Section 6 does not actually validate Table 3. The main analysis uses k=6 for NO2, k=5 for SO2, k=4 for PM10, and k=3 for PM2.5 on the full 2008–2017 record, while Section 6 uses k=8 for NO2 and k=10 for O3 and SO2 on 2010 data only. For SO2, Table 3 reports 45.83% domestic for 2008–2017, but the 2010 k=10 run yields 26.92% domestic—an 18.9-percentage-point difference that is far outside the '±6% margin' claimed in Section 7. The paper never discusses this discrepancy; the conclusion's claim that the method is effective is therefore not supported for the headline results.","section":"Section 6 vs Table 3"},{"comment":"The domestic/transboundary attribution is made by a visual inspection of wind roses without any quantitative decision rule or sensitivity analysis. For example, in the NO2 analysis, NMF1 is called domestic because high pollution is concentrated at 4 m/s, and NMF2 is domestic at 3.5 m/s, but NMF6—with high pollution concentrated at 3.9 m/s—is classified as transboundary with no stated meteorological justification. Changing the implicit threshold by a few m/s would directly alter every percentage in Table 3. A sensitivity analysis, an objective classifier based on wind direction and speed distributions, or at least a clear statement of the applied rule with error margins is needed.","section":"Section 5, classification rule"},{"comment":"The k-selection text is ambiguous and inconsistent with the displayed numbers of components. Section 4.1 describes Figures 5(a)–(d) as 'k=6 and k=5' and 'k=4 and k=3' without stating which pollutant each panel refers to, and the text does not clearly explain which k is chosen for each pollutant. The Results section then uses six components for NO2 (although Figure 6's caption says 'five components'), five for SO2, four for PM10, and three for PM2.5. The link between the cophenetic-correlation panels and the chosen k must be stated explicitly, and the figure captions should match the number of components actually used.","section":"Section 4.1 and Section 5"}],"minor_comments":[{"comment":"The roadmap in the introduction says 'Section 2 outlines the methods, Section 3 describes data sources, Section 4 presents the analysis results, followed by a validation of the proposed method in Section 5. Section 6 concludes the study,' but the actual structure is Section 4 (analysis), Section 5 (results), Section 6 (validation), Section 7 (conclusion). Please correct the roadmap.","section":"Section 1"},{"comment":"The seasonal percentages for SO2 in Table 2 sum to 99.99% rather than 100%; check the rounding and report consistent totals.","section":"Table 2"},{"comment":"The PM2.5 analysis in Section 5 refers to 'Fig. 13 reveals distinct seasonal variations,' but Figure 13 is the PM10 day-to-year variation plot. The PM2.5 day-to-year variation is shown in Figure 16.","section":"Figures 13 and 16"},{"comment":"In the derivation of the multiplicative update rules, the notation '(W H H^T)' should be written as '((W H) H^T)' to avoid ambiguity about the order of multiplication; this is a presentation issue only.","section":"Section 2, Eq. (4)–(6)"},{"comment":"Reference [14] is cited as 'Chen et al.' but the entry lists the journal as 'Science of the Total Environment, Atmospheric Environment'—two journal names—and should be corrected; also consider adding a data-availability or code-availability statement, since the public EPA data link is given but the analysis code is not.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency in the NO2 row of Table 3 and the mismatch between the main SO2 result (45.83% domestic) and its validation run (26.92% domestic) are serious enough that in some venues they would justify rejection. I am recommending major revision rather than reject because both issues are fixable: the authors can re-examine the classification of NMF6, rerun the validation under the same k and time period as the main analysis (or add a thorough sensitivity analysis over k and classification thresholds), and correct the text accordingly. I would also note that the paper's fit with a numerical-analysis journal (math.NA) is marginal, since the NMF implementation is standard and the contribution is an applied data analysis rather than a numerical method. If the journal is open to applied studies, the scope is acceptable; otherwise the editors may want to consider fit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central result, Table 3, doesn't hold up. The NO2 row contradicts the paper's own classifications, and the validation section checks a different analysis than the one reported. This is not a matter of taste; the numbers don't cohere.\n\nWhat's genuinely useful: the paper compiles ten years of hourly data from 14 stations in central Taiwan and applies NMF to separate pollution features. That dataset is real, and the visualizations are thorough. The idea of labeling NMF components by wind speed/direction is a routine extension of existing practice, but it's applied honestly here. The comparison to Chen et al. is a reasonable external anchor.\n\nThe soft spots are load-bearing. The wind-speed classification—components peaking at ~4 m/s called domestic, those reaching 7–12 m/s transboundary—has no thresholds, error bars, or sensitivity analysis. A different cutoff directly changes every percentage in Table 3. More seriously, the validation doesn't validate the main result. The 2010 validation runs use different k values (k=8 for NO2, k=10 for SO2 and O3) and include O3, which isn't in the main analysis. For SO2, the main result is 45.83% domestic (k=5), but the validation gives 26.92% (k=10), close to Chen et al.'s 27%. The paper claims agreement within ±6%, but the gap to the headline number is ~19 points. That gap is never discussed. And there's an internal arithmetic error: in the NO2 main analysis, summing the reported shares of components the text labels domestic gives 49.8%, not the 66.36% in Table 3. You only get 66.36% by counting one component the text explicitly calls transboundary as domestic. These are not cosmetic issues.\n\nWho this is for: researchers working on Taiwan air quality or on source apportionment with matrix factorization might find the regional numbers worth a second look, but only after the errors are fixed. As it stands, the paper is a cautionary example of validation drifting from the headline analysis.\n\nRecommendation: reject in current form. But I'd send it to peer review rather than desk reject, because the dataset and question are real and the errors are fixable in revision. If the authors correct the arithmetic, run sensitivity on the wind-speed thresholds, and validate with the same k and years as the main analysis, this could become a decent regional case study.","headline":"Table 3 is internally inconsistent and unvalidated; the paper's central claim is not supported by its own analysis.","tokens_in":9498,"tokens_out":4235,"would_cite":false,"duration_ms":41690,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A23","65F55"],"pacs":[],"model":"deepseek-v4-flash","headline":"Factoring a decade of hourly readings from 14 stations, the paper concludes that NO2 and PM2.5 in central Taiwan are mostly local, SO2 mostly imported, and PM10 split nearly evenly.","keywords":["non-negative matrix factorization","air pollution","data visualization","source apportionment","domestic vs transboundary pollution","wind-rose analysis","central Taiwan","PM2.5 and NO2"],"falsifier":"Recompute the source-apportionment table with a single explicit wind-speed cutoff — for example the 5.5 m/s boundary of the paper's own wind classification, or 6 m/s — applied uniformly to every NMF component, and rerun the validation using the same component counts as the main analysis: if NO2's domestic share falls below half, or if the validation procedure does not reproduce the main proportions when applied to the full 2008–2017 record, the central claim fails.","tokens_in":8439,"feed_emoji":"🌬️","tokens_out":12371,"duration_ms":113669,"temperature":0.7,"pith_summary":"This paper tries to establish where central Taiwan's air pollution comes from: which pollutants are generated locally and which are carried in from outside the island. The authors factor a decade of hourly concentration readings from 14 monitoring stations into a small number of nonnegative source components, then label each component domestic or transboundary according to the wind speeds and directions at which it peaks. Their central claim is that NO2 and PM2.5 are predominantly domestic, SO2 is predominantly transboundary, and PM10 is split almost evenly. A sympathetic reader would care because source attribution determines who is responsible for cleaning the air in a region that hosts Taiwan's largest thermal power plant, a politically charged emitter.","feed_headline":"Matrix math separates local from imported smog in central Taiwan","feed_subtitle":"Factoring a decade of hourly monitoring data puts NO2 and PM2.5 on local sources, SO2 on imports.","key_machinery":"The load-bearing object is the NMF factorization $A \\approx WH$, with $W$ the time profiles and $H$ the spatial loadings of $k$ source components, computed by multiplicative update rules that keep every entry non-negative; non-negativity is what lets each component be read as an additive, physically meaningful source profile. Around it sits the cophenetic-correlation rule for choosing $k$ and, decisively, the wind-rose inspection that assigns each component to 'domestic' or 'transboundary' by the wind speed at which its pollution load concentrates. The wind-speed signature is what converts a statistical factorization into a geographic attribution.","core_discovery":"The discovery, stated on the paper's own terms, is a source-apportionment table: over 2008–2017 in central Taiwan, NO2 is 66.36% domestic and 33.64% transboundary, PM2.5 is 64.65% domestic, SO2 is 45.83% domestic (54.17% transboundary), and PM10 is 50.95% domestic (49.05% transboundary). These numbers come from decomposing each pollutant's $87{,}672 \\times 14$ hourly matrix into spatial and temporal factors with non-negative matrix factorization, choosing the number of factors by cophenetic correlation, and then reading each factor's wind rose: factors whose pollution concentrates at gentle wind speeds around 3.5–4.8 m/s are classified as domestic, while factors driven by monsoon winds of 7–12 m/s are classified as transboundary. The paper further claims that this classification scheme is validated by comparing a 2010 re-run of the procedure, with larger numbers of factors, against an independently estimated 2010 domestic-ratio benchmark, finding agreement within about six percentage points for NO2, SO2, and O3.","pith_inferences":["A direct test the paper leaves unwritten: refit the classification with one explicit wind-speed cutoff (for instance the 5.5 m/s 'gentle breeze' boundary of its own wind classification) applied uniformly, and watch whether NO2 and PM2.5 stay majority domestic; the percentages in the table move with the cutoff.","The validation section examines 2010 only and uses different numbers of components than the main runs; comparing each separate year of the decade, or re-running the 2010 validation at the main component counts, would show whether the domestic-majority claim is stable or an artifact of the chosen factor count.","Plugging the same factors into back-trajectory or chemical-transport modeling would let each NMF component be named by its actual airmass origin, turning the visual wind-speed heuristic into a falsifiable source attribution.","If the domestic/transboundary proportions survive such tests, they give both Taiwan's local governments and cross-strait negotiators a concrete basis for allocating abatement costs."],"forward_implications":["For NO2 and PM2.5, the policy lever is local: vehicle, industrial, and power-plant controls inside central Taiwan would address the domestic majority of these pollutants.","For SO2, local abatement alone would leave more than half the burden in place, so reducing SO2 requires cooperation with sources across the Taiwan Strait.","For PM10, domestic and transboundary measures carry roughly equal weight, so neither a purely local nor a purely regional strategy is sufficient.","The SO2 peaks recorded at the station downwind of the Taichung thermal power plant are attributed to transboundary flow, implying the plant's local SO2 fingerprint is, by this method, small relative to imported SO2.","If the method is right, the 2010 benchmark agreement (within about six percentage points for NO2, SO2, and O3) indicates the same pipeline can be reused for other pollutants and other years without redesigning the analysis."],"supporting_citations":[{"why":"Frames transboundary pollution as arriving from the Asian continent under the northeast monsoon and supplies the wind-speed classification table used to read the wind roses.","marker":"[4]"},{"why":"Introduces non-negative matrix factorization, the decomposition method the entire analysis is built on.","marker":"[6]"},{"why":"Supplies the multiplicative update rules used to compute the factorization and prove its convergence.","marker":"[7]"},{"why":"Provides the cophenetic-correlation criterion used to select the number of components k for each pollutant.","marker":"[11]"},{"why":"Gives the workflow of decomposing the matrix and interpreting the W and H factors against wind data, which the results section follows.","marker":"[13]"},{"why":"Provides the 2010 domestic pollution ratios for NO2, SO2, and O3 that the paper uses as its independent validation benchmark.","marker":"[14]"}],"fun_headline_variants":["NMF splits Taiwan smog: NO2, PM2.5 local; SO2 imported","Factoring air data reveals local vs imported pollution in Taiwan","Central Taiwan: local sources drive NO2 and PM2.5, SO2 drifts in","Matrix decomposition traces central Taiwan's smog origins","NMF source apportionment: local NO2, PM2.5; transboundary SO2"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a pollutant component's wind speed alone, read visually from a wind rose, cleanly separates 'domestic' from 'transboundary' sources: components peaking near 4 m/s are called domestic and those reaching roughly 7 to 12 m/s are called transboundary, with no quantitative threshold, error margin, or physical transport model behind the split; changing that cutoff directly changes every percentage in the results table.","fun_headline_variants_meta":{"raw":{"variants":["NMF splits Taiwan smog: NO2, PM2.5 local; SO2 imported","Factoring air data reveals local vs imported pollution in Taiwan","Central Taiwan: local sources drive NO2 and PM2.5, SO2 drifts in","Matrix decomposition traces central Taiwan's smog origins","NMF source apportionment: local NO2, PM2.5; transboundary SO2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000588,"raw_usage":{"total_tokens":2732,"prompt_tokens":885,"completion_tokens":1847,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":1742}},"tokens_in":501,"tokens_out":1847,"duration_ms":14365,"temperature":1.0,"reasoning_tokens":1742,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:33:02.121059+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the source-apportionment table with a single explicit wind-speed cutoff — for example the 5.5 m/s boundary of the paper's own wind classification, or 6 m/s — applied uniformly to every NMF component, and rerun the validation using the same component counts as the main analysis: if NO2's domestic share falls below half, or if the validation procedure does not reproduce the main proportions when applied to the full 2008–2017 record, the central claim fails.","supporting_citations":[{"cited_title":"Long- range transport of aerosols and their impact on the air quality of Taiwan","cited_arxiv_id":null,"evidence_quote":"Frames transboundary pollution as arriving from the Asian continent under the northeast monsoon and supplies the wind-speed classification table used to read the wind roses."},{"cited_title":"Learning the parts of objects by nonnegative matrix factorization","cited_arxiv_id":null,"evidence_quote":"Introduces non-negative matrix factorization, the decomposition method the entire analysis is built on."},{"cited_title":"Algorithms for Non-negative Matrix Factorization","cited_arxiv_id":null,"evidence_quote":"Supplies the multiplicative update rules used to compute the factorization and prove its convergence."},{"cited_title":"Metagenes and molecular pattern discovery using matrix factorization","cited_arxiv_id":null,"evidence_quote":"Provides the cophenetic-correlation criterion used to select the number of components k for each pollutant."},{"cited_title":"Trans-boundary air pollution in a city under various atmospheric conditions","cited_arxiv_id":null,"evidence_quote":"Gives the workflow of decomposing the matrix and interpreting the W and H factors against wind data, which the results section follows."},{"cited_title":"Estimation of foreign versus domestic contributions to Taiwan’s air pollution","cited_arxiv_id":null,"evidence_quote":"Provides the 2010 domestic pollution ratios for NO2, SO2, and O3 that the paper uses as its independent validation benchmark."}],"review_version":1}