{"id":"60f3c71f-8e52-46f4-b8a8-988732aca93e","arxiv_id":"1908.02379","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Predictor-based subspace identification is applied to a heated aluminum bar, yielding state-space models with reported validation errors of 1.18 to 5.83 percent, but the main estimation step as written appears mathematically invalid for the stated data length.","lead":"This paper applies predictor-based subspace identification to build state-space models of temperature dynamics on a heated aluminum bar with multiple heaters and thermocouples. The authors report validation errors around one to six percent, and claim that low-order models can capture the dynamics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (20) requires inverting a 264-row regressor Gramian built from at most ~157 data columns; with N=180 and p=24 full row rank is impossible, so the central identification step is invalid as written.","rationale":"The reader's weakest assumption identifies exactly the full-row-rank condition on Z in Eq. (20). I agree that this is the most load-bearing concern: it is a direct dimensional inconsistency, not merely a disagreement with standard practice. With N=180 and p=24, the regressor matrix has 264 rows and at most 157 columns, so it cannot have full row rank and the inverse in Eq. (20) cannot exist. This invalidates the VARX estimation, the AIC-based selection of p, and the subsequent state-sequence and system-matrix estimation. It is stronger than the secondary concern about using validation data for model selection, because it affects the internal derivation before any validation is performed. I also note that the paper provides public code and data and useful experimental characterization, so the empirical claim could in principle be salvageable with a corrected estimator and an independent validation set; however, the manuscript as written does not support its central claim. The final verdict should remain REJECT, consistent with the reader's assessment.","tokens_in":16216,"tokens_out":8230,"duration_ms":97097,"concrete_test":"Using the public identification dataset, form the regressor matrix Z_{0,23}^{(l)} for p=24 and all shifts l compatible with 180 samples; compute rank(Z) and compare it with 264. If rank(Z) < 264, Eq. (20) is undefined. Then re-run the pipeline using a pseudo-inverse or a regularized least-squares step and evaluate the selected model on a third dataset that was not used for model selection; if the 1-6% validation errors do not survive on that untouched dataset, the central claim is unsupported.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim that low-order linear models accurately estimate the temperature dynamics depends on the PBSID estimation chain. The first step, Eq. (20), solves a least-squares problem and writes \\hat{M}_{p-1} = Y Z^T (Z Z^T)^{-1}, explicitly assuming that Z = Z_{0,p-1}^{(l)} has full row rank. The stated identification data length is 180 samples and the selected past window is p=24 for System 1 (Section IV.B). Z has 11p=264 rows. Its columns are shifted regressors z_{k-p,k-1} for k=p,...,p+l, so with N=180 samples there are at most N-p+1=157 columns. A matrix cannot have row rank exceeding its number of columns, so rank(Z) ≤ 157 < 264. The inverse in Eq. (20) therefore does not exist. The same failure invalidates the AIC criterion in Eq. (21), which uses the same least-squares estimate, and it undermines the subsequent construction of \\hat{Q}_{p-1} and the SVD in Eq. (26), so the system matrices in Eq. (34) are not produced by the well-posed estimator claimed in the paper. Section IV.A's statement that the chosen control inputs make Z full rank is inconsistent with these dimensions. This is an internal mathematical failure, not a disagreement with consensus. Even if the public code uses a pseudo-inverse or a regularization, that remedy is not what the paper states, and the VARX step remains underdetermined: 264 regressors per output versus roughly 157 data columns. The reported validation errors are therefore not supported by the derivation as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes an experimental study in which a long aluminum bar is heated by four band heaters and sensed by seven thermocouples, with the goal of identifying a low-order MIMO state-space model of the temperature dynamics. The author combines predictor-based subspace identification (PBSID) with AIC-based order selection and evaluates the resulting models under three validation scenarios: open-loop simulation, simulation driven by measured outputs, and closed-loop simulation. The data and code are publicly available. The paper reports relative validation errors below about 6% for several identified models and concludes that low-order linear state-space representations capture the temperature dynamics with reasonable accuracy.","tokens_in":16558,"tokens_out":3548,"duration_ms":43796,"significance":"If the claims were supported, the paper would provide a useful experimental benchmark for applying PBSID-type methods to slow, spatially distributed thermal systems with limited data, and the public code and data would be a practical asset for the system-identification community. The experimental sections, especially the noise analysis, step-response characterization, and nonlinearity compensation via squared voltage inputs, are thoughtful and well documented. However, the central estimation step is not well posed as written, and the validation protocol makes the reported error rates fitted quantities rather than independent predictive assessments. These issues undermine the main quantitative claims of the paper.","major_comments":[{"comment":"This is a load-bearing issue: the reported identification results and validation errors are not supported by the derivation as written.","section":"§III.B, Eq. (20), and §IV.A/IV.B"},{"comment":"This concern affects all three model-selection methods A, B, and C and the headline error figures.","section":"§III.C1 and Algorithm 1, Step 3"},{"comment":"For System 2, the reported best result (e=4.73%) is obtained with p=26, f=18, and n=13, but the paper gives no detailed prediction plots or eigenvalue analysis for this model. Given that System 2 is used to support the claim that the approach works at a second sampling period, the absence of comparable validation details weakens the empirical support, although this is secondary to the issues above.","section":"§IV.C, paragraphs reporting System 2 results"}],"minor_comments":[{"comment":"The symbol K_{p-1} appears where L_{p-1} is intended: the text says 'where Kp−1∈ R^{n×11p} and Of−1∈ R^{7f×n}', but the preceding definition uses L_{p-1}.","section":"§III.A, after Eq. (12)"},{"comment":"The repository URL contains a typo: 'Temperture' should be 'Temperature'.","section":"Footnote 1"},{"comment":"The vertical axis label 'e' should explicitly refer to the relative error defined in Eq. (39), since the same symbol is used for the innovation process in Eq. (8).","section":"Figure 9(b)"},{"comment":"The notation 'V ARX' and 'VARX' are used inconsistently; one spelling should be adopted throughout.","section":"Throughout"},{"comment":"The definition of identification and validation sequence lengths is slightly ambiguous: the text says '180 and 120 discrete-time samples', while the formal sets I and V use indices k=0,...,N and k=0,...,N_1, so the relationship between N and 180 is not explicit.","section":"§IV.A"}],"recommendation":"reject","confidential_remarks":"The rank condition failure in Eq. (20) is not a minor technicality; with the reported data length and past window, the central least-squares problem is underdetermined and the stated inverse does not exist. Additionally, the validation protocol selects models on the same data used to report errors, so the headline accuracy numbers are not independent estimates. Both issues would require new data collection or a substantially revised experimental and validation design, which is beyond a routine revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me save you time reading the whole thing: this is a useful experimental paper with a fatal flaw in the central estimation step. The author applies PBSID to identify a MIMO thermal model of an aluminum bar with four heaters and seven thermocouples. The experimental work is careful—step response, noise filtering, nonlinearity testing—and the three validation protocols are a nice comparison. The code and data are public, which is a real plus.\n\nThe problem is in equation (20). The paper solves for the VARX parameters by computing (Z Z^T)^{-1} and explicitly assumes Z has full row rank. With the stated identification length of 180 samples and selected past window p=24, Z has 264 rows and at most 157 columns. Full row rank is impossible; the inverse does not exist. This is not a minor typo—the AIC criterion, the state-sequence SVD, and the system matrices all rest on this estimate. The persistency-of-excitation claim in Section IV.A is inconsistent with the dimensions. If the author used a pseudo-inverse or regularized least squares in the code, that's a different method and needs to be stated.\n\nThe second issue is model selection. Methods A, B, and C choose the model that minimizes the validation error on the same validation set. The reported 1-6% relative errors are therefore the best of many models tested on that set, not honest out-of-sample predictions. A separate test set is needed to support the abstract's claim.\n\nThe residual analysis is honest: it shows significant autocorrelation at short lags for some channels, which suggests the 'relatively accurate' summary is optimistic.\n\nThis paper deserves a serious referee because the experimental contribution is real and the flaws are identifiable. But as it stands, the central derivation is invalid and the validation numbers are inflated. If the author can fix the estimation step (e.g., regularize, lower p, or collect longer data) and redo the validation with a held-out test set, it could be a useful publication.","headline":"Careful experimental study, but the identification step cannot work as written—the regressor matrix can't be full row rank with these data sizes—and the validation results are selected, not independent.","tokens_in":17056,"tokens_out":4991,"would_cite":false,"duration_ms":53282,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B30","93E12"],"pacs":[],"model":"deepseek-v4-flash","headline":"Temperature dynamics of a heated aluminum bar can be captured by low-order linear state-space models identified from short data records.","keywords":["system identification","subspace identification","temperature dynamics","heat conduction","MIMO state-space models","Kalman innovation form","model order selection","residual analysis"],"falsifier":"On the published data set, form the regressor $Z_{0,p-1}^{(l)}$ for $N=180$ and $p=24$ and compute its rank; it will be at most 157, which is less than the 264 rows, so equation (20) is not executable as written. The decisive check is whether a correctly implemented estimator—using a smaller $p$ or a pseudoinverse—reproduces the claimed validation errors of about 1 to 6 percent; if it does, the empirical claim stands but the derivation needs revision, and if it does not, the central claim is refuted.","tokens_in":16005,"feed_emoji":"🌡️","tokens_out":8596,"duration_ms":82542,"temperature":0.7,"pith_summary":"This paper tries to show that the temperature dynamics of a physical heat-conduction system—an infinite-dimensional, nonlinear process in principle—can be captured well by low-order linear state-space models estimated from short data records. The author builds a 2-meter aluminum bar heated by four band heaters and sensed by seven thermocouples, then combines predictor-based subspace identification with time-series model-order selection and residual testing. On validation data not used for fitting, the identified models achieve relative errors of roughly 1 to 6 percent depending on the simulation mode, with the best closed-loop model using only six states. If this is right, observer-based temperature modeling for control can avoid high-order PDE discretizations, and short identification experiments may be enough for practical models.","feed_headline":"Aluminum bar temperatures fit low-order models to under 6% error","feed_subtitle":"A subspace method using short data records predicts seven thermocouple readings, with the best observer model using only six states.","key_machinery":"The central object is the predictor-based subspace identification (PBSID) pipeline, specifically the one-step-ahead VARX predictor $y_k \\approx M_{p-1}z_{k-p,k-1}+e_k$, whose Markov-parameter matrix $M_{p-1}$ is estimated by least squares. That matrix is rearranged into $Q_{p-1}$, and a singular value decomposition of $\\hat{Q}_{p-1}Z_{0,p-1}^{(l)}$ yields the state sequence; the system matrices $A$, $B$, $C$, and the Kalman gain $K$ then follow from two further least-squares problems. The machinery also includes an AIC-based choice of the past window $p$, a grid search over state order and future window using validation error, and a white-noise residual test based on autocorrelation bounds $\\pm 2/\\sqrt{N_1}$.","core_discovery":"The central claim is that a seven-output, four-input MIMO temperature process—physically described by the heat equation with nonlinear boundary conditions—can be represented by a linear Kalman innovation state-space model identified from only 180 input-output samples, and that the best model order depends on how the model will be used. For System 1 (96-second sampling), open-loop simulation of the state-space model selects $n=34$ states with a validation error of 3.34 percent; open-loop simulation of the innovation model selects $n=12$ with 5.83 percent; and closed-loop simulation driven by measured outputs selects $n=6$ with 1.18 percent. For System 2 (208-second sampling), the open-loop result is 4.73 percent with $n=13$. The paper interprets these results as evidence that low-order linear models are adequate for temperature prediction, and that using real measurements during simulation reduces the required model order.","pith_inferences":["The stated full-row-rank condition on $Z_{0,p-1}^{(l)}$ in equation (20) cannot hold for the reported $N=180$ and $p=24$: the regressor has 264 rows and at most 157 columns, so the estimates in the paper must have come from a numerically regularized or otherwise repaired version of the equation, and the written algorithm needs correction even if the empirical conclusions survive.","Because the closed-loop model (Method C) uses measured outputs to drive the state, its low order and low error are partly an observer property; a fairer benchmark for open-loop control design is the 34-state open-loop model.","A direct robustness test would be to re-estimate with a smaller past window where full row rank is possible, and to check whether the validation errors stay in the reported range; if they do, the low-order conclusion is not an artifact of the algebraic issue.","The residual correlation in output 2 suggests that a nonlinear or spatially refined model of the local heat transfer near that thermocouple could reduce the remaining error."],"forward_implications":["Temperature control and observer design for this class of mechatronic systems can proceed with low-order linear state-space models rather than high-order PDE discretizations.","Short identification experiments (about 180 samples at 96- to 208-second sampling) are enough to obtain validation errors below 6 percent.","When measured outputs are available during operation, the innovation/observer form requires far fewer states (six for System 1) than a pure open-loop simulator (34 states).","The quadratic actuator nonlinearity can be removed by defining inputs as squared voltages, making a linear model appropriate despite the physically nonlinear heat transfer.","Residual correlations (e.g., channel 2 at short lags) indicate remaining unmodeled dynamics, so the identified models are approximate rather than exact representations."],"supporting_citations":[{"why":"Supplies the least-squares subspace identification framework, the persistence-of-excitation condition, and the VAF validation measure used throughout.","marker":"[18]"},{"why":"Supplies the predictor-based subspace identification method that the paper adapts to the temperature setup.","marker":"[26]"},{"why":"Justifies using a VARX model to estimate the Markov parameters from which the state sequence is recovered.","marker":"[27]"},{"why":"Provides the recursive PBSID formulation that motivates the closed-loop variant used in Method C.","marker":"[28]"},{"why":"Supplies the AIC criterion for VARX order selection and the residual autocorrelation whiteness test.","marker":"[29]"},{"why":"Describes the experimental setup and the preliminary identification results on which this paper builds.","marker":"[30]"}],"fun_headline_variants":["Six-state model tracks seven thermocouples on an aluminum bar","Subspace method yields low-order temperature models from sparse data","Aluminum bar heat dynamics captured with just six states","Six-state innovation model predicts seven thermocouple temperatures","Temperature dynamics from an aluminum bar: low-order models accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the data matrix $Z_{0,p-1}^{(l)}$ in equation (20) has full row rank, so the least-squares inverse written there exists; with $N=180$ identification samples and $p=24$, that matrix has 264 rows and at most 157 columns, so the inverse cannot exist as stated.","fun_headline_variants_meta":{"raw":{"variants":["Six-state model tracks seven thermocouples on an aluminum bar","Subspace method yields low-order temperature models from sparse data","Aluminum bar heat dynamics captured with just six states","Six-state innovation model predicts seven thermocouple temperatures","Temperature dynamics from an aluminum bar: low-order models accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000975,"raw_usage":{"total_tokens":4154,"prompt_tokens":968,"completion_tokens":3186,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":3105}},"tokens_in":584,"tokens_out":3186,"duration_ms":25943,"temperature":1.0,"reasoning_tokens":3105,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:46:14.406092+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the published data set, form the regressor $Z_{0,p-1}^{(l)}$ for $N=180$ and $p=24$ and compute its rank; it will be at most 157, which is less than the 264 rows, so equation (20) is not executable as written. The decisive check is whether a correctly implemented estimator—using a smaller $p$ or a pseudoinverse—reproduces the claimed validation errors of about 1 to 6 percent; if it does, the empirical claim stands but the derivation needs revision, and if it does not, the central claim is refuted.","supporting_citations":[{"cited_title":"Verhaegen and V","cited_arxiv_id":null,"evidence_quote":"Supplies the least-squares subspace identification framework, the persistence-of-excitation condition, and the VAF validation measure used throughout."},{"cited_title":"V ARMAX- based closed-loop subspace model identification,","cited_arxiv_id":null,"evidence_quote":"Supplies the predictor-based subspace identification method that the paper adapts to the temperature setup."},{"cited_title":"Recursive predictor-based subspace identification with application to the real-time closed-loop tracking of flutter,","cited_arxiv_id":null,"evidence_quote":"Provides the recursive PBSID formulation that motivates the closed-loop variant used in Method C."},{"cited_title":"L ¨utkepohl, New Introduction to Multiple Time Series Analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the AIC criterion for VARX order selection and the residual autocorrelation whiteness test."},{"cited_title":"Identi- fication of temperature dynamics using subspace and machine learning techniques,","cited_arxiv_id":null,"evidence_quote":"Describes the experimental setup and the preliminary identification results on which this paper builds."}],"review_version":1}