Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Variance-based variable selection in sensor calibration with strong interferents -- application to air pollution monitoring with a carbon nanotube sensor array

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Minimizing a complexity-penalized prediction variance recovers the environmental variables a pollution sensor truly responds to, and the sparse calibration model that results out-predicts the complete model in new environments.

desk verdict A sensible variance-based selection method with strong simulation support, undermined by an i.i.d. assumption in the outdoor validation that the data clearly violate. read the letter →

arxiv 2507.05001 v1 pith:QBAJZUSG submitted 2025-07-07 stat.AP stat.ME

classification stat.APstat.ME MSC 46N3062-0762F0762F1562F4062J05 PACS 07.07.Df81.16.Be81.16.Fg61.48.De
keywords sensitivityanalysissensorcalibrationinverseproblemexplainabilitycarbonnanotubesmonoxidevariableselectionpredictionvariance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a way to decide which environmental variables — the 'interferents' such as temperature, humidity, and other gases — should enter a pollution sensor's calibration model. Its claim is that minimizing a complexity-penalized prediction variance, in the form of the Bayesian Information Criterion (BIC) applied to the model's expected prediction variance, recovers the variables that truly drive the sensor, and that pruning the rest improves accuracy and generalization when the sensor moves to a new environment. On simulated data with strong input correlations, noise, and an unmeasured confounder, the criterion selects the three genuinely influential variables in almost every bootstrap run while rarely selecting correlated non-influential ones, and the selected model beats the complete model, especially when the deployment environment differs from the calibration environment. On a 28-day outdoor carbon-nanotube sensor array, the method finds that functionalized sensors truly respond to CO and O3, with CO dominant, while unfunctionalized sensors carry about 34% unexplained variance, pointing to unmeasured interferents such as volatile organic compounds. The practical stake is a white-box calibration that states what a sensor actually measures, with a field CO resolution near ten times better than laboratory benchmarks.

What carries the argument

The load-bearing object is the variance decomposition $V(\alpha, f) = m_f^\top C_\beta m_f + \theta^2 + \operatorname{Tr}(C_f C_\beta)$, which splits a sensor model's prediction variance into a finite-data term that shrinks with sample size, a model-error term that shrinks as the model grows, and a trace term that grows with complexity and encodes the robustness cost of overfitting. The method estimates $C_\beta$ and $\theta^2$ by bootstrap resampling of the calibration points, builds candidate polynomial models for every subset of interferents by greedily deleting features with the smallest coefficients, and applies the Bayesian Information Criterion $n\log V + k\log n$ to penalize complexity, so the winning pair $(\alpha, f)$ balances variance against parsimony. A Pareto front over subset size ranks variables by influence, and Proportional Marginal Effects (PME) indices — a variance decomposition valid under dependent inputs — turn the selected model into a quantitative sensitivity breakdown; bootstrapping the whole procedure yields selection frequencies and stability statistics.

What would settle it

In a controlled chamber, expose a functionalized sensor of the same type as deployed outdoors to CO and O3 at concentrations spanning the outdoor ranges but deliberately decorrelated from each other; if the sensor's O3 response (roughly 13% of the selected model's variance) shrinks or vanishes when CO is held fixed, the field attribution of O3 sensitivity is an artifact of outdoor CO–O3 correlation. A complementary computational check would replace the single-point bootstrap with a block bootstrap preserving the time structure of the 28-day series and compare the selected variable sets with the paper's day-split results.

Watch

Extended reading notes

Core claim

The paper's central claim is that variable selection for sensor calibration should minimize not the fitted error but the BIC-penalized expected prediction variance, $$V(\$\alpha$, f) = m_f^\top C_\$\beta$ m_f + \$theta^{2}$ + \operatorname{Tr}(C_f C_\$\beta$),$$ where $\alpha$ indexes the subset of measured interferents, $f$ is the polynomial feature vector, $m_f$ and $C_f$ are the mean and covariance of the features over the calibration distribution, $C_\beta$ is the covariance of the fitted coefficients, and $\theta^2$ is the model error. The three terms separate finite-data estimation uncertainty, irreducible model error, and the robustness cost of over-complexity, so the criterion embodies the bias-variance trade-off before the parsimony penalty is added. Minimizing this criterion selects the causally influential variables: in simulation the three influential interferents are selected in close to 100% of bootstrap runs even at high noise and correlation, the correlated non-influential variables are rarely selected, and the resulting model outperforms the complete model and generalizes significantly better when training and deployment environments differ. On the outdoor data, the method identifies CO and O3 as the real targets of the functionalized sensors, with CO responsible for about half of the explained variance, and attributes a third of the unfunctionalized sensor's variance to unmeasured variables.

Load-bearing premise

The calibration points are treated as exchangeable: uncertainty in the model coefficients is estimated by resampling individual measurement points with replacement, and performance is judged on random 50/50 splits, even though the outdoor dataset is a 28-day autocorrelated time series — the paper's own time-structured splits show clearly degraded predictive performance.

Editorial extensions

If this is right

  • Including every available measurement is a net liability: the selected sparse model matches or beats the complete model on accuracy, credibility-interval length, and interval coverage in every simulation scenario, and its advantage grows when training and deployment environments differ.
  • The functionalized carbon-nanotube sensors can be characterized in the field as genuine CO sensors, with CO explaining about 49% of the selected model's variance and O3 as the next target; relative humidity acts as a proxy that drops out of larger models.
  • About 34% of the unfunctionalized sensor's variance remains unexplained by all measured variables, which the paper interprets as evidence of unmeasured interferents, possibly volatile organic compounds, that degrade generalization.
  • Agnostic selection can replace expert judgement: treating all variables as potential interferents recovers plausible sensor targets (CO, O3) and interferents without prior assumptions about the sensor chemistry.
  • Field CO resolution at the $3\sigma$ threshold is 86–98 ppb over the 7–7.6 ppm range across five functionalized sensors, about ten times better than laboratory literature values, with the caveat that the laboratory studies span much broader concentration ranges.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The three-term variance identity could serve as a live diagnostic beyond selection: a practitioner might decide whether a new measurement channel is worth adding by checking whether the gain comes from lowering $\theta^2$ (real signal) or from inflating $\operatorname{Tr}(C_f C_\beta)$ (overfitting risk).
  • Because target and interferent variables are treated symmetrically, the same machinery could rank which pollutants are recoverable at all from a given non-selective array, yielding an uncertainty budget per pollutant rather than per candidate model.
  • The single-point bootstrap ignores the autocorrelation of the 28-day outdoor series; a block-bootstrap version of the same criterion would test whether the gap between random-split and time-structured performance seen in the paper's Table E.10 is driven by that independence assumption.
  • Outdoor CO and O3 are strongly correlated in the dataset, so the field O3 attribution for the functionalized sensor may be partly inherited from CO; a controlled co-exposure experiment with decorrelated CO and O3 would separate direct from proxy sensitivity, and the PME and resolution tools make that test directly interpretable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a variable selection method for sensor calibration based on minimizing a BIC-penalized prediction variance over polynomial models and variable subsets. The method combines a variance decomposition (Eq. 8) into parameter-uncertainty, model-error, and complexity terms, estimated with a bootstrap, and uses PME indices for interpretability. It is validated on simulated data with known ground truth and applied to an outdoor carbon nanotube sensor array dataset. The central claim is that the selected model identifies truly influential interferents and generalizes better than the complete model, especially in unknown environments.

Significance. If the method holds, it offers a white-box, computationally moderate solution for interferent selection in a domain where existing methods (Sobol, Shapley, Lasso) are limited by noise and correlation. The simulation design is thoughtful, covering noise, correlation, and unmeasured variables, and the analytical variance decomposition in Eq. (8) is a useful contribution. However, the experimental validation is undermined by the i.i.d. assumption on a time series, and the lack of baseline comparison limits the evidence for the method's advantage.

major comments (4)
  1. [4.2.2 and Table E.10] The experimental support for the central generalization claim is contradicted by the time-structured splits reported in Table E.10. For CO prediction with sensor 1H, the random 50/50 split R² of 0.77 (Table 7) drops to 0.60, 0.54, and 0.028 in scenarios 1-3, with the prior R² at -2.81 in scenario 3. These splits are still not chronological train-then-future splits, and the main text (Section 4.2.2) states that the selected model's performance is 'further confirmed' by Table E.10, which is not supported by the numbers. The authors should either present a proper forward-chaining evaluation or substantially qualify the generalization claim.
  2. [3.2 and 4.2] The bootstrap procedure in Section 3.2 resamples n points 'independently and with replacement,' and the evaluation in Section 4.2.1 averages over 100 random 50/50 splits. This assumes the calibration data are i.i.d., but the outdoor dataset is a 28-day time series with strong autocorrelation (Figure E.5). For positively autocorrelated data, the i.i.d. bootstrap underestimates the uncertainty of parameter estimates, and the BIC penalty k(f)log n in Eq. (11) uses n instead of the effective sample size. The paper should either use a block bootstrap, explicitly model the temporal dependence, or state the limitation clearly.
  3. [4.1] The simulation study compares the proposed method only against the simple and complete models. Given the introduction's claim that conventional variable selection methods (Lasso, stepwise, Shapley effects) suffer from major challenges in this setting, the paper should include at least one or two baseline methods in the simulation comparison to demonstrate that the proposed method actually outperforms or matches them. Without such a benchmark, the added value of the method over established alternatives is not established.
  4. [3.3] The 'Causal Relevance' property of the PME indices is stated as a bullet point without proof or a precise citation to a theorem. The property holds for the true model, but the indices in Section 4 are computed on an estimated polynomial model. Non-causal variables can receive nonzero PME due to finite-sample estimation and model misspecification, as evidenced by the selection frequencies of non-influential variables in Table 3 (up to 26%). The interpretation of RH as a 'proxy' for functionalized sensors in Section 4.2.1 should therefore be presented as an empirical observation with this caveat.
minor comments (5)
  1. [Table 3b] The caption of Table 3b states 'for ρ = 5% and σmes = 0.8'; this should read 'for ρ = 0.8 and σmes = 5%.'
  2. [Figure 3] The sentence 'The percentage below the selected model represents the percentage of selections of the model' is ambiguous; please specify which model is meant.
  3. [Appendix D] In Eq. (D.2), the symbol k is reused for the number of polynomial functions in Section 3.2 and for the resolution level; please use a different symbol to avoid confusion.
  4. [Reproducibility] The manuscript does not include a data availability or code availability statement, which would aid reproducibility.
  5. [Section 2] The sentence about 'd × q sensors' and 'd sensors with an output of dimension q' is confusing; consider rewriting for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the variance decomposition is derived analytically and validated on an external simulated ground truth; self-citations to prior Bayesian work are not load-bearing for the selection criterion.

full rationale

The central selection criterion is not equivalent to its own input. Equation (8) is obtained in Appendix A from the law of total variance applied to the linear predictor (6), giving V = m_f^T C_beta m_f + theta^2 + Tr(C_f C_beta). The unknown moments are estimated by a standard bootstrap from D_n (Section 3.2), which is a Monte Carlo estimation step, not a fitted value of the target quantity. The BIC penalty in Eq. (11) is an independent parsimony correction applied to that variance. The simulated validation uses an external ground-truth generator (Eq. 15) containing log, arctangent, and cosine terms that are not in the polynomial family P(p,alpha), so recovering z1,z2,z3 is a genuine empirical check rather than a tautology. The experimental prediction protocol uses held-out random and time-structured splits and compares against a prior that uses only the environmental variables (Table 7 and Table E.10), so the reported improvements are not forced by construction. The posterior inversion formula (Eq. 13) is attributed to the authors' prior work [5] and to [24], but it is re-derived in Appendix B and is used for concentration estimation, not for defining variable relevance; the selection criterion in Section 3.2 does not depend on [5]. The time-structured splits in Table E.10 reveal a real generalization risk caused by temporal dependence in the 28-day outdoor series, but that is a correctness/validity concern about the i.i.d. bootstrap and random-split evaluation, not a circularity, because the degradation is exhibited by an independent evaluation protocol rather than by the construction of the criterion. No equation or fitted parameter is reused as its own prediction, and no load-bearing claim rests on an unverified self-citation.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The method relies on standard statistical assumptions rather than new entities. The main free parameter is the polynomial degree p; the most fragile assumptions are the representativeness of the calibration data and the causal interpretation of PME.

free parameters (1)
  • polynomial order p = 3
    User-chosen hyperparameter; not tuned to data but controls the model family, and the greedy removal explores nested models.
assumptions (6)
  • domain assumption The true calibration function M can be approximated by a polynomial of total degree at most p in x and z.
    Invoked in Section 3.1, Eq. (2). If this fails, the model error theta-squared absorbs misspecification and variable selection can be distorted.
  • domain assumption Measurement and model errors are mutually independent, with model error centered and of variance 1.
    Section 3.1 identifiability assumptions. Required for the variance decomposition in Eq. (8).
  • domain assumption The training data D_n are representative of the random fluctuations of (X_mes, Z_mes_alpha).
    Used to estimate m_f and C_f in Eq. (9). If the deployment environment differs, the selected variables may be inappropriate; the paper partially tests this via rho_train different from rho_test.
  • domain assumption Bootstrap resampling of n individual points from D_n yields valid estimates of C_beta and E[theta-hat^2].
    Bootstrap requires exchangeability of D_n; outdoor time-series data are autocorrelated, which undermines this and is not addressed.
  • domain assumption The posterior PDF of x_star can be approximated by a Gaussian with linearized mean and covariance.
    Appendix B, Eq. (B.2). Standard approximation but adds error to concentration estimates.
  • domain assumption PME indices identify causal relevance even when inputs are correlated.
    Section 3.3. The paper asserts a non-causal variable has PME zero; this is not generally proven for correlated inputs and is used to interpret experimental sensitivities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Variance-based variable selection in sensor calibration with strong interferents -- application to air pollution monitoring with a carbon nanotube sensor array." pith.science (2026). https://pith.science/paper/QBAJZUSG

@misc{pith2026250705001,
  author       = {Pith},
  title        = {Pith review of: Variance-based variable selection in sensor calibration with strong interferents -- application to air pollution monitoring with a carbon nanotube sensor array},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBAJZUSG}},
  note         = {Machine review of arXiv:2507.05001}
}
read the original abstract

Air and water pollution are major threats to public health, highlighting the need for reliable environmental monitoring. Low-cost multisensor systems are promising but suffer from limited selectivity, because their responses are influenced by non-target variables (interferents) such as temperature and humidity. This complicates pollutant detection, especially in data-driven models with noisy, correlated inputs. We propose a method for selecting the most relevant interferents for sensor calibration, balancing performance and cost. Including too many variables can lead to overfitting, while omitting key variables reduces accuracy. Our approach evaluates numerous models using a bias-variance trade-off and variance analysis. The method is first validated on simulated data to assess strengths and limitations, then applied to a carbon nanotube-based sensor array deployed outdoors to characterize its sensitivity to air pollutants.

Figures

Figures reproduced from arXiv: 2507.05001 by the authors.

Figure 1
Figure 1. Left - PME-based variance decomposition of the sensor output computed using the real model, [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. A 10x2 chemiresistor array on a silicon chip (bottom left) fully integrated into a replaceable sensor [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 3
Figure 3. Average Pareto front obtained for the 1A - unfunctionalized CNT - sensor (Top) and 1H - func [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages

  1. [1]

    Society for Industrial and Applied Mathematics, 1996

    Åke Björck.Numerical Methods for Least Squares Problems. Society for Industrial and Applied Mathematics, 1996

  2. [2]

    Feature selection based on the shapley value

    Shay Cohen, Eytan Ruppin, and Gideon Dror. Feature selection based on the shapley value. InProceedings of the 19th IJCAI, page 665–670, San Francisco, CA, USA, 2005. Morgan Kaufmann Publishers Inc

  3. [3]

    SIAM, 2021

    Sébastien Da Veiga, Fabrice Gamboa, Bertrand Iooss, and Clémentine Prieur.Basics and trends in sensitivity analysis: Theory and practice in R. SIAM, 2021

  4. [4]

    In situ calibration algorithms for environmental sensor networks: A review.IEEE Sensors Journal, 19(15):5968–5978, 2019

    Florentin Delaine, Bérengère Lebental, and Hervé Rivano. In situ calibration algorithms for environmental sensor networks: A review.IEEE Sensors Journal, 19(15):5968–5978, 2019

  5. [5]

    Optimizing sensor calibra- 27 tion in open environments: A bayesian approach for non-specific multisensory systems

    Marine Dumon, Bérengère Lebental, and Guillaume Perrin. Optimizing sensor calibra- 27 tion in open environments: A bayesian approach for non-specific multisensory systems. The SMAI Journal of Computational Mathematics, 10:305–324, 2024

  6. [6]

    Joint outdoor ozone and carbon monoxide prediction with a carbon nanotube sensor array calibrated using a bayesian framework

    Marine Dumon, Guillaume Perrin, and Bérengère Lebental. Joint outdoor ozone and carbon monoxide prediction with a carbon nanotube sensor array calibrated using a bayesian framework. In2024 IEEE SENSORS, pages 1–4, 2024

  7. [7]

    B. Efron. Bootstrap Methods: Another Look at the Jackknife.The Annals of Statistics, 7(1):1 – 26, 1979

  8. [8]

    Doris Entner and Patrik O. Hoyer. On causal discovery from time series data using fci. In Proceedings of the 5th European Workshop on Probabilistic Graphical Models, pages 121–128, Finland, 2010. HIIT

Show all 28 references
  1. [9]

    Efficient input variable selection for soft-senor de- sign based on nearest correlation spectral clustering and group lasso.ISA Transactions, 58:367–379, 2015

    Koichi Fujiwara and Manabu Kano. Efficient input variable selection for soft-senor de- sign based on nearest correlation spectral clustering and group lasso.ISA Transactions, 58:367–379, 2015

  2. [10]

    Pc algorithm for nonparanormal graphical models

    Naftali Harris and Mathias Drton. Pc algorithm for nonparanormal graphical models. Journal of Machine Learning Research, 14(105):3365–3383, 2013

  3. [11]

    Propor- tional marginal effects for global sensitivity analysis.SIAM/ASA Journal on Uncer- tainty Quantification, 12(2):667–692, 2024

    Margot Herin, Marouane Il Idrissi, Vincent Chabridon, and Bertrand Iooss. Propor- tional marginal effects for global sensitivity analysis.SIAM/ASA Journal on Uncer- tainty Quantification, 12(2):667–692, 2024

  4. [12]

    Shapley effects for sensitivity analysis with cor- related inputs: comparisons with sobol’ indices, numerical estimation and applications

    Bertrand Iooss and Clémentine Prieur. Shapley effects for sensitivity analysis with cor- related inputs: comparisons with sobol’ indices, numerical estimation and applications. International Journal for Uncertainty Quantification, 9(5), 2019

  5. [13]

    Springer Texts in Statistics

    Gareth James, Daniela Witten, Trevor Hastie, and Robert Tibshirani.An Introduction to Statistical Learning: with Applications in R. Springer Texts in Statistics. Springer, 2nd edition, 2021. 28

  6. [14]

    Kelly and J.C

    F.J. Kelly and J.C. Fussell. Air pollution and public health: emerging hazards and improved understanding of risk.Environmental Geochemistry and Health, 37:631–649, 2015

  7. [15]

    J. M . Marin and C. P. Robert.Bayesian core. Springer-Verlag, New York, 2007

  8. [16]

    McDonald

    Gary C. McDonald. Ridge regression. WIREs Computational Statistics, 1(1):93–100, 2009

  9. [17]

    Nonlinear Multiobjective Optimization, volume 12 of International Series in Operations Research and Management Science

    Kaisa Miettinen. Nonlinear Multiobjective Optimization, volume 12 of International Series in Operations Research and Management Science. Springer, 1999

  10. [18]

    Gas sensing technologies–status, trends, perspectives and novel applications

    Angelo Milone, Anna Grazia Monteduro, Silvia Rizzato, Angelo Leo, and Giuseppe Maruccio. Gas sensing technologies–status, trends, perspectives and novel applications. arXiv preprint arXiv:2110.13637, 2021

  11. [19]

    Lasso: A feature selection technique in predictive modeling for machine learning

    R Muthukrishnan and R Rohini. Lasso: A feature selection technique in predictive modeling for machine learning. In2016 IEEE ICACA, pages 18–20, 2016

  12. [20]

    Comparison of calibration strategies for a high sensitivity pei-based rf humidity sensor.IEEE Sensors Journal, 2024

    Bernard Bobby Ngoune, Marine Dumon, Balakumara Vignesh, Benoit Bondu, Senthilmurugan Subbiah, Guillaume Perrin, Stéphane Bila, Corinne Dejous, Berengère Lebental, and Hamida Hallil. Comparison of calibration strategies for a high sensitivity pei-based rf humidity sensor.IEEE S...

  13. [21]

    Identifying the limitation of stepwise selection for variable selection in regression anal- ysis

    Akinwande Michael Olusegun, Hussaini Garba Dikko, and Shehu Usman Gulumbe. Identifying the limitation of stepwise selection for variable selection in regression anal- ysis. American Journal of Theoretical and Applied Statistics, 4:414, 2015

  14. [22]

    A systematic review on the detection and monitoring of toxic gases using carbon nanotube-based biosensors.Sensing and Bio-Sensing Research, 34:100463, 2021

    Robert Birundu Onyancha, Kingsley Eghonghon Ukhurebor, Uyiosa Osagie Aigbe, Otolorin Adelaja Osibote, Heri Septya Kusuma, Handoko Darmokoesoemo, and Vin- cent Aizebeoje Balogun. A systematic review on the detection and monitoring of toxic gases using carbon nanotube-based bios...

  15. [23]

    Public health hazards due to unsafe drinking water.Air and Water Borne Diseases, 7, 05 2018

    Mahendra Pal, Yodit Ayele, Angesom Hadush, Sumitra Panigrahi, and Vijay Jadhav. Public health hazards due to unsafe drinking water.Air and Water Borne Diseases, 7, 05 2018

  16. [24]

    Perrin, C

    G. Perrin, C. Soize, and N. Ouhbi. Data-driven kernel representations for sampling with an unknown block dependence structure under correlation constraints.Journal of Computational Statistics and Data Analysis, 119:139–154, 2018

  17. [25]

    M. B. Priestley.Spectral Analysis and Time Series. Academic Press, 1981

  18. [26]

    Globalsensitivityindicesfornonlinearmathematicalmodelsandtheirmonte carlo estimates

    I.MSobol’. Globalsensitivityindicesfornonlinearmathematicalmodelsandtheirmonte carlo estimates. Mathematics and Computers in Simulation, 55(1):271–280, 2001

  19. [27]

    Global sensitivity analysis with dependence measures.Journal of Statistical Computation and Simulation, 85(7):1283–1305, 2015

    Sebastien Da Veiga. Global sensitivity analysis with dependence measures.Journal of Statistical Computation and Simulation, 85(7):1283–1305, 2015

  20. [28]

    Highly sensitive room temperature carbon monoxide detection using sno2 nanoparticle-decorated semiconducting single-walled carbon nanotubes

    Yang Zhang, Shumao Cui, Jingbo Chang, Leonidas E Ocola, and Junhong Chen. Highly sensitive room temperature carbon monoxide detection using sno2 nanoparticle-decorated semiconducting single-walled carbon nanotubes. Nanotechnol- ogy, 24(2):025503, dec 2012. Appendix A. Approxim...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.