pith. sign in

arxiv: 1906.08869 · v1 · pith:5BLZMAZRnew · submitted 2019-06-20 · 📡 eess.SP · cs.LG· eess.IV

A data-driven approach to sampling matrix selection for compressive sensing

Elin Farnell , Henry Kvinge , John P. Dixon , Julia R. Dupuis , Michael Kirby , Chris Peterson , Elizabeth C. Schundler , Christian W. Smith This is my paper

Pith reviewed 2026-05-25 19:02 UTC · model grok-4.3

classification 📡 eess.SP cs.LGeess.IV
keywords compressive sensingsampling matrix orderingWalsh-Hadamardchemical detectionadaptive coherence estimatorvariance capturedepth imaging
0
0 comments X

The pith

A data-driven ordering of the Walsh-Hadamard sampling basis by captured variance supports accurate sensing at 90 percent compression.

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

The authors develop a technique to rank sampling vectors according to how much variance each explains in a training dataset. This produces an ordering that lets any prefix of the matrix serve as an effective sensing matrix. Tests on chemical release data show reliable detection with the adaptive coherence estimator even at 90 percent compression. On depth images the method raises peak signal-to-noise ratio by more than 30 percent relative to random or sequency orderings. The approach is intended for bases that hardware can implement directly.

Core claim

Ranking the rows of the Walsh-Hadamard matrix by the variance they capture from training data yields a sampling order that permits meaningful compressive measurements at any desired compression level; when combined with ACE detection this ordering achieves successful chemical detection at a 90 percent compression rate and improves PSNR by over 30 percent on a depth-image test set.

What carries the argument

The variance-ordered Walsh-Hadamard sampling matrix that places rows capturing the most data variance first.

If this is right

  • Chemical detection succeeds at 90 percent compression using the adaptive coherence estimator.
  • Peak signal-to-noise ratio rises by more than 30 percent on depth images compared with conventional orderings.
  • The ordering can be fixed once from representative data and applied at multiple compression ratios.
  • The method works for any fixed sampling basis that satisfies hardware constraints.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If operational data exhibits different variance patterns, the ordering may lose its advantage and require recomputation from new training examples.
  • Hardware implementations could use this ordering to reduce the number of measurements collected without sacrificing detection reliability.
  • The same variance-ranking idea might be combined with other reconstruction or detection algorithms beyond those tested here.

Load-bearing premise

The variance patterns learned from the simulant chemical release and depth image training sets will continue to describe the variance in future data.

What would settle it

A new collection of chemical release or depth image measurements in which the variance-ordered sampling matrix produces worse detection accuracy or lower reconstruction quality than random sampling at the same compression rate.

Figures

Figures reproduced from arXiv: 1906.08869 by Chris Peterson, Christian W. Smith, Elin Farnell, Elizabeth C. Schundler, Henry Kvinge, John P. Dixon, Julia R. Dupuis, Michael Kirby.

Figure 1
Figure 1. Figure 1: Comparison of the number of connected components in reshaped sampling vectors (drawn from rows of a 4096 × 4096 Walsh-Hadamard matrix) for various orders. Top to bottom: standard order, maximal-variance order for the Swiss Ranger dataset, and frequency order. In Figs. 3 and 4, we provide quantitative comparisons of the performance of sampling and reconstruction with the various sampling orders on the Swiss… view at source ↗
Figure 2
Figure 2. Figure 2: Example reconstructions of a test image (top left) sampled by various orders of rows of the shifted Walsh￾Hadamard matrix. The reconstructed images labeled by sampling basis (with corresponding PSNR) are: standard ordering (49.4), random ordering (49.6), sequency order (58.0), frequency order (62.7), and maximal-variance order (62.8). Note that the standard and random sampling orders produce artifacts in t… view at source ↗
Figure 3
Figure 3. Figure 3: A comparison of the quality of reconstruction of depth images from the Swiss Ranger test set after using different sampling orders from the shifted Walsh-Hadamard matrix. Each subplot is a histogram of the PSNR values for reconstructed images sampled with that particular order. The organization of the subplots coincides with that in [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A comparison of the quality of reconstruction of depth images from the Swiss Ranger test set after using different sampling orders from the shifted Walsh-Hadamard matrix. Left: PSNR for each test image. Right: RMSE for each test image. All sampling was done at 25% compression on 64 × 64 depth images. All reconstructions use split Bregman iteration with a Haar wavelet sparsity-promoting basis. We next consi… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of performance of various sampling orderings on Fabry-P´erot GAA data. Each plot shows the number of pixels over the threshold as a function of the frame in the hyperspectral video. From left to right, the plots show detection by: ACE, ACE with bulk coherence, and ACE with bulk coherence and persistence [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Left: ACE on an example reconstruction from Fabry-P´erot GAA (Cube 100). Right: ACE with bulk coherence and persistence on the same reconstructed cube. Note that unwanted artifacts occur in the reconstructions after CS with standard, random, and sequency orders. Furthermore, the corresponding reconstructions make it difficult or impossible to localize the presence of the chemical simulant, especially in th… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of performance of various sampling orderings on Johns Hopkins SF6 data. Each plot shows the number of pixels over the threshold as a function of the frame in the hyperspectral video. From left to right, the plots show detection by: ACE, ACE with bulk coherence, and ACE with bulk coherence and persistence. While every ordering results in chemical detection that matches that seen in the uncompress… view at source ↗
Figure 8
Figure 8. Figure 8: Left: ACE on an example reconstruction from Johns Hopkins SF6 data (Cube 30). Right: ACE, bulk coherence, and persistence on the same reconstructed cube. Unwanted artifacts occur in the reconstructions after CS with standard, random, and sequency orders, and these orders make it impossible to localize the presence of the chemical in the reconstruction. Note that all figures are displayed on a range of [0, … view at source ↗
Figure 9
Figure 9. Figure 9: Left: ACE on an example reconstruction from Johns Hopkins SF6 data (Cube 90). Right: ACE, bulk coherence, and persistence applied to the same reconstructed cubes. Unwanted artifacts occur in the reconstructions sampled with standard, random, and sequency orders, and these orders make it impossible to localize the presence of the chemical in the reconstruction. Note that all figures are displayed on a range… view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of cumulative percent variance explained by successive inclusion of ordered sampling basis vectors in maximal-variance order for data that has and has not been mean-subtracted. Left: Variance explained by sampling vectors on original Swiss-Ranger data. Right: Variance explained by sampling vectors on mean￾subtracted Swiss-Ranger data. Mean-subtraction has little effect on the maximal-variance o… view at source ↗
Figure 11
Figure 11. Figure 11: Examples of Walsh-Hadamard sampling basis vectors that appear at the beginning and at the end of the maximal￾variance order for the Swiss Ranger depth training data, along with corresponding approximate percent variance. Left: First 25 sampling basis vectors. Right: Last 25 sampling basis vectors. In this case, we trained the order using the Walsh-Hadamard matrix rather than the shifted Walsh-Hadamard mat… view at source ↗
read the original abstract

Sampling is a fundamental aspect of any implementation of compressive sensing. Typically, the choice of sampling method is guided by the reconstruction basis. However, this approach can be problematic with respect to certain hardware constraints and is not responsive to domain-specific context. We propose a method for defining an order for a sampling basis that is optimal with respect to capturing variance in data, thus allowing for meaningful sensing at any desired level of compression. We focus on the Walsh-Hadamard sampling basis for its relevance to hardware constraints, but our approach applies to any sampling basis of interest. We illustrate the effectiveness of our method on the Physical Sciences Inc. Fabry-P\'{e}rot interferometer sensor multispectral dataset, the Johns Hopkins Applied Physics Lab FTIR-based longwave infrared sensor hyperspectral dataset, and a Colorado State University Swiss Ranger depth image dataset. The spectral datasets consist of simulant experiments, including releases of chemicals such as GAA and SF6. We combine our sampling and reconstruction with the adaptive coherence estimator (ACE) and bulk coherence for chemical detection and we incorporate an algorithmic threshold for ACE values to determine the presence or absence of a chemical. We compare results across sampling methods in this context. We have successful chemical detection at a compression rate of 90%. For all three datasets, we compare our sampling approach to standard orderings of sampling basis such as random, sequency, and an analog of sequency that we term `frequency.' In one instance, the peak signal to noise ratio was improved by over 30% across a test set of depth images.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes ordering the rows of a Walsh-Hadamard sampling matrix by their empirical variance computed on a training portion of each dataset (chemical simulant releases and depth images). This fixed ordering is then used for compressive sensing at arbitrary compression ratios, followed by reconstruction and ACE-based chemical detection or PSNR evaluation. The central empirical claims are successful ACE detection at 90% compression and a >30% PSNR gain on one depth-image test set relative to random, sequency, and frequency orderings.

Significance. If the variance-derived ordering remains stable under modest distribution shift and the reported gains are statistically reliable, the method offers a simple, hardware-compatible way to adapt compressive sensing to domain-specific data statistics. The work supplies concrete numbers on three real sensor datasets and combines the sampler with an existing detector (ACE), which is a practical strength. However, the absence of error bars, explicit data-split protocols, and out-of-distribution tests leaves the magnitude and robustness of the gains uncertain.

major comments (2)
  1. [Abstract and experimental results] Abstract and experimental results section: the headline claims (90% compression detection success; >30% PSNR lift) are presented without reported error bars, number of Monte-Carlo trials, exact train/test split ratios, or any statistical significance test. Because these numbers constitute the primary evidence that the variance ordering outperforms sequency and random baselines, the lack of quantitative uncertainty measures is load-bearing for the central performance claim.
  2. [Method and evaluation] Method and evaluation sections: the ordering is obtained by ranking Walsh-Hadamard vectors according to their sample variance on the training partition; no experiment evaluates whether an ordering learned on one chemical release or scene type retains its advantage when applied to a different release, background, or sensor. Given that the method is explicitly data-driven, the absence of any cross-dataset or distribution-shift test directly limits the scope of the optimality claim.
minor comments (2)
  1. [Method] Notation for the variance-based ranking and the precise definition of the 'frequency' ordering should be stated explicitly in an equation or algorithm box so that the procedure is reproducible from the text alone.
  2. [Detection pipeline] The manuscript should clarify whether the ACE threshold is fixed across all sampling methods or tuned separately; if tuned separately, that choice affects the fairness of the 90% compression comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the scope and presentation of our results. We address each major comment below and indicate the revisions planned for the manuscript.

read point-by-point responses

  1. Referee: [Abstract and experimental results] Abstract and experimental results section: the headline claims (90% compression detection success; >30% PSNR lift) are presented without reported error bars, number of Monte-Carlo trials, exact train/test split ratios, or any statistical significance test. Because these numbers constitute the primary evidence that the variance ordering outperforms sequency and random baselines, the lack of quantitative uncertainty measures is load-bearing for the central performance claim.

    Authors: We agree that the original manuscript does not report error bars, the number of Monte-Carlo trials, exact train/test split ratios, or statistical significance tests. In the revised manuscript we will add the precise train/test split ratios used for each of the three datasets, state the number of Monte-Carlo trials performed for the ACE detection and PSNR evaluations, include error bars on the reported metrics, and add a statistical significance test (e.g., paired t-test) comparing the variance ordering against the sequency and random baselines. These additions will appear in the experimental results section and will be summarized concisely in the abstract. revision: yes

  2. Referee: [Method and evaluation] Method and evaluation sections: the ordering is obtained by ranking Walsh-Hadamard vectors according to their sample variance on the training partition; no experiment evaluates whether an ordering learned on one chemical release or scene type retains its advantage when applied to a different release, background, or sensor. Given that the method is explicitly data-driven, the absence of any cross-dataset or distribution-shift test directly limits the scope of the optimality claim.

    Authors: The method is intentionally data-driven and produces a dataset-specific ordering; the paper already demonstrates the procedure on three distinct sensor modalities (multispectral chemical simulant data, hyperspectral LWIR data, and depth imagery) with consistent gains relative to standard orderings. We acknowledge that explicit cross-dataset transfer experiments (training the ordering on one modality and testing on another) are absent. In revision we will add a dedicated paragraph in the discussion section that (i) clarifies the intended use case as domain-specific adaptation rather than universal optimality and (ii) explicitly notes the lack of cross-dataset robustness tests as a limitation on the generality of the claims. revision: partial

Circularity Check

0 steps flagged

No significant circularity; ordering is an explicit external computation from training-set variance statistics.

full rationale

The paper defines its sampling order by ranking Walsh-Hadamard vectors according to their empirical variance on the supplied training portion of each dataset (chemical releases or depth images). This step is performed once on external data and then held fixed; it does not reduce by any equation in the paper to a fitted target metric, a self-citation, or an ansatz. Subsequent claims (ACE detection at 90 % compression, PSNR gains) are measured on held-out samples drawn from the identical distributions, but the ordering itself is not tautological with those performance numbers. No load-bearing uniqueness theorem, self-citation chain, or renaming of a known result appears in the derivation. The method is therefore a standard data-driven preprocessing step whose validity rests on the representativeness of the training distribution—an empirical assumption, not a circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on computing an ordering from variance in the three listed datasets; no explicit free parameters or invented entities are named, but the approach implicitly treats training-data variance as the optimality criterion.

free parameters (1)
  • ACE detection threshold
    An algorithmic threshold on ACE values is used to decide chemical presence or absence and is likely chosen or tuned on the data.
axioms (1)
  • domain assumption Variance captured by basis functions in the training data serves as a reliable proxy for reconstruction and detection utility.
    The ordering is defined to be optimal with respect to variance capture.

pith-pipeline@v0.9.0 · 5842 in / 1325 out tokens · 41888 ms · 2026-05-25T19:02:38.022546+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. More chemical detection through less sampling: amplifying chemical signals in hyperspectral data cubes through compressive sensing

    eess.IV 2019-06 unverdicted novelty 6.0

    Compressive sensing reconstruction amplifies chemical signals in hyperspectral cubes, with greater amplification at lower sampling rates, demonstrated on two real chemical simulant datasets using ACE detection.

  2. Total variation vs L1 regularization: a comparison of compressive sensing optimization methods for chemical detection

    eess.IV 2019-06 unverdicted novelty 3.0

    Empirical comparison on two real chemical-release datasets shows L1 regularization yields better ACE-based chemical detection than total variation at 90% compression.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · cited by 2 Pith papers

  1. [1]

    Single-pixel imaging via compressive sampling,

    Duarte, M. F., Davenport, M. A., Takhar, D., Laska, J. N., Sun, T., Kelly, K. F., and Baraniuk, R. G., “Single-pixel imaging via compressive sampling,” IEEE signal processing magazine 25(2), 83–91 (2008)

    work page 2008

  2. [2]

    Sparse MRI: The application of compressed sensing for rapid MR imaging,

    Lustig, M., Donoho, D., and Pauly, J. M., “Sparse MRI: The application of compressed sensing for rapid MR imaging,” Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 58(6), 1182–1195 (2007)

    work page 2007

  3. [3]

    Decoding by linear programming,

    Candes, E. J. and Tao, T., “Decoding by linear programming,” IEEE transactions on information the- ory 51(12), 4203–4215 (2005)

    work page 2005

  4. [4]

    Uncertainty principles and ideal atomic decomposition,

    Donoho, D. L. and Huo, X., “Uncertainty principles and ideal atomic decomposition,” IEEE transactions on information theory 47(7), 2845–2862 (2001)

    work page 2001

  5. [5]

    A new compressive imaging camera architecture using optical-domain compression,

    Takhar, D., Laska, J. N., Wakin, M. B., Duarte, M. F., Baron, D., Sarvotham, S., Kelly, K. F., and Baraniuk, R. G., “A new compressive imaging camera architecture using optical-domain compression,” in [Computational Imaging IV ], 6065, 606509, International Society for Optics and Photonics (2006)

    work page 2006

  6. [6]

    Compressive sensing [lecture notes],

    Baraniuk, R. G., “Compressive sensing [lecture notes],” IEEE signal processing magazine 24(4), 118–121 (2007)

    work page 2007

  7. [7]

    Model-based compressive sensing,

    Baraniuk, R. G., Cevher, V., Duarte, M. F., and Hegde, C., “Model-based compressive sensing,” IEEE Transactions on Information Theory 56(4), 1982–2001 (2010)

    work page 1982

  8. [8]

    AIRIS standoff multispectral sensor,

    Cosofret, B. R., Chang, S., Finson, M. L., Gittins, C. M., Janov, T. E., Konno, D., Marinelli, W. J., Levreault, M. J., and Miyashiro, R. K., “AIRIS standoff multispectral sensor,” in [ Chemical, Biological, Radiological, Nuclear, and Explosives (CBRNE) Sensing X ], 7304, 73040Y, International Society for Optics and Photonics (2009)

    work page 2009

  9. [9]

    A primer for chemical plume detection using LWIR sensors,

    Broadwater, J. B., Limsui, D., and Carr, A. K., “A primer for chemical plume detection using LWIR sensors,” Technical Paper, National Security Technology Department, Las Vegas, NV (2011)

    work page 2011

  10. [10]

    Sparsity and structure in hyper- spectral imaging: Sensing, reconstruction, and target detection,

    Willett, R. M., Duarte, M. F., Davenport, M. A., and Baraniuk, R. G., “Sparsity and structure in hyper- spectral imaging: Sensing, reconstruction, and target detection,” IEEE signal processing magazine 31(1), 116–126 (2014)

    work page 2014

  11. [11]

    The split Bregman method for L1-regularized problems,

    Goldstein, T. and Osher, S., “The split Bregman method for L1-regularized problems,” SIAM J. Imaging Sci. 2(2), 323–343 (2009)

    work page 2009

  12. [12]

    The restricted isometry property and its implications for compressed sensing,

    Candes, E. J., “The restricted isometry property and its implications for compressed sensing,” Comptes rendus mathematique 346(9-10), 589–592 (2008)

    work page 2008

  13. [13]

    Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information,

    Cand` es, E. J., Romberg, J., and Tao, T., “Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information,” IEEE Transactions on information theory 52(2), 489–509 (2006)

    work page 2006

  14. [14]

    Compressed sensing,

    Donoho, D. L., “Compressed sensing,” IEEE Transactions on information theory 52(4), 1289–1306 (2006)

    work page 2006

  15. [15]

    The Johnson-Lindenstrauss lemma meets compressed sensing. 2006,

    Baraniuk, R., Davenport, M., DeVore, R., and Wakin, M., “The Johnson-Lindenstrauss lemma meets compressed sensing. 2006,” Preprint

    work page 2006

  16. [16]

    Existing and emerging image quality metrics,

    Dosselmann, R. and Yang, X. D., “Existing and emerging image quality metrics,” in [ Electrical and Com- puter Engineering, 2005. Canadian Conference on ], 1906–1913, IEEE (2005)

    work page 2005

  17. [17]

    Adaptive matched subspace detectors and adaptive coherence estima- tors,

    Scharf, L. L. and McWhorter, L. T., “Adaptive matched subspace detectors and adaptive coherence estima- tors,” in [ Signals, Systems and Computers, 1996. Conference Record of the Thirtieth Asilomar Conference on], 1114–1117, IEEE (1996)

    work page 1996

  18. [18]

    Adaptive subspace detectors,

    Kraut, S., Scharf, L. L., and McWhorter, L. T., “Adaptive subspace detectors,” IEEE Transactions on signal processing 49(1), 1–16 (2001)

    work page 2001

  19. [19]

    Multipulse adaptive coherence for detection in wind turbine clutter,

    Pakrooh, P., Scharf, L. L., Cheney, M., Homan, A., and Ferrara, M., “Multipulse adaptive coherence for detection in wind turbine clutter,” IEEE Transactions on Aerospace and Electronic Systems 53(6), 3091– 3103 (2017)

    work page 2017

  20. [20]

    The adaptive coherence estimator for detection in wind turbine clutter,

    Pakrooh, P., Scharf, L., Cheney, M., Homan, A., and Ferrara, M., “The adaptive coherence estimator for detection in wind turbine clutter,” in [ Radar Conference (RadarConf), 2017 IEEE ], 1793–1798, IEEE (2017)

    work page 2017

  21. [21]

    Multipulse subspace detectors,

    Scharf, L. L. and Pakrooh, P., “Multipulse subspace detectors,” in [ Conference record-Asilomar Conference on Signals, Systems, & Computers ], (2017)

    work page 2017

  22. [22]

    Total variation vs L1 regularization: a comparison of compressive sensing optimization methods for chemical detection,

    Farnell, E., Kvinge, H., Dupuis, J. R., Kirby, M., Peterson, C., and Schundler, E. C., “Total variation vs L1 regularization: a comparison of compressive sensing optimization methods for chemical detection,” (2019). Preprint

    work page 2019

  23. [23]

    Sylvester, J. J., “LX. thoughts on inverse orthogonal matrices, simultaneous signsuccessions, and tessellated pavements in two or more colours, with applications to Newton’s rule, ornamental tile-work, and the theory of numbers,” The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science 34(232), 461–475 (1867)

    work page

  24. [24]

    Unified matrix treatment of the fast Walsh-Hadamard transform,

    Fino, B. J. and Algazi, V. R., “Unified matrix treatment of the fast Walsh-Hadamard transform,” IEEE Transactions on Computers (11), 1142–1146 (1976)

    work page 1976

  25. [25]

    G., [ Applications of Walsh and related functions: with an introduction to sequency theory ], Academic press (1984)

    Beauchamp, K. G., [ Applications of Walsh and related functions: with an introduction to sequency theory ], Academic press (1984)

    work page 1984

  26. [26]

    Walsh transforms,

    Beer, T., “Walsh transforms,” American Journal of Physics 49(5), 466–472 (1981)

    work page 1981