arxiv: 2604.05211 · v2 · submitted 2026-04-06 · 🧮 math.NA · cs.NA

Recognition: 2 theorem links

· Lean Theorem

Learned Dictionaries with Total Variation and Non-Negativity for Single-Cell Microscopy: Convergence Theory and Deterministic Multi-Channel Cell Feature Unification

Erdem Altuntac

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:39 UTC · model grok-4.3

classification 🧮 math.NA cs.NA

keywords dictionary learningtotal variation regularizationnon-negativity constraintPDHG algorithmsingle-cell microscopymulti-channel feature unificationvariational source condition

0 comments

The pith

A variational dictionary learning method with total variation and non-negativity converges under an explicit step-size bound and unifies multi-channel cell features from microscopy data.

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

The paper develops a dictionary learning algorithm that couples least-squares data fidelity with total-variation regularization and a non-negativity constraint, subject to a unitary constraint on the dictionary, to obtain edge-preserving and physically interpretable reconstructions of single-cell signals. It solves the resulting optimization problem by an alternating proximal-gradient scheme based on PDHG and proves that the iterates converge to the regularized minimizer whenever the product of the primal and dual step sizes satisfies tau sigma less than 1/8. Under the further assumption of a variational source condition, the method recovers the true solution at the optimal rate O(delta) when the regularization parameter is set proportional to the noise level delta. The framework is then used to learn separate unitary dictionaries for each of five imaging channels and to form a single channel-agnostic cell descriptor by concatenating the corresponding sparse codes.

Core claim

The paper proves that the PDHG algorithm for the constrained dictionary learning problem with total-variation and non-negativity penalties converges to the regularized minimizer under the explicit step-size condition tau sigma less than 1/8. It further shows that, when a variational source condition holds for the true solution, the reconstruction error decays as O(delta) for noise level delta by choosing the regularization parameter lambda proportional to delta. On the BSCCM dataset the same construction yields per-channel unitary dictionaries whose sparse codes, when concatenated, produce a deterministic cell descriptor that achieves reconstruction fidelities of 97.06 to 97.54 percent on D1

What carries the argument

The hybrid variational cost functional that combines least-squares fidelity, total-variation regularization, non-negativity, and a unitary dictionary constraint, minimized by the PDHG proximal scheme.

If this is right

The deterministic optimization produces bit-identical iterates across independent runs.
Channel-specific dictionaries adapt to the distinct optical physics of each imaging modality.
Concatenated sparse codes yield a reproducible, channel-agnostic descriptor suitable for downstream biological analysis.
Unsupervised lymphoid-versus-myeloid separation is obtained with ARI equal to 0.575 on the test data.

Where Pith is reading between the lines

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

The deterministic pipeline may simplify auditability requirements in clinical cell-imaging applications.
The same convergence guarantees could be tested on other linear inverse problems that employ total-variation and non-negativity penalties.
Scaling the per-channel dictionary construction to larger cell cohorts would provide a direct check of computational stability.

Load-bearing premise

The variational source condition is assumed to hold for the underlying true solution.

What would settle it

A numerical test in which the reconstruction error fails to scale as O(delta) for successively smaller noise levels delta when lambda is chosen proportional to delta, or in which the PDHG iterates diverge for any step-size product below but close to 1/8.

Figures

Figures reproduced from arXiv: 2604.05211 by Erdem Altuntac.

**Figure 2.** Figure 2: Convergence curves for the per-channel dictionary learning run on BSCCM-tiny ( [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗

**Figure 3.** Figure 3: Reconstruction results for cell #30 from BSCCM-tiny ( [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗

**Figure 4.** Figure 4: Per-channel cell reconstructions across five representative cells (rows) and all five [PITH_FULL_IMAGE:figures/full_fig_p043_4.png] view at source ↗

**Figure 5.** Figure 5: Unified single-cell representation for cell #344, [PITH_FULL_IMAGE:figures/full_fig_p044_5.png] view at source ↗

**Figure 6.** Figure 6: Class label distribution for the labeled subset of BSCCM-tiny ( [PITH_FULL_IMAGE:figures/full_fig_p044_6.png] view at source ↗

**Figure 7.** Figure 7: Labeled cells - channel: DPC Left. Rows alternate between original patch (with [PITH_FULL_IMAGE:figures/full_fig_p045_7.png] view at source ↗

**Figure 8.** Figure 8: Labeled cells - channel: DPC Right. 45 [PITH_FULL_IMAGE:figures/full_fig_p045_8.png] view at source ↗

**Figure 9.** Figure 9: Labeled cells - channel: DPC Top [PITH_FULL_IMAGE:figures/full_fig_p046_9.png] view at source ↗

**Figure 10.** Figure 10: Labeled cells - channel: DPC Bottom. 46 [PITH_FULL_IMAGE:figures/full_fig_p046_10.png] view at source ↗

**Figure 11.** Figure 11: Labeled cells - channel: Brightfield [PITH_FULL_IMAGE:figures/full_fig_p047_11.png] view at source ↗

**Figure 12.** Figure 12: Biological validation metrics on the BSCCM-tiny labelled subset. Leftmost panel: [PITH_FULL_IMAGE:figures/full_fig_p047_12.png] view at source ↗

**Figure 13.** Figure 13: Brightfield ground truth (top row in each panel) versus unified reconstruction [PITH_FULL_IMAGE:figures/full_fig_p048_13.png] view at source ↗

**Figure 14.** Figure 14: DPC Left ground truth versus unified reconstruction [PITH_FULL_IMAGE:figures/full_fig_p048_14.png] view at source ↗

**Figure 15.** Figure 15: DPC Right ground truth versus unified reconstruction [PITH_FULL_IMAGE:figures/full_fig_p049_15.png] view at source ↗

**Figure 16.** Figure 16: DPC Top ground truth versus unified reconstruction [PITH_FULL_IMAGE:figures/full_fig_p049_16.png] view at source ↗

**Figure 17.** Figure 17: DPC Bottom ground truth versus unified reconstruction [PITH_FULL_IMAGE:figures/full_fig_p050_17.png] view at source ↗

read the original abstract

We introduce a variational dictionary learning algorithm with hybrid penalization for single-cell microscopy signals. The cost functional couples least-squares data fidelity with total-variation (TV) regularization and a non-negativity constraint, promoting edge-preserving, physically meaningful reconstructions. The learning task is formulated with an explicit unitary constraint on the dictionary, ensuring well-conditioned representations. The optimization is solved by an alternating proximal-gradient scheme; we prove PDHG iterates converge to the regularized minimizer under an explicit step-size condition (tau*sigma < 1/8), and that under a variational source condition (VSC) the regularized solution converges to the true solution at the optimal O(delta) rate with lambda proportional to delta. Beyond reconstruction, we address multi-channel cell feature unification: given five imaging channels of the BSCCM dataset (DPC Left, Right, Top, Bottom, Brightfield), we learn a family of per-channel unitary dictionaries, each adapted to its channel's optical physics, and concatenate the per-channel sparse codes into a single channel-agnostic cell descriptor. This deterministic approach is mathematically transparent, reproducible, and compatible with clinical AI auditability requirements. On BSCCM-tiny (N=1000 cells, K=512 atoms) the framework reaches reconstruction fidelities of 97.06-97.54% on DPC channels and 94.79% on Brightfield, with bit-identical iterates across runs. Biological validation yields unsupervised lymphoid-vs-myeloid separation at ARI=0.575, NMI=0.471 (permutation p<0.0001).

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.

Desk Editor's Note private letter to a colleague

The paper gives a clean PDHG scheme for dictionary learning under TV plus non-negativity with an explicit step-size bound, plus a deterministic way to fuse multi-channel cell codes, but the O(delta) rate still rests on an unverified variational source condition.

read the letter

The core contribution is an alternating proximal-gradient method for learning unitary dictionaries on single-cell signals, where the objective combines least-squares fidelity with total variation and a non-negativity constraint. They prove the PDHG iterates reach the regularized minimizer whenever tau sigma stays below 1/8, and they state the standard source-condition argument that yields an O(delta) rate to the ground truth when lambda scales with delta. They then learn separate dictionaries per imaging channel on the BSCCM data and concatenate the resulting sparse codes into one descriptor for downstream tasks. On the tiny subset they report reconstruction fidelities around 97 percent on the DPC channels and 94.8 percent on brightfield, with bit-identical runs and an unsupervised lymphoid-myeloid split at ARI 0.575. The deterministic concatenation step is straightforward and avoids some of the opacity that comes with end-to-end learned features. The explicit step-size restriction is also practical for anyone who wants to implement the scheme without tuning heuristics. The main soft spot is the rate claim. The O(delta) guarantee is conditional on the variational source condition holding for the true cell images under this particular regularizer, yet the work supplies neither an analytic argument nor a numerical check that the condition is satisfied on the microscopy signals. If the source condition fails, the rate falls back to the generic slower bound. The reported numbers also lack error bars and direct comparisons to non-dictionary baselines, so it is hard to quantify how much the learned dictionaries actually improve over simpler TV reconstruction. This is useful reading for people who work on variational methods in imaging or who need reproducible feature pipelines for single-cell data. The math is grounded on standard tools and the application is concrete enough that the paper should go to peer review rather than a desk reject.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a variational dictionary learning algorithm for single-cell microscopy that combines least-squares fidelity with total-variation regularization and non-negativity, subject to an explicit unitary constraint on the dictionary. Optimization is performed by an alternating proximal-gradient (PDHG) scheme. The authors prove that the iterates converge to the regularized minimizer under the step-size restriction τσ < 1/8 and, under an additional variational source condition (VSC), that the regularized solution converges to the ground truth at the optimal rate O(δ) when λ ∝ δ. The framework is then applied to the BSCCM dataset to learn per-channel unitary dictionaries and concatenate the resulting sparse codes into a single cell descriptor, yielding reported reconstruction fidelities of 97.06–97.54 % (DPC channels) and 94.79 % (Brightfield) together with unsupervised lymphoid-vs-myeloid separation at ARI = 0.575.

Significance. If the stated convergence results hold, the work supplies an explicit, verifiable step-size condition and a standard source-condition argument for optimal regularization rates in a dictionary-learning setting. The deterministic, channel-agnostic feature unification is reproducible and transparent, which is a practical strength for downstream clinical use. The provision of bit-identical iterates across runs further supports reproducibility.

major comments (2)

[Abstract] Abstract (convergence claims): the O(δ) rate is derived under the variational source condition (VSC) on the true solution with respect to the TV+non-negativity regularizer. No analytic argument or numerical diagnostic is supplied showing that the VSC holds for the BSCCM cell images; if the condition is violated the rate guarantee reduces to the generic O(√δ) bound and the headline optimality claim does not follow.
[Abstract] Abstract (experimental claims): reconstruction fidelities and the ARI = 0.575 are reported without error bars, without comparison to standard dictionary-learning or feature-unification baselines, and without a statement of how the unitary constraint is enforced in the PDHG scheme. These omissions make it impossible to assess whether the numerical results support the claimed advantage of the method.

minor comments (1)

[Abstract] The manuscript states that the PDHG scheme handles the unitary constraint, but the precise projection or alternating step used to enforce it is not described in sufficient detail for independent verification of the convergence proof.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough review and valuable feedback on our manuscript. We address each major comment below with proposed revisions that strengthen clarity and transparency without altering the core contributions.

read point-by-point responses

Referee: [Abstract] Abstract (convergence claims): the O(δ) rate is derived under the variational source condition (VSC) on the true solution with respect to the TV+non-negativity regularizer. No analytic argument or numerical diagnostic is supplied showing that the VSC holds for the BSCCM cell images; if the condition is violated the rate guarantee reduces to the generic O(√δ) bound and the headline optimality claim does not follow.

Authors: We appreciate this observation. The O(δ) rate is explicitly conditional on the variational source condition (VSC) holding for the ground-truth image with respect to the TV+non-negativity regularizer, as stated in the theorem. The manuscript does not claim or prove that VSC is satisfied by the BSCCM data; it presents the standard source-condition argument for optimal rates when the assumption holds. In revision we will (i) rephrase the abstract to emphasize the conditional nature of the rate and (ii) add a short paragraph in the theory section discussing practical numerical diagnostics for VSC (e.g., checking the source-condition residual on representative patches). These changes improve transparency while preserving the theoretical result. revision: partial
Referee: [Abstract] Abstract (experimental claims): reconstruction fidelities and the ARI = 0.575 are reported without error bars, without comparison to standard dictionary-learning or feature-unification baselines, and without a statement of how the unitary constraint is enforced in the PDHG scheme. These omissions make it impossible to assess whether the numerical results support the claimed advantage of the method.

Authors: We agree that these details aid assessment. In the revised manuscript we will: (1) report means and standard deviations for reconstruction fidelities and ARI computed over multiple independent runs (the algorithm is deterministic for fixed initialization, so variability arises only from random dictionary initialization); (2) add brief comparisons in the experiments section to K-SVD (without TV/non-negativity) and to simple concatenation of per-channel PCA codes; (3) insert an explicit sentence describing enforcement of the unitary constraint via orthogonal projection onto the Stiefel manifold after each dictionary update within the alternating PDHG scheme. These additions will be summarized concisely in the abstract as well. revision: yes

Circularity Check

0 steps flagged

No circularity: standard convergence proofs and conditional rates rest on external theory and dataset

full rationale

The claimed PDHG convergence to the regularized minimizer under tau*sigma < 1/8 follows from the standard analysis of the primal-dual hybrid gradient algorithm for convex problems and does not reduce to any quantity fitted from the BSCCM data or defined by the learned dictionaries. The O(delta) rate under a variational source condition is the textbook result from regularization theory, invoked conditionally without any derivation that the true cell signals satisfy the VSC for TV+nonnegativity or any reduction of the rate to a data-dependent fit. The multi-channel unification step is a deterministic concatenation of per-channel sparse codes; reported fidelities and ARI/NMI values are empirical outcomes on an external dataset and are not shown to be equivalent to the inputs by construction. No self-definitional loops, fitted-input predictions, load-bearing self-citations, uniqueness theorems imported from the authors, ansatz smuggling, or renaming of known results appear in the derivation chain.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claims rest on the variational source condition for the rate result and on standard convexity/qualification assumptions needed for proximal-gradient convergence; no new physical entities are postulated.

free parameters (2)

lambda
Regularization parameter set proportional to noise level delta
tau and sigma
Step sizes required to satisfy tau*sigma < 1/8

axioms (2)

domain assumption Variational source condition holds for the true solution
Invoked to obtain the O(delta) convergence rate
standard math The data fidelity term is convex and the regularizers are proper convex
Required for proximal-gradient convergence theory

pith-pipeline@v0.9.0 · 5601 in / 1478 out tokens · 45119 ms · 2026-05-10T18:39:12.553695+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking (D=3 forcing) echoes
τσ < 1/∥K∥₂ ≤ 1/8 ... ∥∇∥₂ ≤8 (Lemma 3, Theorem 4)
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean absolute_floor_iff_bare_distinguishability unclear
under a variational source condition (VSC) the regularized solution converges ... at the optimal O(δ) rate

Reference graph

Works this paper leans on

25 extracted references · 22 canonical work pages

[1]

Altunta¸ c

E. Altunta¸ c. Variational dictionary learning with hybridℓ1 and non-negativity penalization for single-cell microscopy. Zenodo preprint, 2026. DOI: 10.5281/zenodo.18735456

work page doi:10.5281/zenodo.18735456 2026
[2]

Altunta¸ c.New Pair of Primal-Dual Algorithms for Bregman Iterated Variational Reg- ularization

E. Altunta¸ c.New Pair of Primal-Dual Algorithms for Bregman Iterated Variational Reg- ularization. arXiv preprint arXiv:1903.07392, 2019

work page arXiv 1903
[3]

Altunta¸ c

E. Altunta¸ c. Choice of the parameters in a primal-dual algorithm for Bregman iterated variational regularization.Numerical Algorithms, 2020. DOI: 10.1007/s11075-020-00909-6

work page doi:10.1007/s11075-020-00909-6 2020
[4]

Giovanneschi, A

F. Giovanneschi, A. Nittur Ramesh, M. A. Gonzalez Huici, and E. Altunta¸ c. Convolutional sparse coding and dictionary learning for LiDAR depth completion in automotive scenar- ios. In2023 Photonics & Electromagnetics Research Symposium (PIERS), Prague, Czech Republic, July 2023. DOI: 10.1109/PIERS59004.2023.10221515

work page doi:10.1109/piers59004.2023.10221515 2023
[5]

Cwalina, C

S. Cwalina, C. Kottke, V. Jungnickel, R. Freund, P. Runge, P. Rustige, T. Knieling, S. Gu-Stoppel, J. Albers, N. Laske, F. Senger, L. Wen, F. Giovanneschi, E. Altunta¸ c, A. N. Ramesh, M. A. Gonzalez Huici, A. Kuter, and S. Reddy. Fiber-based frequency modulated LiDAR with MEMS scanning capability for long-range sensing in automo- tive applications. In202...

work page doi:10.1109/metroautomotive50197.2021.9502868 2021
[6]

Altunta¸ c, X

E. Altunta¸ c, X. Hu, B. A. Emery, S. Khanzada, G. Kempermann, and H. Amin. Bottom-up neurogenic-inspired computational model. In2023 IEEE BioSensors Conference (BioSen- sors), London, UK, July 2023. DOI: 10.1109/BioSensors58001.2023.10280794

work page doi:10.1109/biosensors58001.2023.10280794 2023
[7]

Pinkard, C

H. Pinkard, C. Liu, F. Nyatigo, D. A. Fletcher, and L. Waller. The Berkeley Sin- gle Cell Computational Microscopy (BSCCM) dataset. arXiv preprint arXiv:2402.06191,

work page arXiv
[8]

Dataset DOI (Dryad): 10.5061/dryad.sxksn038s

Project page:https://waller-lab.github.io/BSCCM/. Dataset DOI (Dryad): 10.5061/dryad.sxksn038s

work page doi:10.5061/dryad.sxksn038s
[9]

Chen and I

Y. Chen and I. Loris. On the choice of parameters in primal-dual splitting methods. Numerical Algorithms, 79:889-909, 2018. DOI: 10.1007/s11075-018-0616-x. 40

work page doi:10.1007/s11075-018-0616-x 2018
[10]

Chambolle and T

A. Chambolle and T. Pock. A first-order primal-dual algorithm for convex problems with applications to imaging.J Math Imaging Vis, 40(1):120-145, 2011. DOI: 10.1007/s10851- 010-0251-1

work page doi:10.1007/s10851- 2011
[11]

L. Condat. A primal-dual splitting method for convex optimization involving Lipschitzian, proximable and linear composite terms.J Optim Theory Appl, 158:460-479, 2013. DOI: 10.1007/s10957-012-0245-9

work page doi:10.1007/s10957-012-0245-9 2013
[12]

Mairal, F

J. Mairal, F. Bach, J. Ponce, and G. Sapiro. Online learning for matrix factorization and sparse coding.Journal of Machine Learning Research, 11:19-60, 2010

2010
[13]

2010 , edition =

M. Elad.Sparse and Redundant Representations: From Theory to Applications in Signal and Image Processing. Springer, 2010. DOI: 10.1007/978-1-4419-7011-4

work page doi:10.1007/978-1-4419-7011-4 2010
[14]

Incorporating the image formation process into deep learning improves network performance

R. Lopez, J. Regier, M. B. Cole, M. I. Jordan, and N. Yosef. Deep generative modeling for single-cell transcriptomics.Nat Methods, 15(12):1053-1058, 2018. DOI: 10.1038/s41592- 018-0229-2. PMC: PMC6289068

work page doi:10.1038/s41592- 2018
[15]

Gayoso, Z

A. Gayoso, Z. Steier, R. Lopez, J. Regier, K. L. Nazor, A. Streets, and N. Yosef. Joint probabilistic modeling of single-cell multi-omic data with totalVI.Nat Methods, 18:272-282,
[16]

DOI: 10.1038/s41592-020-01050-x

work page doi:10.1038/s41592-020-01050-x
[17]

Gayoso, R

A. Gayoso, R. Lopez, G. Xing, P. Boyeau, V. Valiollah Pour Amiri, J. Hong, K. Wu, M. Jayasuriya, E. Mehlman, M. Langevin, Y. Liu, J. Samaran, G. Misrachi, A. Nazaret, O. Clivio, C. Xu, T. Ashuach, M. Lotfollahi, V. Svensson, E. Beltrame, V. Kleshchevnikov, C. Talavera-Lopez, L. Pachter, F. J. Theis, A. Streets, M. I. Jordan, J. Regier, and N. Yosef. A Pyt...

work page doi:10.1038/s41587-021-01206-w 2022
[18]

Moshkov, M

N. Moshkov, M. Bornholdt, S. Benoit, M. Smith, C. McQuin, A. Goodman, R. A. Senft, Y. Han, M. Babadi, P. Horvath, B. A. Cimini, A. E. Carpenter, S. Singh, and J. C. Caicedo. Learning representations for image-based profiling of perturbations.Nat Commun, 15:1594,
[19]

DOI: 10.1038/s41467-024-45999-1

work page doi:10.1038/s41467-024-45999-1
[20]

Burgess, J

J. Burgess, J. J. Nirschl, M.-C. Zanellati, A. Lozano, S. Cohen, and S. Yeung-Levy. Orientation-invariant autoencoders learn robust representations for shape profiling of cells and organelles.Nat Commun, 15:1022, 2024. DOI: 10.1038/s41467-024-45362-4

work page doi:10.1038/s41467-024-45362-4 2024
[21]

Donhauser, K

K. Donhauser, K. Ulicna, G. E. Moran, A. Ravuri, K. Kenyon-Dean, C. Eastwood, and J. Hartford. Towards scientific discovery with dictionary learning: extracting biological concepts from microscopy foundation models.arXiv preprint arXiv:2412.16247, 2024. URL: https://arxiv.org/abs/2412.16247

work page arXiv 2024
[22]

Acevedo, S

A. Acevedo, S. Alf´ erez, A. Merino, L. Puigv´ ı, and J. Rodellar. Recognition of peripheral blood cell images using convolutional neural networks.Comput Methods Programs Biomed, 180:105020, 2019. DOI: 10.1016/j.cmpb.2019.105020. PMID: 31425939

work page doi:10.1016/j.cmpb.2019.105020 2019
[23]

Matek, S

C. Matek, S. Schwarz, K. Spiekermann, and C. Marr. Human-level recognition of blast cells in acute myeloid leukaemia with convolutional neural networks.Nat Mach Intell, 1:538–544, 2019. DOI: 10.1038/s42256-019-0101-9

work page doi:10.1038/s42256-019-0101-9 2019
[24]

Kratz, S.-H

A. Kratz, S.-H. Bengtsson, J. E. Casey, J. M. Keefe, G. H. Beatrice, D. Y. Grzy- bek, M. B. Lewandrowski, and K. B. Lewandrowski. Performance evaluation of the CellaVision DM96 system: WBC differentials by automated digital image analysis sup- ported by an artificial neural network.Am J Clin Pathol, 124(5):770–781, 2005. DOI: 10.1309/XMB9K0J41LHLATAY. 41

work page doi:10.1309/xmb9k0j41lhlatay 2005
[25]

European Parliament and Council of the European Union. Regulation (EU) 2017/746 on in vitro diagnostic medical devices, Annex I, Chapter II, Section 9.1.Official Journal of the European Union, L 117:176–332, 2017.https://eur-lex.europa.eu/legal-content/ EN/TXT/?uri=CELEX:32017R0746 Acknowledgements The author thanks Henry Pinkard and Laura Waller (Waller ...

work page doi:10.5061/dryad.sxksn038sand 2017