Pith. sign in

REVIEW 4 major objections 5 minor 20 references

Efficiently Generating Multidimensional Calorimeter Data with Tensor Decomposition Parameterization

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper argues that having GANs and diffusion models output compact canonical-polyadic tensor factors instead of full 3D tensors can cut output parameters by roughly 80–90% while keeping Fréchet Inception Distance scores flat.

desk verdict A useful efficiency idea for generative models on tensor data, but the core plateau claim rests on an under-specified FID evaluation and a non-standard GAN baseline. read the letter →

arxiv 2508.19443 v1 pith:YLPBLNKV submitted 2025-08-26 cs.LG

Paimon Goulart , Shaan Pakala , Evangelos Papalexakis This is my paper
classification cs.LG
keywords tensordecompositioncanonicalpolyadicgenerativeadversarialnetworksdiffusionmodelscalorimetersimulationparameterefficiencyFréchetInceptionDistancesyntheticdata
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

The paper tries to show that generative models for multidimensional simulation data do not need to output the full tensor. Instead, the generator predicts several small factor matrices of a canonical polyadic decomposition, and the full tensor is assembled from those factors. On 25×51×51 calorimeter shower images, both a GAN and two diffusion variants reach a sharp FID drop followed by a plateau at about 10–20% of the full output parameters. If the claim holds, high-resolution simulation datasets could be generated with far fewer output parameters and lower overall model cost, without sacrificing sample quality. The practical case matters because calorimeter simulations are expensive to produce.

What carries the argument

Canonical Polyadic Decomposition (CPD): the paper replaces direct output of a full tensor with output of rank-R outer-product factors. The generator (or denoiser) predicts factor matrices A ∈ R^{I×R}, B ∈ R^{J×R}, C ∈ R^{K×R}; combining them via Σ a_r ∘ b_r ∘ c_r reconstructs the generated tensor. This is what converts I×J×K output parameters into (I+J+K)×R parameters. The paper pairs this with standard GAN adversarial training (discriminator viewing 2D slices of the tensor) and with DDIM-style diffusion, including a tensor-to-factor variant that keeps the output-size reduction while avoiding upfront decomposition of every training sample.

What would settle it

Recompute FID with an explicit protocol: choose the 2D slices (or 3D-aware features), the feature extractor, and the sample count; then compare the rank ordering of models against a physics-based test such as using the generated showers to train a classifier of electron vs. pion vs. photon. If the physics-based ordering disagrees with the FID ordering, the plateau claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that parameterizing a generative model's output by a rank-R canonical polyadic decomposition preserves sample quality while drastically shrinking the output layer. For a tensor X ∈ R^{I×J×K}, CPD writes X ≈ Σ_{r=1}^R a_r ∘ b_r ∘ c_r, so the model emits (I+J+K)R parameters instead of IJK. Experiments on electron, pion, and photon calorimeter images show FID improving sharply and then plateauing at roughly 10–20% of the full output parameters for both GANs and diffusion models. For diffusion, the paper compares two designs: factor-to-factor, where three separate denoisers produce the three factors, and tensor-to-factor, where one model consumes a noisy full tensor and emit

Load-bearing premise

The paper's central claim rests on Fréchet Inception Distance being a valid quality measure for 25×51×51 calorimeter tensors, even though that score was designed for 2D natural images and the paper does not specify how it is computed on this 3D data.

Editorial extensions

If this is right

  • Low-rank CPD output can match full-output FID with roughly 80–90% fewer output parameters on the tested calorimeter data.
  • Both GAN and diffusion backbones benefit, so the parameterization is not tied to one training objective.
  • The tensor-to-factor diffusion variant dominates factor-to-factor in FID, making it the more promising design for scalable generation.
  • Because output size is the bottleneck in high-resolution tensor generation, reducing tensor rank gives a direct handle on memory and computational cost.

Reading between the lines

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

  • A natural extension is to measure wall-clock time and memory usage, since the paper reports parameter counts and FID, not training or sampling speed; the rank at which speed actually improves may not match the FID plateau.
  • The same internal-decomposition idea could be tried on higher-order tensors (4D+), where I×J×K×L parameter counts explode and (I+J+K+L)R savings grow; CPD's expressiveness at fixed rank becomes the key question.
  • A downstream test—using generated showers as training data for a particle-identification or energy-estimation task—would convert the FID plateau into a physics-utility claim, which is the direction the paper lists as future work.
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 to reduce the cost of generative models for multidimensional simulation data by making the generator predict CPD factor matrices instead of the full tensor. For a tensor X ∈ R^{I×J×K}, the generator outputs (I+J+K)R parameters instead of IJK parameters. The method is applied to GANs and to two diffusion variants, factor-to-factor and tensor-to-factor. Experiments on a public 25×51×51 calorimeter dataset compare FID as a function of the output-parameter fraction relative to the full tensor. The authors report a sharp FID decrease followed by a plateau at roughly 10–20% of full output parameters, concluding that low-rank tensor decomposition can maintain similar performance while substantially reducing parameters.

Significance. If the central claim holds, the idea is simple, broadly applicable, and potentially impactful: any generative model with tensor-valued output can be reparameterized through CPD to reduce output size and memory. The algebraic parameter-count reduction is direct and not circular, and the authors release code. However, the evidence presented is not yet sufficient. FID computation for 3D calorimeter tensors is unspecified, the GAN baseline is not a standard full-output GAN, the FID curves lack error bars or multiple seeds, and no downstream physics task is used. The authors' own future-work section acknowledges the absence of downstream utility tests. The central claim is plausible but not currently established.

major comments (4)
  1. [Section 3.3 / Figures 5 and 6] The evaluation section states only "we choose to use FID" but never specifies how a 25×51×51 tensor is processed before computing the Fréchet Inception Distance. Inception-V3 expects 2D RGB images, so the paper must specify the exact preprocessing: slice selection, projection, channel replication, resizing, or another procedure, as well as the feature extractor and the number of real/generated samples used. All Section 4.1 performance claims and the plateau conclusion in Figure 5 and Figure 6 rest entirely on this metric. Without this specification, the curves are not reproducible and their meaning is unclear.
  2. [Section 4.1, Figures 5 and 6] The reported FID curves show no error bars, no multiple seeds, and no statistical tests. The claim of a "sharp decrease followed by a plateau at around 10-20%" is based on single-run comparisons, and the observed differences between nearby ranks could be noise. Please provide at least three seeds with error bars, or otherwise quantify the uncertainty, and state the sample size used for each FID estimate.
  3. [Section 4.1, GAN baseline] The red dashed line for the GAN comparison is not a standard full-output GAN; the text says "we used a high-rank tensor decomposition such that the output parameters match the full tensor." This means the baseline is itself a decomposed GAN, not the usual generator that outputs the full tensor. The conclusion "maintain similar performance to generative models with more parameters" is therefore not comparing against the standard baseline. Please include a true full-output GAN baseline, or explicitly justify why the existing baseline is the appropriate comparator and discuss how this affects the interpretation.
  4. [Section 5 / Conclusion] The conclusion states that the generated data "remains useful" and that the method reduces "overall model parameters." However, the experiments measure only FID, not downstream utility, and the x-axis of Figures 5 and 6 reports output parameters as a percentage of the full tensor, not total trainable parameters or runtime. Please either temper these claims or add supporting measurements: total model parameters, generation time, and ideally a downstream task (e.g., using generated showers to train a classifier) or a domain-specific physics metric.
minor comments (5)
  1. [Section 3.2.1, Eq. (1)] The notation "Et,A0,B0,C0,ϵ" should be written as a proper expectation with subscripts, and the predicted factors in the loss should be defined explicitly (e.g., A_pred, B_pred, C_pred). Currently the equation is ambiguous.
  2. [Section 3.2.1] The text says "three models in parallel" but Eq. (1) presents a single combined loss. Clarify whether one model with three output heads is used or three independent models, and how the combined loss is computed in the latter case.
  3. [Section 3.1.2] The discriminator description is underspecified: "perform 2D convolutions on each slice ... pools these outputs and passes this through a dense layer" should include layer sizes, number of filters, and how the 25 slices are combined. This is needed for reproducibility.
  4. [Throughout] The dataset shape is written as "R25x51x51" in multiple places; use consistent mathematical notation such as \mathbb{R}^{25\times 51\times 51}.
  5. [Figures 7–9] The sample images are helpful, but the figure captions could state more explicitly which rows correspond to which ranks and how the selected slices relate to the tensor axes. Currently the reader must infer this.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the parameter-count reduction is algebraic and the quality claims are empirical.

full rationale

The paper's central claim is that generating tensor factors instead of full tensors reduces output parameters while maintaining FID. This reduction is an algebraic identity: full tensor output is i*i*j parameters versus (i+i+j)*r for CPD factors (Sections 3.1, 3.2), and no fitted parameter is relabeled as a prediction. The FID curves in Section 4.1 are empirical measurements comparing decomposition-rank variants against full-parameter baselines; they are not derived from the model's own training loss. The losses in Eq. (1) and Eq. (2) compare model predictions against ground-truth factors or tensors, which is a standard training objective, not a circular construction. The main self-citation, TenGAN [17], is used to motivate generating tensor factors and is not the load-bearing evidence for the efficiency result; the parameter-count comparison is made directly in the paper. The other self-citations are standard tensor-decomposition surveys. The lack of detail on how FID is computed for 3D calorimeter data is a legitimate evaluation-validity concern, but it is not a circularity: the metric is externally defined and the reported comparisons are between outputs of different models, not between an input and a renamed version of the same input. Therefore no circular step is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central contribution is architectural and empirical, not physical. No new entities are postulated. The only tuned quantity is the CPD rank, and the main unstated inputs are the validity of FID for 3D physics tensors and the adequacy of low-rank CPD for this data.

free parameters (1)
  • CPD rank r = Values tested but not listed; varied to produce Figures 5 and 6, from near-full output parameters down to low fractions
    The rank controls output parameter count and is the primary design variable swept in experiments. No principled selection rule is given; lower ranks are chosen to show the FID plateau.
assumptions (3)
  • domain assumption Low-rank CPD captures enough of the calorimeter tensor structure that samples drawn from factor space reconstruct realistic tensors
    Section 3.2 and Figures 5 and 6 assume reduced-rank CPD factors are a valid generative parameterization; the 'plateau' claim depends on low-rank sufficiency.
  • domain assumption FID is a valid evaluation metric for 25x51x51 calorimeter tensors
    Section 3.3 states FID is used but not how. FID's Inception features are trained on 2D natural images, and applying it to 3D physics images without a specified adaptation is an unverified assumption.
  • standard math Predicting x0 in DDIM is equivalent to predicting noise, per Luo (2022)
    Section 3.2 relies on this equivalence to justify predicting clean samples; it is accepted in the diffusion literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficiently Generating Multidimensional Calorimeter Data with Tensor Decomposition Parameterization." pith.science (2026). https://pith.science/paper/YLPBLNKV

@misc{pith2026250819443,
  author       = {Pith},
  title        = {Pith review of: Efficiently Generating Multidimensional Calorimeter Data with Tensor Decomposition Parameterization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YLPBLNKV}},
  note         = {Machine review of arXiv:2508.19443}
}
read the original abstract

Producing large complex simulation datasets can often be a time and resource consuming task. Especially when these experiments are very expensive, it is becoming more reasonable to generate synthetic data for downstream tasks. Recently, these methods may include using generative machine learning models such as Generative Adversarial Networks or diffusion models. As these generative models improve efficiency in producing useful data, we introduce an internal tensor decomposition to these generative models to even further reduce costs. More specifically, for multidimensional data, or tensors, we generate the smaller tensor factors instead of the full tensor, in order to significantly reduce the model's output and overall parameters. This reduces the costs of generating complex simulation data, and our experiments show the generated data remains useful. As a result, tensor decomposition has the potential to improve efficiency in generative models, especially when generating multidimensional data, or tensors.

Figures

Figures reproduced from arXiv: 2508.19443 by the authors.

Figure 1
Figure 1. Example GAN with internal tensor decomposition. The [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Factor to Factor diffusion sampling process. Each factor [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Progressive denoising of each factor matrix over time (from left to right). The three rows correspond to factor matrices A, B, and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 6
Figure 6. Figure 6: For the diffusion models, we display the Fr [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 5
Figure 5. Figure 5: For the GANs, we display the Frechet Inception Distance ´ (FID) versus the output parameters. Output parameters shown as percent of full tensor (vs. smaller factor outputs). For both GAN & diffusion plots, we notice a sharp de￾crease in the FID, followed by a plateau a…
Figure 8
Figure 8. Figure 8: We display some sample images. The first row is the [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 7
Figure 7. Figure 7: We display some sample images. The first row is the [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 9
Figure 9. Figure 9: We display some sample images. The first row is the [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

  1. [1]

    Denoising diffusion models with geometry adaptation for high fidelity calorimeter simu- lation, 2023

    Oz Amram and Kevin Pedro. Denoising diffusion models with geometry adaptation for high fidelity calorimeter simu- lation, 2023. 1

    work page 2023

  2. [2]

    Tucker: Tensor factorization for knowledge graph comple- tion

    Ivana Bala ˇzevi´c, Carl Allen, and Timothy M Hospedales. Tucker: Tensor factorization for knowledge graph comple- tion. arXiv preprint arXiv:1901.09590, 2019. 7

    arXiv 1901

  3. [3]

    Overview of the coupled model intercomparison project phase 6 (cmip6) experimental design and organiza- tion

    Veronika Eyring, Sandrine Bony, Gerald A Meehl, Cather- ine A Senior, Bjorn Stevens, Ronald J Stouffer, and Karl E Taylor. Overview of the coupled model intercomparison project phase 6 (cmip6) experimental design and organiza- tion. Geoscientific Model Development , 9(5):1937–1958,

    work page 1937

  4. [4]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commu- nications of the ACM, 63(11):139–144, 2020. 1, 2

    work page 2020

  5. [5]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. Advances in neural information processing systems , 30, 2017. 1, 3

    work page 2017

  6. [6]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. In Advances in Neural Infor- mation Processing Systems, pages 6840–6851. Curran Asso- ciates, Inc., 2020. 1, 2

    work page 2020

  7. [7]

    com, Johannes Albrecht, Antonio Augusto Alves, Guilherme Amadio, Giuseppe Andronico, Nguyen Anh-Ky, Laurent Aphecetche, John Apostolakis, Makoto Asai, Luca Atzori, et al

    HEP Software Foundation hsf-editorial-secretariat@ google- groups. com, Johannes Albrecht, Antonio Augusto Alves, Guilherme Amadio, Giuseppe Andronico, Nguyen Anh-Ky, Laurent Aphecetche, John Apostolakis, Makoto Asai, Luca Atzori, et al. A roadmap for hep software and computing r&d for the 2020s. Computing and software for big science, 3:1–49, 2019. 1

    work page 2019

  8. [8]

    Fast simulation of a high granularity calorimeter by generative adversarial networks

    Gul Rukh Khattak, Sofia Vallecorsa, Federico Carminati, and Gul Muhammad Khan. Fast simulation of a high granularity calorimeter by generative adversarial networks. The Euro- pean Physical Journal C, 82(4):386, 2022. 1

    work page 2022

Show all 20 references
  1. [9]

    Kolda and B.W

    T.G. Kolda and B.W. Bader. Tensor decompositions and ap- plications. SIAM review, 51(3), 2009. 1, 7

    2009

  2. [10]

    Learning skillful medium-range global weather forecasting

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416–1421, 2023. 1

    2023

  3. [11]

    Understanding diffusion models: A unified per- spective, 2022

    Calvin Luo. Understanding diffusion models: A unified per- spective, 2022. 2

    2022

  4. [12]

    The well: a large-scale collection of diverse physics simulations for machine learning

    Ruben Ohana, Michael McCabe, Lucas Meyer, Rudy Morel, Fruzsina Agocs, Miguel Beneitez, Marsha Berger, Blakesly Burkhart, Stuart Dalziel, Drummond Fielding, et al. The well: a large-scale collection of diverse physics simulations for machine learning. Advances in Neural Informa...

    2024

  5. [13]

    Pierini and M

    M. Pierini and M. Zhang. Clic calorimeter 3d images: Elec- tron showers at random angle, 2020. 3

    2020

  6. [14]

    Pierini and M

    M. Pierini and M. Zhang. Clic calorimeter 3d images: Neu- tral pion showers at random angle, 2020. 3

    2020

  7. [15]

    Pierini and M

    M. Pierini and M. Zhang. Clic calorimeter 3d images: Pho- ton showers at random angle, 2020. 3

    2020

  8. [16]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. CoRR, abs/2112.10752,

    arXiv

  9. [17]

    Tengan: ad- versarially generating multiplex tensor graphs

    William Shiao, Benjamin A Miller, Kevin Chan, Paul Yu, Tina Eliassi-Rad, and Evangelos E Papalexakis. Tengan: ad- versarially generating multiplex tensor graphs. Data Mining and Knowledge Discovery, 38(1):1–21, 2024. 1, 2

    2024

  10. [18]

    How good is my gan? In Proceedings of the Eu- ropean conference on computer vision (ECCV) , pages 213– 229, 2018

    Konstantin Shmelkov, Cordelia Schmid, and Karteek Ala- hari. How good is my gan? In Proceedings of the Eu- ropean conference on computer vision (ECCV) , pages 213– 229, 2018. 3

    2018

  11. [19]

    Tensor decomposition for signal processing and ma- chine learning

    Nicholas D Sidiropoulos, Lieven De Lathauwer, Xiao Fu, Kejun Huang, Evangelos E Papalexakis, and Christos Falout- sos. Tensor decomposition for signal processing and ma- chine learning. IEEE Signal Processing Magazine, 2016. 1

    2016

  12. [20]

    Denois- ing diffusion implicit models, 2022

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models, 2022. 1, 2 5 Appendix Sample Images Figure 7. We display some sample images. The first row is the real data, the last row is a GAN using the full output parameters, and the remaining rows are...

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.