Pith. sign in

REVIEW 2 major objections 5 minor 47 references

A graph transformer on superpixel graphs matches top image-classification accuracy while adding translation invariance and dropping the need for boundary coordinates.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 20:31 UTC pith:BYRYMERH

load-bearing objection Solid CIFAR-10 match to ShapeGNN via GraphGPS on mean-centered SLIC RAGs, plus better low-data/noise behavior, but the title/abstract over-sell rotation invariance that the authors themselves drop after it costs 2.5 points. the 2 major comments →

arxiv 2607.04262 v1 pith:BYRYMERH submitted 2026-07-05 cs.LG

On Preserving Geometrical Invariance for Superpixel Image Classification using Graph Transformer

Sarabeshwar Balaji , Shubham Mohanty , Akash Anil This is my paper
classification cs.LG
keywords Graph Transformersuperpixelstranslation invariancerotation invarianceSLICimage classificationCIFAR-10region adjacency graph
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Dense pixel grids force CNNs and vision transformers to spend computation on redundant local information, and ordinary graph neural networks on superpixel graphs still miss long-range relations and geometric invariances. This paper shows that a Graph Transformer run on region-adjacency graphs of SLIC superpixels, after simple mean-centering of the superpixel centroids, reaches 80.19 percent accuracy on CIFAR-10. That figure is essentially identical to the previous best superpixel model yet does not require the expensive boundary-point coordinates that model used. The same architecture also stays more accurate when only a fraction of the training images are available and when the test graphs are deliberately corrupted by random edge or node edits. The result suggests that global attention plus a lightweight geometric pre-processing step can replace both dense convolutions and elaborate shape encodings for efficient, robust image classification.

Core claim

SuperGT, a GraphGPS-style transformer that mixes local message passing with global attention on SLIC region-adjacency superpixel graphs whose centroids have been mean-centered, attains 80.19 percent CIFAR-10 accuracy, matching the prior state-of-the-art ShapeGNN while remaining translation-invariant and free of any need for superpixel boundary coordinates.

What carries the argument

SuperGT: GraphGPS layers (GAT message-passing plus Performer global attention) whose node features are the usual superpixel statistics concatenated with GPSE-GEOM embeddings, after the two centroid coordinates have been replaced by their mean-centered versions.

Load-bearing premise

Mean-centering the superpixel centroids is enough geometric pre-processing for robust classification, even though the authors themselves found that the fuller PCA scheme that also enforces rotation invariance lowers accuracy by roughly 2.5 percent and therefore abandoned it.

What would settle it

Train SuperGT and ShapeGNN on the same CIFAR-10 split, then evaluate both on a test set whose images have been randomly rotated by multiples of 90 degrees (or continuously); if SuperGT's accuracy falls below ShapeGNN's by more than the original 0.01-point margin, the claimed practical value of the translation-only pre-processing is refuted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Superpixel-graph classifiers can drop the costly extraction of boundary-point coordinates without losing accuracy.
  • Global-attention layers make superpixel models markedly more sample-efficient than pure message-passing GNNs when labeled images are scarce.
  • The same global-attention architecture degrades more slowly under random edge addition, edge deletion, or node deletion, suggesting greater robustness to segmentation noise.
  • A simple mean-centering step is already sufficient to inject translation invariance into any superpixel pipeline that previously used raw centroid coordinates.

Where Pith is reading between the lines

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

  • Because the full PCA rotation-invariant transform hurts accuracy, the field may need equivariant rather than invariant positional features if both translation and rotation robustness are required.
  • The large gap between ordinary MPNNs and SuperGT on the same RAG graphs implies that long-range attention, not the choice of superpixel algorithm, is the main missing ingredient in earlier superpixel classifiers.
  • If the same mean-centered Graph Transformer transfers to higher-resolution datasets, it could become a practical drop-in replacement for lightweight CNNs in edge-vision settings where pixel grids are too costly.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes SuperGT, a GraphGPS-style Graph Transformer applied to SLIC region-adjacency superpixel graphs for image classification. Node features include mean color, color variance, area fraction and centroid coordinates; the latter are mean-centered to obtain translation invariance. A PCA-based canonicalization that would also enforce rotation invariance is derived (Eqs. 1–3) but is abandoned after an ablation shows a ~2.5 % accuracy drop. With GPSE-GEOM positional encodings and Performer attention the model reaches 80.19 % on CIFAR-10, matching ShapeGNN (80.18 %) without boundary-point coordinates, and outperforms strong GNN baselines under reduced training-set sizes and under structural noise (edge/node addition/removal).

Significance. If the empirical picture holds, SuperGT supplies a practical demonstration that global attention on superpixel RAGs can match the current superpixel SOTA while remaining simpler (no boundary coordinates) and more sample-efficient. The multi-seed PE/SE ablation (Table 1), limited-data curves and structural-noise tests (Figure 3) are useful additions to the superpixel-GNN literature. The elementary PCA proofs are correct and the code is released. The geometric-invariance framing that motivates the title and abstract is, however, only partially realized, which limits the conceptual advance relative to ordinary mean-centering already common in point-cloud pipelines.

major comments (2)
  1. Title, abstract and research question (p. 1) advertise a pre-processing scheme that “preserves translation/rotation invariance.” Section 3.2.1 derives a formally correct PCA canonicalization (Eqs. 1–3) that is both translation- and rotation-invariant, yet the authors immediately report that it lowers accuracy by ~2.5 % (Table 3: 77.68 % vs 80.19 %) and therefore discard it: “all the subsequent experiments employ only mean-centering … preserving only the translation invariance.” Consequently the headline numbers that match ShapeGNN (Table 2) and the robustness curves (Fig. 3) are obtained without the rotation-invariance component that the paper claims to deliver. Mean-centering alone is a trivial centering step; it does not justify the geometric-invariance framing that motivates the work. Either the full PCA scheme must be retained and its performance gap closed, or the title/abstract/cla
  2. All quantitative claims rest on a single dataset (CIFAR-10). No results are reported on MNIST, Fashion-MNIST, STL-10, ImageNet-scale subsets or any other standard superpixel-GNN benchmark. Without at least one additional dataset it is impossible to judge whether the 80.19 % figure and the limited-data/noise advantages generalize, or whether they are CIFAR-10-specific. A second dataset (or a clear statement of scope limitation) is required for the central empirical claim to be load-bearing.
minor comments (5)
  1. SLIC superpixel count and compactness are never stated; they are free parameters that affect graph size and should be reported for reproducibility.
  2. Figure 1 uses an ImageNet example while all experiments are on CIFAR-10; a CIFAR-10 illustration would be more consistent.
  3. Table 1 and Table 2 use color coding that may not survive grayscale printing; numerical ranking or bold/underline would be clearer.
  4. The Hu-moment ablation (Table 3) is described as “marginal” yet is never compared against ShapeGNN’s own shape encoding; a short discussion would strengthen the ablation section.
  5. Typographical inconsistencies appear (“SuperGT 1”, missing spaces after periods, “ElstaticPE”). A careful proof-read is needed.

Circularity Check

0 steps flagged

No circular derivation: SuperGT accuracy is measured on held-out CIFAR-10; PCA algebra is independent of the target metric and is not used for the headline numbers.

full rationale

The paper's derivation chain is self-contained and non-circular. Superpixel graphs are built by SLIC + RAG (Section 3.1); node features are mean-centered (or optionally PCA-canonicalized via Eqs. 1–3) before a GraphGPS-style transformer with PE/SE is trained by standard supervised cross-entropy on the CIFAR-10 training split and evaluated on the fixed 10 k test split (Tables 1–3, Fig. 3). The PCA algebra is a pure linear-algebra identity that does not embed classification accuracy; the authors themselves discard the full rotation-invariant version after an empirical drop (Table 3) and report only mean-centering for the main results. PE/SE variants are selected by validation performance, not by tautology. No parameter is fitted to a subset and then re-presented as a prediction of a closely related quantity; no uniqueness theorem or load-bearing premise is imported solely via self-citation. The geometric-invariance framing is incomplete relative to the title/abstract, but that is a correctness/overclaim issue, not circularity. Score 0 is therefore the honest finding.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central empirical claim rests on standard SLIC segmentation, the GraphGPS architecture, and the modeling choice that mean-centering centroids is the right geometric pre-processing. No new physical entities are postulated; free parameters are ordinary deep-learning hyper-parameters selected by Optuna on a training subset.

free parameters (3)
  • learning rate / weight decay / dropout / batch size
    Selected by Bayesian optimization (Optuna) on a 10 k stratified subset of the CIFAR-10 training split; final values (lr 9e-4, wd 7.7e-4, etc.) are reported but remain free parameters of the fit.
  • number of GraphGPS layers (7) and PE/SE choice (GPSE-GEOM + Performer)
    Architectural choices fixed after Table 1 ablation; not derived from first principles.
  • SLIC superpixel count and compactness (implicit)
    Not fully specified in the text; controls the size of every graph and therefore the measured accuracy.
axioms (3)
  • domain assumption SLIC superpixel segmentation is sufficiently translation- and rotation-invariant that only the centroid coordinates need geometric correction.
    Stated in Section 3.2; required for the claim that mean-centering (or PCA) yields image-level geometric invariance.
  • domain assumption Region-adjacency graphs plus global attention capture the long-range dependencies needed for CIFAR-10 classification better than pure message-passing.
    Motivates the entire SuperGT design (Introduction and Section 3.3).
  • ad hoc to paper Mean-centering of centroids is an adequate practical substitute for full rotation+translation invariance.
    Adopted after the PCA scheme that enforces both invariances is observed to drop accuracy 2.5% (Table 3); used for all main results.

pith-pipeline@v1.1.0-grok45 · 16866 in / 2771 out tokens · 27047 ms · 2026-07-11T20:31:54.342922+00:00 · methodology

0 comments
read the original abstract

Convolutional Neural Network (CNN) and Vision Transformer (ViT) for image classification exploit a dense grid of pixels containing redundant information. Consequently, for a larger image dataset, CNNs and ViTs face deployability challenges due to high computational complexity. Representing images as graphs of superpixels offers an efficient alternative that preserves key information while eliminating pixel-level redundancy. Graph Neural Networks (GNNs) have been utilized on such graphs to perform image classification. However, GNNs are known to struggle with capturing long-range dependencies which is important in the domain of image classification. Furthermore, a majority of these superpixel-based image classification approaches do not explicitly preserve translation/rotation invariance. Nevertheless, preserving translation/rotation invariance is important for robust image classification. Thus, this paper proposes SuperGT, a Graph Transformer-based framework for image classification, which captures the long range dependencies, along with a pre-processing scheme that preserves translation/rotation invariance. We evaluate SuperGT on CIFAR-10 dataset and observe that it performs significantly better than many baselines. Furthermore, we note that the overall performance of SuperGT is comparable to the previous state-of-the-art model, namely, ShapeGNN, without relying on coordinates of the boundary points of each superpixel required by ShapeGNN.

Figures

Figures reproduced from arXiv: 2607.04262 by Akash Anil, Sarabeshwar Balaji, Shubham Mohanty.

Figure 1
Figure 1. Figure 1: Illustration of the superpixel graph construction (a) Original input image (from ImageNet [25]). (b) SLIC segmentation boundaries overlaid on the image. (c) Region Adjacency Graph (RAG), where blue line segments indicate edges between neighboring superpixels [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Proposed framework. Each image is first segmented into superpixels, from which a Region Adjacency Graph (RAG) is constructed. For visualization, here the graph nodes are colored according to the mean color of their corresponding superpixels. We then apply the proposed feature preprocessing scheme to incorporate geometric invariance, including translation and rotation invariance. Then positional and structu… view at source ↗
Figure 3
Figure 3. Figure 3: Robustness analysis. be noted that unlike ShapeGNN, SuperGT is invariant to trans￾lations of the underlying image and does not require coordi￾nates of the boundary points of each superpixel. The supe￾rior performance of SuperGT can be attributed to the ability of GlobalAttnℓ to capture long-range dependencies. Performance under limited training data. Now we evaluate SuperGT on the same test split while tra… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 5 linked inside Pith

  1. [1]

    Simonyan, A

    K. Simonyan, A. Zisserman, Very deep convolutional net- works for large-scale image recognition, in: 3rd Inter- national Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, Association for Computing Machin- ery, 2015

    2015

  2. [2]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  3. [3]

    S. Yu, J. Ma, Deep learning for geophysics: Current and future trends, Reviews of Geophysics 59 (3) (2021) e2021RG000742

    2021

  4. [4]

    S. Min, B. Lee, S. Yoon, Deep learning in bioinformatics, Briefings in bioinformatics 18 (5) (2017) 851–869

    2017

  5. [5]

    Maggiori, Y

    E. Maggiori, Y . Tarabalka, G. Charpiat, P. Alliez, Con- volutional neural networks for large-scale remote-sensing image classification, IEEE Transactions on geoscience and remote sensing 55 (2) (2016) 645–657

    2016

  6. [6]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, Advances in neural information processing systems 30 (2017)

    2017

  7. [7]

    R. A. Cosma, L. Knobel, P. van der Linden, D. M. Knigge, E. J. Bekkers, Geometric superpixel representations for efficient image classification with graph neural networks, in: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, 2023, pp. 109–118

    2023

  8. [8]

    J. Long, Z. yan, H. chen, A graph neural network for superpixel image classification, in: Journal of Physics: Conference Series, V ol. 1871, IOP Publishing, 2021, p. 012071

    2021

  9. [9]

    P. H. Avelar, A. R. Tavares, T. L. da Silveira, C. R. Jung, L. C. Lamb, Superpixel image classification with graph at- tention networks, in: 2020 33rd SIBGRAPI conference on Graphics, patterns and images (SIBGRAPI), IEEE, 2020, pp. 203–209

    2020

  10. [10]

    Bae, G.-H

    J.-H. Bae, G.-H. Yu, J.-H. Lee, D. T. Vu, L. H. Anh, H.- G. Kim, J.-Y . Kim, Superpixel image classification with graph convolutional neural networks based on learnable positional embedding, Applied Sciences 12 (18) (2022) 9176

    2022

  11. [11]

    C.-H. Youn, et al., Dynamic graph neural network for super-pixel image classification, in: 2021 International Conference on Information and Communication Technol- ogy Convergence (ICTC), IEEE, 2021, pp. 1095–1099

    2021

  12. [12]

    J. P. Rodrigues, J. L. Carbonera, Graph convolutional net- works for image classification: Comparing approaches for building graphs from images., in: ICEIS (1), 2024, pp. 437–446

    2024

  13. [13]

    K. Oono, T. Suzuki, Graph neural networks exponentially lose expressive power for node classification, in: Interna- tional Conference on Learning Representations, 2020

    2020

  14. [14]

    U. Alon, E. Yahav, On the bottleneck of graph neural networks and its practical implications, in: 9th Inter- national Conference on Learning Representations, ICLR 2021, 2021

    2021

  15. [15]

    Topping, F

    J. Topping, F. Di Giovanni, B. P. Chamberlain, X. Dong, M. M. Bronstein, Understanding over-squashing and bottlenecks on graphs via curvature, arXiv preprint arXiv:2111.14522 (2021)

    Pith/arXiv arXiv 2021

  16. [16]

    V . P. Dwivedi, X. Bresson, A generalization of transformer networks to graphs, arXiv preprint arXiv:2012.09699 (2020)

    Pith/arXiv arXiv 2012

  17. [17]

    Kreuzer, D

    D. Kreuzer, D. Beaini, W. Hamilton, V . Létourneau, P. Tossou, Rethinking graph transformers with spectral at- tention, Advances in neural information processing sys- tems 34 (2021) 21618–21629

    2021

  18. [18]

    Rampášek, M

    L. Rampášek, M. Galkin, V . P. Dwivedi, A. T. Luu, G. Wolf, D. Beaini, Recipe for a general, powerful, scal- able graph transformer, Advances in Neural Information Processing Systems 35 (2022) 14501–14515

    2022

  19. [19]

    Z. Wu, P. Jain, M. Wright, A. Mirhoseini, J. E. Gonza- lez, I. Stoica, Representing long-range context for graph neural networks with global attention, Advances in neural information processing systems 34 (2021) 13266–13279

    2021

  20. [20]

    Hotelling, Analysis of a complex of statistical vari- ables into principal components., Journal of educational psychology 24 (6) (1933) 417

    H. Hotelling, Analysis of a complex of statistical vari- ables into principal components., Journal of educational psychology 24 (6) (1933) 417. 7

    1933

  21. [21]

    Krizhevsky, V

    A. Krizhevsky, V . Nair, G. Hinton, et al., The cifar-10 dataset (2014)

    2014

  22. [22]

    Veli ˇckovi´c, G

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Liò, Y . Bengio, Graph attention networks, in: Interna- tional Conference on Learning Representations, 2018

    2018

  23. [23]

    Achanta, A

    R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, S. Süsstrunk, Slic superpixels compared to state-of-the-art superpixel methods, IEEE transactions on pattern analysis and machine intelligence 34 (11) (2012) 2274–2282

    2012

  24. [24]

    V . G. Satorras, E. Hoogeboom, M. Welling, E (n) equiv- ariant graph neural networks, in: International conference on machine learning, PMLR, 2021, pp. 9323–9332

    2021

  25. [25]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei- Fei, Imagenet: A large-scale hierarchical image database, in: 2009 IEEE conference on computer vision and pattern recognition, Ieee, 2009, pp. 248–255

    2009

  26. [26]

    Bezick, V

    M. Bezick, V . Giammarino, A. H. Qureshi, Robust point cloud reinforcement learning via pca-based canonicaliza- tion, arXiv preprint arXiv:2510.20974 (2025)

    arXiv 2025

  27. [27]

    Z. Xiao, H. Lin, R. Li, L. Geng, H. Chao, S. Ding, En- dowing deep 3d models with rotation invariance based on principal component analysis, in: 2020 IEEE international conference on multimedia and expo (ICME), IEEE, 2020, pp. 1–6

    2020

  28. [28]

    P. W. Battaglia, J. B. Hamrick, V . Bapst, A. Sanchez- Gonzalez, V . Zambaldi, M. Malinowski, A. Tacchetti, D. Raposo, A. Santoro, R. Faulkner, et al., Relational in- ductive biases, deep learning, and graph networks, arXiv preprint arXiv:1806.01261 (2018)

    Pith/arXiv arXiv 2018

  29. [29]

    V . P. Dwivedi, C. K. Joshi, A. T. Luu, T. Laurent, Y . Ben- gio, X. Bresson, Benchmarking graph neural networks, Journal of Machine Learning Research 24 (43) (2023) 1– 48

    2023

  30. [30]

    Cantürk, R

    S. Cantürk, R. Liu, O. Lapointe-Gagné, V . Létourneau, G. Wolf, D. Beaini, L. Rampášek, Graph positional and structural encoder, in: Forty-first International Conference on Machine Learning, 2024

    2024

  31. [31]

    V . P. Dwivedi, A. T. Luu, T. Laurent, Y . Bengio, X. Bres- son, Graph neural networks with learnable structural and positional representations, in: International Conference on Learning Representations, 2022

    2022

  32. [32]

    D. Chen, L. O’Bray, K. Borgwardt, Structure-aware trans- former for graph representation learning, in: International conference on machine learning, PMLR, 2022, pp. 3469– 3489

    2022

  33. [33]

    M. S. Hussain, M. J. Zaki, D. Subramanian, Global self-attention as a replacement for graph convolution, in: Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, 2022, pp. 655–665

    2022

  34. [34]

    W. Park, W. Chang, D. Lee, J. Kim, S.-w. Hwang, Grpe: Relative positional encoding for graph transformer, arXiv preprint arXiv:2201.12787 (2022)

    Pith/arXiv arXiv 2022

  35. [35]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regular- ization, in: International Conference on Learning Repre- sentations, 2019

    2019

  36. [36]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Sgdr: Stochastic gradient de- scent with warm restarts, in: International Conference on Learning Representations, 2017

    2017

  37. [37]

    Akiba, S

    T. Akiba, S. Sano, T. Yanase, T. Ohta, M. Koyama, Optuna: A next-generation hyperparameter optimization framework, in: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 2623–2631

    2019

  38. [38]

    W. Hu, M. Fey, M. Zitnik, Y . Dong, H. Ren, B. Liu, M. Catasta, J. Leskovec, Open graph benchmark: Datasets for machine learning on graphs, Advances in neural infor- mation processing systems 33 (2020) 22118–22133

    2020

  39. [39]

    K. M. Choromanski, V . Likhosherstov, D. Dohan, X. Song, A. Gane, T. Sarlos, P. Hawkins, J. Q. Davis, A. Mohiuddin, L. Kaiser, et al., Rethinking attention with performers, in: International Conference on Learn- ing Representations, 2021

    2021

  40. [40]

    Axelrod, R

    S. Axelrod, R. Gomez-Bombarelli, Geom, energy- annotated molecular conformations for property predic- tion and molecular generation, Scientific data 9 (1) (2022) 185

    2022

  41. [41]

    T. N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, in: International Con- ference on Learning Representations, 2017

    2017

  42. [42]

    Brody, U

    S. Brody, U. Alon, E. Yahav, How attentive are graph at- tention networks?, in: International Conference on Learn- ing Representations, 2022

    2022

  43. [43]

    Hamilton, Z

    W. Hamilton, Z. Ying, J. Leskovec, Inductive representa- tion learning on large graphs, Advances in neural infor- mation processing systems 30 (2017)

    2017

  44. [44]

    Corso, L

    G. Corso, L. Cavalleri, D. Beaini, P. Liò, P. Veli ˇckovi´c, Principal neighbourhood aggregation for graph nets, Ad- vances in neural information processing systems 33 (2020) 13260–13271

    2020

  45. [45]

    K. Xu, W. Hu, J. Leskovec, S. Jegelka, How powerful are graph neural networks?, in: International Conference on Learning Representations, 2019

    2019

  46. [46]

    Hu, Visual pattern recognition by moment invari- ants, IRE transactions on information theory 8 (2) (1962) 179–187

    M.-K. Hu, Visual pattern recognition by moment invari- ants, IRE transactions on information theory 8 (2) (1962) 179–187

    1962

  47. [47]

    Ramer, An iterative procedure for the polygonal ap- proximation of plane curves, Computer graphics and im- age processing 1 (3) (1972) 244–256

    U. Ramer, An iterative procedure for the polygonal ap- proximation of plane curves, Computer graphics and im- age processing 1 (3) (1972) 244–256. 8

    1972