pith. sign in

arxiv: 2606.18440 · v1 · pith:RMEDJYPYnew · submitted 2026-06-16 · 🧮 math.AG

Algebraic Networks and Architectural Degenerations

Giacomo Graziani This is my paper

Pith reviewed 2026-06-26 22:10 UTC · model grok-4.3

classification 🧮 math.AG
keywords algebraic networksneurovarietiesarchitectural degeneracysingular locusfully connected networksrealization mapspolynomial neural networksidentifiability
0
0 comments X

The pith

For fully connected networks with non-increasing widths and scalar output, non-degenerate parameters yield smooth points on the associated neurovariety.

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

The paper builds an algebraic framework for polynomial neural networks that uses monomial activations and no bias terms. It defines realization maps from parameter spaces to function spaces, producing affine neurovarieties whose points are the functions realized by the network. The central result ties the geometry of these varieties to the network structure: under stated layerwise regularity conditions, any parameter that avoids rank-deficient layers or inactive neurons lands at a smooth point of the neurovariety. Consequently the singular locus sits inside the architectural degeneracy locus, the set of functions that admit at least one degenerate representation.

Core claim

For fully connected networks with non-increasing widths and scalar output, full parameters give smooth points of the corresponding neurovariety under explicit layerwise regularity assumptions. In particular, for these architectures, the singular locus is contained in the architectural degeneracy locus.

What carries the argument

The architectural degeneracy locus, the set of functions that admit a representation with a rank-deficient layer or an inactive hidden neuron; it is shown to contain all singular points of the neurovariety for the specified architectures.

If this is right

  • Singularities of the neurovariety can be detected by checking whether a function admits a degenerate parameter representation.
  • The symmetry groups and quotient parameter spaces defined in the framework classify distinct realizations of the same function.
  • Geometric identifiability and reducibility become questions about the fibers of the realization map and the components of the neurovariety.
  • The containment of the singular locus inside the degeneracy locus holds only for the listed width and output conditions together with the regularity assumptions.

Where Pith is reading between the lines

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

  • The same containment might be tested numerically by sampling parameters and checking the rank of the Jacobian of the realization map at those points.
  • Extending the framework beyond monomial activations would require replacing the polynomial realization map with a different algebraic map while preserving the degeneracy notion.
  • If the containment holds more generally, then training algorithms that avoid degenerate parameters would automatically avoid singular points in parameter space.

Load-bearing premise

The networks must have non-increasing widths, scalar output, and satisfy the explicit layerwise regularity assumptions.

What would settle it

Exhibit a parameter point with full-rank layers and no inactive neurons whose image lies on a singular point of the neurovariety for one of the networks covered by the theorem.

read the original abstract

We study the geometry of polynomial neural networks with monomial activation functions and no bias. We introduce a general framework of algebraic networks, together with their realization maps and associated affine neurovarieties. In this setting we define morphisms, subnetworks, symmetry groups and quotient parameter spaces and we discuss geometric notions of identifiability and reducibility. Our main goal is to relate the singularities of neurovarieties to degenerations of the underlying architecture. For fully connected networks, we define the architectural degeneracy locus as the locus of functions admitting a representation by parameters with a rank-deficient layer or an inactive hidden neuron. We prove that, for fully connected networks with non-increasing widths and scalar output, full parameters give smooth points of the corresponding neurovariety under explicit layerwise regularity assumptions. In particular, for these architectures, the singular locus is contained in the architectural degeneracy locus.

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

1 major / 1 minor

Summary. The paper introduces a framework of algebraic networks with monomial activations (no bias), defines realization maps, affine neurovarieties, morphisms, subnetworks, symmetry groups, and quotient spaces, and discusses identifiability and reducibility. Its central result is that, for fully connected networks with non-increasing widths and scalar output, parameters with full-rank layers and active neurons are smooth points of the neurovariety under explicit layerwise regularity assumptions; consequently the singular locus is contained in the architectural degeneracy locus (parameters with a rank-deficient layer or inactive hidden neuron).

Significance. If the containment holds, the work supplies a precise geometric link between singularities of the neurovariety and architectural degenerations, extending algebraic-geometry techniques to polynomial networks and offering tools for studying identifiability. The introduction of the general algebraic-network formalism and the explicit treatment of symmetry groups and quotients are concrete contributions that could be reused beyond the fully-connected scalar-output case.

major comments (1)
  1. [Abstract / main theorem] Abstract and main theorem (presumably §4 or §5): the smoothness statement is conditioned on separate layerwise regularity assumptions, yet the subsequent claim that 'the singular locus is contained in the architectural degeneracy locus' is stated unconditionally. If regularity can fail for some full-rank, active-neuron parameters, then singularities at those points would lie outside the degeneracy locus, falsifying the containment. The manuscript must either prove that regularity is automatic for full parameters or revise the containment statement to account for the regularity locus.
minor comments (1)
  1. [Introduction / §2] Notation for the realization map and the neurovariety could be introduced earlier and used consistently; several definitions (e.g., architectural degeneracy locus) appear only after the main claim is stated.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading of the manuscript and for identifying this important point of clarification in the presentation of the main result. We address the comment below.

read point-by-point responses

  1. Referee: [Abstract / main theorem] Abstract and main theorem (presumably §4 or §5): the smoothness statement is conditioned on separate layerwise regularity assumptions, yet the subsequent claim that 'the singular locus is contained in the architectural degeneracy locus' is stated unconditionally. If regularity can fail for some full-rank, active-neuron parameters, then singularities at those points would lie outside the degeneracy locus, falsifying the containment. The manuscript must either prove that regularity is automatic for full parameters or revise the containment statement to account for the regularity locus.

    Authors: We agree that the current wording of the abstract and the main theorem statement risks being read as claiming the containment unconditionally. The smoothness result is proven only under the stated layerwise regularity assumptions in addition to the full-rank and active-neuron hypotheses. The 'in particular' clause in the abstract is intended to indicate that the containment follows from this conditional smoothness, but the logical dependence is not made fully explicit. We will revise both the abstract and the theorem statement to condition the containment claim explicitly on the regularity assumptions as well. This change will be incorporated in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: definitions and theorem are independent of inputs

full rationale

The paper defines algebraic networks, realization maps, neurovarieties, architectural degeneracy locus, and related notions from first principles, then states and proves a containment result (singular locus inside degeneracy locus) for fully connected networks under explicit layerwise regularity assumptions plus non-increasing widths and scalar output. No quoted step reduces a claimed prediction or uniqueness result to a fitted parameter, self-citation chain, or definitional tautology; the central claim is a standard containment theorem whose hypotheses are stated separately from the conclusion.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper introduces new objects (algebraic networks, neurovarieties, architectural degeneracy locus) whose definitions rest on standard algebraic geometry background; no free parameters or invented physical entities are described.

pith-pipeline@v0.9.1-grok · 5661 in / 1117 out tokens · 30683 ms · 2026-06-26T22:10:49.343240+00:00 · methodology

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

34 extracted references · 4 canonical work pages

  1. [1]

    Aluffi , title =

    P. Aluffi , title =. 2013 , pages =

    2013

  2. [2]

    Ballico , title =

    E. Ballico , title =. Linear and Multilinear Algebra , year =

  3. [3]

    M. M. Bronstein and J. Bruna and T. Cohen and P. Veli. arXiv preprint arXiv:2104.13478 , year =

    Pith/arXiv arXiv

  4. [4]

    H. S. arXiv preprint arXiv:2508.02723 , year =

    arXiv

  5. [5]

    Borovik and H

    V. Borovik and H. Friedman and S. Hosten and M. Pfeffer , title =. arXiv preprint arXiv:2512.06939 , year =

    Pith/arXiv arXiv

  6. [6]

    Breiding and S

    P. Breiding and S. Timme , title =. 2018 , pages =

    2018

  7. [7]

    Usevich and R

    K. Usevich and R. Borsoi and C. D. 2025 , note =. 2506.17093 , archivePrefix =

    arXiv 2025

  8. [8]

    Chen and I

    T. Chen and I. Goodfellow and J. Shlens , title =. 2016 , eprint =

    2016

  9. [9]

    Draisma and E

    J. Draisma and E. Horobet and G. Ottaviani and B. Sturmfels and R. R. Thomas , title =. Foundations of Computational Mathematics , year =

  10. [10]

    Finkel and J

    B. Finkel and J. I. Rodriguez and C. Wu and T. Yahl , title =. Algebraic Statistics , year =. doi:10.2140/astat.2025.16.113 , eprint =

    work page doi:10.2140/astat.2025.16.113 2025

  11. [11]

    Fulton , title =

    W. Fulton , title =. 1998 , address =

    1998

  12. [12]

    Gavranovi

    B. Gavranovi. 2024 , url =

    2024

  13. [13]

    T. L. G. Proceedings of the Indian Academy of Sciences. Mathematical Sciences , year =. doi:10.1007/BF02829538 , eprint =

    work page doi:10.1007/bf02829538

  14. [14]

    Graziani , title =

    G. Graziani , title =. 2026 , eprint =

    2026

  15. [15]

    N. W. Henry and G. L. Marchetti and K. Kohn , title =. International Conference on Learning Representations , year =

  16. [16]

    Hollering and E

    B. Hollering and E. Mazzucchelli and M. Parisi and B. Sturmfels , title =. arXiv preprint arXiv:2511.21333 , year =

    arXiv

  17. [17]

    Kubjas and J

    K. Kubjas and J. Li and M. Wiesmann , title =. Algebraic Statistics , year =. doi:10.2140/astat.2024.15.295 , eprint =

    work page doi:10.2140/astat.2024.15.295 2024

  18. [18]

    Kohn and T

    K. Kohn and T. Merkh and G. Mont. SIAM Journal on Applied Algebra and Geometry , year =. doi:10.1137/21M1441183 , url =

    work page doi:10.1137/21m1441183

  19. [19]

    Kozhasov and A

    K. Kozhasov and A. Muniz and Y. Qi and L. Sodomaco , title =. Mathematics of Computation , year =. 2309.15105 , archivePrefix =

    arXiv

  20. [20]

    Kileel and M

    J. Kileel and M. Trager and J. Bruna , title =. 2019 , eprint =

    2019

  21. [21]

    J. M. Landsberg , title =. 2012 , address =

    2012

  22. [22]

    Massarenti and M

    A. Massarenti and M. Mella , title =. arXiv preprint arXiv:2511.19703 , year =

    arXiv

  23. [23]

    G. L. Marchetti and V. Shahverdi and S. Mereta and M. Trager and K. Kohn , title =. arXiv preprint arXiv:2501.18915 , year =

    arXiv

  24. [24]

    Mumford and J

    D. Mumford and J. Fogarty and F. Kirwan , title =

  25. [25]

    Piene , title =

    R. Piene , title =. Annales Scientifiques de l'

  26. [26]

    Pragacz , title =

    P. Pragacz , title =. Annales Scientifiques de l'

  27. [27]

    I. R. Shafarevich , title =. 1994 , address =

    1994

  28. [28]

    Shahverdi , title =

    V. Shahverdi , title =. arXiv preprint arXiv:2401.16613 , year =

    arXiv

  29. [29]

    Sonoda and N

    S. Sonoda and N. Murata , title =. Neural Networks , year =

  30. [30]

    2601.21645 , archivePrefix =

    2026 , note =. 2601.21645 , archivePrefix =

    Pith/arXiv arXiv 2026

  31. [31]

    A. J. Sommese and C. W. Wampler , title =. 2005 , address =

    2005

  32. [32]

    Watanabe , title =

    S. Watanabe , title =. 2009 , series =

    2009

  33. [33]

    Weyman , title =

    J. Weyman , title =. 2003 , address =

    2003

  34. [34]

    Shahverdi and G

    V. Shahverdi and G. L. Marchetti and K. Kohn , title =. 2026 , eprint =

    2026