Pith. sign in

REVIEW 3 major objections 5 minor 2 references

Ubiquitous Symmetry at Critical Points Across Diverse Optimization Landscapes

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read The paper claims that every critical point observed in four new optimization settings — projective spaces over finite fields, octahedral and perfect matching graphs, and particle attraction — has non-trivial symmetry, extending the…

desk verdict A plausible extension of the least-symmetry-breaking program to four new landscapes, with a genuinely new symmetry measure, but the 'always' claims outrun the trajectory-based evidence and no artifacts are provided. read the letter →

arxiv 2506.01959 v1 pith:T4KRSFRY submitted 2025-05-04 cs.LG cs.AIphysics.atom-ph

classification cs.LGcs.AIphysics.atom-ph
keywords symmetrybreakinglosslandscapecriticalpointsisotropygroupstabilizergradientdescentNewton'smethodedge
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

This paper claims that the 'least symmetry breaking' principle observed in neural-network and tensor-decomposition loss landscapes holds in four further optimization settings: real-valued functions on projective spaces over finite fields, vertex labellings of the octahedral graph, vertex labellings of a perfect matching graph, and systems of attracting particles in the plane. In every reported run, gradient descent and Newton's method converge to critical points with non-trivial stabilizer subgroups of the loss's symmetry group. The paper introduces the edge isotropy group, which tracks symmetry of pairwise kernel interactions and stays non-trivial even in the few configurations where vertex-level isotropy is trivial. If the claim is right, symmetry at critical points is a general feature of kernel-based pairwise optimization landscapes, and any non-symmetric minimum must occupy a basin so narrow that tens of millions of random starts cannot find it.

What carries the argument

The carrying mechanism is pairwise-kernel loss functions together with stabilizer analysis. Each loss is built as a sum over pairs of kernel evaluations — on hyperplane restrictions, on directed graph edges, or on particle positions — so the loss is automatically invariant under a known symmetry group, and the object of study is the stabilizer of a critical point inside that group. Two stabilizers are used: the vertex isotropy group IV, which fixes the values attached to vertices, and the new edge isotropy group IE, defined as all permutations that preserve every kernel value between edge pairs (or particle pairs), with IV always a subgroup of IE. The pattern the paper reports is that these stabilizers are consistently non-trivial, and IE reveals symmetry when IV does not.

What would settle it

Run an exhaustive or very large random search on any one of the four systems and exhibit a critical point with trivial stabilizer. A concrete target: in the four-particle attraction model with kernel $\|a-b\|^{12}-\|a-b\|^8$, the paper found two points whose vertex isotropy is trivial but whose edge isotropy is $C_2$ and $S_3$; finding a critical point with trivial edge isotropy there, or any non-symmetric critical point in the projective-space case, would refute the ubiquity claim.

Watch

Extended reading notes

Core claim

The paper's central discovery, stated as a uniform empirical law, is that every critical point found in all four settings has non-trivial symmetry. For the projective case, the loss is a sum of powers of inner products between restrictions of functions to hyperplanes, invariant under PGL(n+1,F_q), and every minimum found has a stabilizer containing a substantial subgroup that fixes a projective hyperplane. For the octahedral graph, the perfect matching graph, and the particle system, the loss is a sum of kernel values over pairs of directed edges or particle pairs, invariant under the relevant automorphism group, and again every observed critical point has non-trivial isotropy. In the particle case with four particles, two critical points have trivial vertex isotropy but retain non-trivial edge isotropy, which motivates the paper's new measure IE. Across a 30-million-run experiment with a degree-16 polynomial kernel, only the previously identified symmetric minima were reached.

Load-bearing premise

The load-bearing premise is that the finite collection of optimization runs is representative of the entire landscape: if a non-symmetric critical point exists but was never sampled by the chosen initializations and optimizers, the ubiquity conclusion collapses.

Editorial extensions

If this is right

  • If a non-symmetric minimum exists in any of these landscapes, its basin of attraction is so small that it was missed in roughly 30 million runs, making symmetric minima the practically relevant solutions.
  • The least-symmetry-breaking principle extends: critical points of kernel-based pairwise losses tend to sit at stabilizer subgroups that are substantial subgroups of the loss's symmetry group.
  • The edge isotropy group IE is a finer invariant than IV: it stays non-trivial in configurations where IV is trivial, including kernels with repulsive terms deliberately chosen to kill vertex symmetry.
  • The phenomenon is robust to the kernel's form: distance-based, inner-product-based, high-degree polynomial, and randomly chosen bounded-below polynomial kernels all produced symmetric critical points.

Reading between the lines

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

  • One could exploit the implied symmetry by quotienting the search space by the symmetry group before optimization; if all observable minima are symmetric, the reduced search space would contain representatives of every basin that matters.
  • The IE measure suggests a hierarchical view: vertex-level symmetry breaks first, while edge-level pairwise symmetry persists; an analogous hierarchy might exist for higher-order interactions, and testing IE on triples or hyperedges would be a natural next experiment.
  • If the claim generalizes, it would offer a practical certification heuristic: for kernel-based pairwise losses, one may check only symmetric candidates when searching for local minima, and approximate basin volume by the size of the stabilizer.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper reports an empirical study of symmetry at critical points of real-valued loss functions in four optimization settings: functions on projective space over finite fields, the octahedral graph, the perfect matching graph, and a particle attraction model. For each setting, the author runs gradient descent and Newton's method from random initializations, records the critical points found, and computes vertex isotropy groups IV under the natural symmetry group of the problem. The central claim is that all observed critical points have non-trivial symmetry, extending prior observations in ReLU networks and tensor decomposition; the paper also introduces a new 'edge isotropy group' IE intended to capture symmetries not visible from IV, and presents evidence that IE is often larger than IV, including cases where IV is trivial.

Significance. If the central claim were fully established, the paper would provide a meaningful extension of the 'least symmetry breaking' principle to several new landscapes and introduce a useful diagnostic, IE, for detecting hidden symmetries at critical points. The author deserves credit for reporting detailed tables of critical points, for testing multiple kernel families, and for being explicit in Section 5.4 that non-symmetric minima cannot be theoretically ruled out. The 30-million-run experiment in Appendix D is an unusually large sampling effort. However, the paper is purely empirical and the strength of the stated conclusion currently exceeds the evidence: all experiments are trajectory-based, and the manuscript itself concedes that non-symmetric minima with very narrow basins would be undetectable by this protocol. Since no code or data are provided, the reproducibility of the tables is also limited. The contribution is therefore a promising empirical observation rather than a demonstrated universal phenomenon.

major comments (3)
  1. [Section 5.4 / Abstract / Section 7] The abstract's 'all the critical points observed have non-trivial symmetry' is carefully qualified, but the Summary in Section 7 states that 'both gradient descent and Newton's method always converge to local minima that exhibit substantial non-trivial symmetry,' which is stronger than the evidence. Every experiment is trajectory-based: random initialization followed by an optimizer. A non-symmetric critical point that is a repeller, a saddle, or a minimum with measure-zero basin is invisible to this protocol regardless of run count. Section 5.4 explicitly concedes this, saying non-symmetric minima 'if they do exist their basins of attraction must be extraordinarily narrow.' To make the 'always' claim load-bearing, the paper needs either an exhaustive algebraic enumeration of stationary solutions of ∇L=0 in at least one small case (e.g., the octahedral graph or perfect matching graph with fixed kernel parameters) or a rigorous basin-volume/sampling argument. In the meantime, the conclusions should be restricted to 'all observed critical points'.
  2. [Section 5.2 / Section 6.3 / Tables 4 and 6] There is an internal inconsistency in the identification of which particle-attraction critical points have trivial IV. In Table 4, rows 4 and 5 have IV order 1, while row 6 has IV group C2 of order 2; yet Section 5.2 says 'the only points (number 5, 6) with trivial symmetry,' and Section 6.3 repeats 'for the 5, 6 critical points.' Additionally, Table 6 reorders the rows relative to Table 4: row 3 in Table 6 has loss 0 and group S4, whereas row 3 in Table 4 has loss -0.88889 and group S3, with the loss-0 point appearing as row 8 in Table 4. Because the paper's demonstration that IE exceeds IV depends on correctly identifying which points have trivial IV, these discrepancies must be corrected and reconciled.
  3. [Sections 2-5 and Appendix D] The experimental protocol is not described sufficiently for reproducibility. The paper does not provide code, data, initialization distributions, learning rates, convergence tolerances, or the number of runs per configuration; Section 5.4 mentions 'roughly 1,000 trials per case' but no per-case table is given. The 30-million-run experiment in Appendix D uses one loss, one optimizer, and reports only the four attracting minima found, not an enumeration of stationary points. I recommend adding a detailed protocol section and, if possible, releasing code and data, so that the tables in Sections 2-6 can be independently checked and the sampling claim can be evaluated.
minor comments (5)
  1. [Section 6.4] The kernel κ(a,b)=∥a−b∥^2 + 1/∥a−b∥^2 is undefined at a=b; the paper defines the value to be 0 there, but this creates a discontinuity that should be discussed, especially since the loss sums over i=j terms.
  2. [Section 5.3] The claim that inner-product kernels raised to the 50th power produce critical points with non-trivial symmetry is stated without any table or numerical data; either add the results or mark the claim as preliminary.
  3. [Tables 5 and 7] Group names such as 'S2 3' and 'C2 2' are nonstandard notation; consider using S3^2, C2^2, or explicit direct products to avoid ambiguity.
  4. [Abstract] The phrase 'We will introduce four more cases' is informal and should be revised to present tense or a statement of results.
  5. [Appendix E] The random polynomial experiments are described with the coefficients listed, but the mapping between the five experiments and the five polynomial expressions should be made explicit to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the symmetry findings are empirical observations, not derived from the symmetry measure or from self-citation.

full rationale

The manuscript makes no closed-form derivation; its contribution is a set of direct optimization experiments followed by stabilizer computations. The 'ubiquitous symmetry' claim is an empirical generalization ('all critical points observed'), explicitly bounded in §5.4 by the admission that non-symmetric minima may exist with narrow basins. No parameter is fitted to a subset and then 'predicted' on a related quantity. The cited prior work (Arjevani–Field 2019; Arjevani et al. 2021) is by other authors and is used as background, not as a load-bearing uniqueness theorem. The edge-isotropy measure IE is defined independently as the kernel-value-preserving automorphism group; the relation IV⊂IE is stated as an obvious definitional fact, and the experimental content is the observation that IE exceeds IV at the found critical points. That observation does not reduce to the definition by construction, since it is in principle possible for all found critical points to have trivial IE. The internal presentation tension (Table 4 contains IV-trivial critical points while the Abstract says all observed critical points have non-trivial symmetry) is resolved by the paper's own later IE analysis and concerns consistency of exposition, not circular derivation. The sampling limitation is a correctness/robustness caveat, not a circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 1 invented entities

The paper's central claim rests on the constructed loss functions (kernel sums over group actions or edge pairs), on the implicit assumption that gradient descent and Newton's method from random starts sample all critical point types, and on standard group-theoretic computations. The kernels and constants (c=7, various degrees) are chosen by hand, and the paper itself notes in Section 5.3 that robustness 'needs more experiments.' The edge isotropy group is a new definition rather than an independently evidenced entity.

free parameters (5)
  • c in octahedral kernel = 7
    Hand-picked constant in kernel κ(a,b) = <a,b>^p - c <a,b>^q (Section 3.1); not fitted but chosen by the author.
  • p, q in octahedral kernel = p=6, q=4
    Hand-picked polynomial degrees in the octahedral kernel (Section 3.1).
  • c in reformulated projective loss = unspecified c > 0
    A positive constant introduced in Section 2.4 without a specified value; the behavior likely depends on it.
  • c in perfect matching kernel = unspecified
    Constant in κ(a,b) = ||a-b||^{2p} - c ||a-b||^{2q} (Section 4.1), not specified.
  • kernel degree choices = various (e.g., d12-d8, d2+1/d2, inner product powers 8 and 50)
    The specific kernels are chosen by hand; the paper notes in Section 5.3 that robustness 'needs more experiments'.
assumptions (3)
  • domain assumption Gradient descent and Newton's method from random initializations explore the loss landscape sufficiently to find representatives of all critical point types.
    The empirical claim that no asymmetric critical points were found rests on the search being representative; the paper itself flags this in Section 5.4.
  • standard math Standard group actions and stabilizer computations are correct (PGL(n+1,F_q) action, Aut(G) actions).
    The symmetry computations rely on standard group theory, not proved in the paper.
  • ad hoc to paper The chosen loss functions are representative of optimization landscapes in general.
    The losses are constructed ad hoc (kernel sums over group orbits or edge pairs), and there is no argument that they are generic.
invented entities (1)
  • Edge isotropy group I_E
    purpose: A new symmetry measure that records permutations preserving kernel interaction values between all edge pairs, intended to capture symmetries invisible to the vertex isotropy group I_V.
    It is a well-defined mathematical construction applicable to any graph configuration, but it makes no external predictions and is demonstrated only on the same experiments used to motivate it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ubiquitous Symmetry at Critical Points Across Diverse Optimization Landscapes." pith.science (2026). https://pith.science/paper/T4KRSFRY

@misc{pith2026250601959,
  author       = {Pith},
  title        = {Pith review of: Ubiquitous Symmetry at Critical Points Across Diverse Optimization Landscapes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T4KRSFRY}},
  note         = {Machine review of arXiv:2506.01959}
}
read the original abstract

Symmetry plays a crucial role in understanding the properties of mathematical structures and optimization problems. Recent work has explored this phenomenon in the context of neural networks, where the loss function is invariant under column and row permutations of the network weights. It has been observed that local minima exhibit significant symmetry with respect to the network weights (invariance to row and column permutations). And moreover no critical point was found that lacked symmetry. We extend this line of inquiry by investigating symmetry phenomena in real-valued loss functions defined on a broader class of spaces. We will introduce four more cases: the projective case over a finite field, the octahedral graph case, the perfect matching case, and the particle attraction case. We show that as in the neural network case, all the critical points observed have non-trivial symmetry. Finally we introduce a new measure of symmetry in the system and show that it reveals additional symmetry structures not captured by the previous measure.

Figures

Figures reproduced from arXiv: 2506.01959 by the authors.

Figure 1
Figure 1. Typical minima observed for ReLU networks with an identity target matrix. The stabilizer under [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Octahedral Graph We study the octahedral graph, a 6-vertex, 12-edge undirected graph representing an octahedron. Let G = (V, E) be a graph with vertex set V = {0, 1, 2, 3, 4, 5}, 7 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Perfect Matching Graph with n = 3 We now shift our focus from the octahedral graph to the perfect matching graph. For vectors a, b ∈ R 2 , we define the kernel function as κ(a, b) = ∥a − b∥ 2p − c ∥a − b∥ 2q , where p and q are parameters and c is a constant. The loss function L is then obtained by summing the kernel over all directed edges (i, j) and (k, l) in the edge set E: L = X (i,j)∈E X (k,l)∈E κ(aij , akl). 4… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualization of critical points from Table [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 13
Figure 13. Figure 13: Visualization of critical points from Table [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 44
Figure 44. Figure 44: Visualization of critical points from Table [PITH_FULL_IMAGE:figures/full_fig_p025_44.png]
Figure 48
Figure 48. Figure 48: Visualization of critical points from Table [PITH_FULL_IMAGE:figures/full_fig_p026_48.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

  1. [2019]

    On the Principle of Least Symmetry Breaking in Shallow ReLU Models

    URL https://arxiv.org/pdf/1912.11939. Yossi Arjevani, Joan Bruna, Michael Field, Joe Kileel, Matthew Trager, and Francis Williams. Symmetry breaking in symmetric tensor decomposition. arXiv preprint arXiv:2103.06234 ,

  2. [2021]

    URL https:// arxiv.org/pdf/2103.06234. 31

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.