pith. sign in

arxiv: 2502.08980 · v6 · pith:MPLQWPYXnew · submitted 2025-02-13 · 🧮 math.MG

Distinguishing finite metric spaces via similarity spectra

Jun O'Hara This is my paper

Pith reviewed 2026-05-23 03:55 UTC · model grok-4.3

classification 🧮 math.MG
keywords spectral invariantsfinite metric spacesq-spectrumtransition q-spectrumsimilarity matricesdistinguishing metric spacesgraph spectra
0
0 comments X

The pith

The q-spectrum from similarity matrices distinguishes all finite metric spaces on at most four points and a large class of larger ones.

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

The paper defines the q-spectrum and transition q-spectrum as invariants of finite metric spaces built from similarity matrices. These reduce to the adjacency and Laplacian spectra of graphs in the limit as q approaches zero. It establishes that the q-spectrum separates every metric space with four or fewer points along with a broad collection of larger spaces, while the transition q-spectrum separates spaces whose pairwise distances form a multiset linearly independent over the rationals and every space with three or fewer points. Numerical checks show the transition version often separates more examples in practice even though its proven range is narrower.

Core claim

The q-spectrum completely distinguishes a large class of finite metric spaces and all metric spaces on at most 4 points. The transition q-spectrum distinguishes spaces for which the multiset of pairwise distances is independent over the rational numbers, along with all spaces on at most 3 points.

What carries the argument

Similarity matrices of a finite metric space whose eigenvalues form the q-spectrum and transition q-spectrum.

If this is right

  • Every pair of distinct metric spaces on four or fewer points produces different q-spectra.
  • Spaces whose distance multisets are linearly independent over the rationals produce different transition q-spectra.
  • The new invariants reduce exactly to the classical graph spectra when the space is the vertex set of a graph and q tends to zero.

Where Pith is reading between the lines

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

  • The invariants could be computed directly from distance matrices to compare point sets without first embedding them into Euclidean space.
  • If the transition q-spectrum continues to separate most random examples, it may serve as a practical filter before more expensive isomorphism checks on larger point clouds.

Load-bearing premise

The similarity matrices are constructed so that their spectra encode enough metric information to separate the claimed classes.

What would settle it

Two non-isometric finite metric spaces on five points that share the same q-spectrum.

Figures

Figures reproduced from arXiv: 2502.08980 by Jun O'Hara.

Figure 1
Figure 1. Figure 1: (iii) Take t3 ∈ S3(a). Assume t3 = a + r + s, where r, s ∈ S1 \ {a, b, c, p, q}. We can determine the way how △(a, r, s) is attached to the edge BC in the same way as above. Let A3 be the vertex of △(a, r, s) which is different from B and C. Then the length AA3 is given by the unique element u ∈ S1 \ {a, b, c, p, q, r, s} that satisfies c+r+u ∈ S3 or b+r+u ∈ S3 depending on the way how △(a, r, s) is attach… view at source ↗
Figure 2
Figure 2. Figure 2: and the lengths of 4-cycles that consist of two single 2-cycles are given by 2α, 2β and 2γ. Therefore τ4(q), the constant term of ¯pX(q; µ), is given by −2q α+β − 2q β+γ − 2q α+γ + q 2α + q 2β + q 2γ . (2.5) There are following four cases (i) - (iv) according to the types of τ4(q). (1-i) τ4(q) is of the form −2 P3 i=1 q Ai + P3 i=j q Bj , where Ai , Bj are mutually distinct. This case can occur if and only… view at source ↗
Figure 3
Figure 3. Figure 3: Suppose S3(X1) = S3(X2). There are 24 possibilities for the system of linear equations obtained from S3(X1) = S3(X2), each of which implies that X1 is isometric to X2. To be precise, for each permutation σ ∈ S4 of four letters {1, 2, 3, 4}, a system of linear equations    ϕ1(a, b, c, d, f, g) = ψσ(1)(a, b, c, d, f, g), ϕ2(a, b, c, d, f, g) = ψσ(2)(a, b, c, d, f, g), ϕ3(a, b, c, d, f, g) = ψσ(3)(a, b… view at source ↗
Figure 4
Figure 4. Figure 4: X X a￾b e c+e a b+e c a￾b e c+e b+e c a 11 12 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: • The system of linear equations ϕ1 = ψ1, ϕ2 = ψ3, ϕ3 = ψ2, ϕ4 = ψ4, i.e.    2a + c = 2a + b, a + b + c − ε = a + 2c + ε, a + 2b + ε = a + b + c − ε, b + 2c + 2ε = 2b + c + 2ε has a solution b = c and ε = 0, which contradicts our assumption that ε 6= 0. ✷ Since there is a pair of non-isometric four-point spaces with the same magnitude (homology groups) as illustrated in [PITH_FULL_IMAGE:figures/ful… view at source ↗
Figure 7
Figure 7. Figure 7: Two wedge sums of •−•−• and •−• Two graphs in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Two wedge sums of three cycle graphs [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Metric fibration The characteristic polynomials of the two graphs in [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Two graphs that differ by a Whitney twist [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Another example of Whitney twist The difference of the coefficients of λ 8 comes from the difference of the number of pairs of points with distance 4 and 5. (4) The notion of isomer was introduced in [9] to construct a space that has the same magnitude with, but is not isometric with the regular n-gon or an n-cycle graph. To be precise, an isomer of a regular n-gon can be constructed for n = 6 and n ≥ 8, … view at source ↗
Figure 12
Figure 12. Figure 12: Thin single lines represent length d1, double lines represent length d2, and dotted lines represent length d3. They represent a regular hexagon R6 and its isomer R′ 6 when (d1, d2, d3) = (1, √ 3, 2), and a 6-cycle graph C6 and its isomer C ′ 6 when (d1, d2, d3) = (1, 2, 3). Regular hexagon and its isomer have different characteristic polynomials since they have different maximal lengths of 3-cycles. In fa… view at source ↗
Figure 13
Figure 13. Figure 13: Thin single lines, double lines, dotted lines and thick lines repr [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Direct computation shows mΣ1 (q) = mΣ2 (q) = mΣ3 (q) 6= mΣ4 (q) 6= mΣ5 (q) = mΣ6 (q) 6= mΣ1 (q), and the characteristic polynomials of ZtΣi are all mutually distinct. In the case of Σ1 and Σ2 the difference in the characteristic polynomials appears first in the coefficients of λ 9 , otherwise in the coefficients of λ 10 . 4 Cospectral graphs Two non-isomorphic graphs are said to be cospectral or isospectr… view at source ↗
Figure 15
Figure 15. Figure 15: different characteristic polynomials of similarity matrices and different magnitudes. On the other hand, [PITH_FULL_IMAGE:figures/full_fig_p013_15.png] view at source ↗
Figure 17
Figure 17. Figure 17 [PITH_FULL_IMAGE:figures/full_fig_p013_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Isospectral enneahedra (picture taken from wiki) [PITH_FULL_IMAGE:figures/full_fig_p013_18.png] view at source ↗
read the original abstract

We study spectra and characteristic polynomials of similarity matrices associated with finite metric spaces, where the similarity matrix of a finite metric space $X=\{x_1,\dots,x_n\}$ is given by $\displaystyle Z(q)=(q^{d(x_i,x_j)})_{i,j},$ where $d(x_i,x_j)$ denotes the distance between $x_i$ and $x_j$. % We introduce two spectral invariants of finite metric spaces, the $q$-spectrum and the normalized $q$-spectrum, defined respectively from $Z(q)$ and its normalized transition matrix. In the case of graphs, these invariants recover the adjacency spectrum and the Laplacian spectrum in the limit $q\to0$. Our main result shows that the $q$-spectrum determines a large class of finite metric spaces under a natural nondegeneracy condition. We also prove that all four-point metric spaces are determined by their $q$-spectra. % The key observation is that the coefficients of the characteristic polynomial of $Z(q)$ encode cycle structures of the underlying metric space. We further investigate the normalized $q$-spectrum and present computational examples comparing these invariants with classical graph spectra.

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

3 major / 2 minor

Summary. The paper introduces the q-spectrum and transition q-spectrum as spectral invariants for finite metric spaces, constructed from parameterized similarity matrices that generalize the adjacency and Laplacian spectra of graphs (recovered in the q→0 limit). It proves that the q-spectrum distinguishes all metric spaces on at most 4 points and a large unspecified class of spaces, while the transition q-spectrum distinguishes all spaces on at most 3 points and those whose multiset of pairwise distances is linearly independent over the rationals. Computational experiments are reported indicating that the transition q-spectrum often exhibits stronger distinguishing power in practice.

Significance. If the distinction theorems hold, the work supplies new, computable invariants for classifying finite metric spaces that extend classical graph spectra in a natural way. The explicit results for small cardinalities (≤3 and ≤4 points) and the rational-independence condition are concrete and falsifiable, while the computational observations provide empirical evidence of practical utility. These features would make the invariants potentially useful in metric geometry and related areas, though the scope is limited to the classes where distinction is proved.

major comments (3)
  1. [§2] §2 (Definition of the similarity matrix and q-spectrum): The distinction claims in the main theorems rest on the assumption that the eigenvalues of the specific q-parameterized similarity matrix encode enough metric information to separate non-isometric spaces. No independent argument or comparison is supplied showing why this matrix construction (as opposed to, e.g., the spectrum of the distance matrix itself or other kernels) is informationally sufficient for the stated classes.
  2. [Theorem on q-spectrum for ≤4 points] Theorem on q-spectrum distinguishing all 4-point spaces: The proof appears to proceed by case analysis on configurations, but it inherits the unverified encoding assumption from the matrix definition; without an explicit check that the eigenvalues separate distances in a manner guaranteed by the construction (rather than by enumeration alone), the result remains dependent on the matrix choice.
  3. [Theorem on transition q-spectrum for rational independence] Theorem on transition q-spectrum for rationally independent distances: The separation is claimed for spaces whose distance multisets are Q-linearly independent, yet the argument does not include a verification that the transition matrix eigenvalues detect this independence independently of the matrix parameterization; this is load-bearing for the theorem's scope.
minor comments (2)
  1. [§2] The precise functional form of the similarity matrix entries (involving q) is introduced without an accompanying small example computing both the q-spectrum and the classical graph spectra in the limit, which would aid readability.
  2. [Computational experiments section] The computational experiments section would benefit from explicit reporting of the range of q values sampled and the numerical method used to compute the transition spectrum.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading of our manuscript and the insightful comments. We respond to each major comment in turn.

read point-by-point responses

  1. Referee: [§2] §2 (Definition of the similarity matrix and q-spectrum): The distinction claims in the main theorems rest on the assumption that the eigenvalues of the specific q-parameterized similarity matrix encode enough metric information to separate non-isometric spaces. No independent argument or comparison is supplied showing why this matrix construction (as opposed to, e.g., the spectrum of the distance matrix itself or other kernels) is informationally sufficient for the stated classes.

    Authors: The q-parameterized similarity matrix is specifically constructed to extend the adjacency and Laplacian matrices of graphs, with the limit q → 0 recovering the classical graph spectra. This provides the foundational motivation for the choice. The distinction theorems then demonstrate that this construction is sufficient for the classes considered, via direct proofs. While we do not provide an exhaustive comparison to alternative kernels such as the distance matrix spectrum, the natural generalization from the graph case justifies the construction. If the editor deems a brief discussion of alternatives beneficial, we can include it. revision: no

  2. Referee: [Theorem on q-spectrum for ≤4 points] Theorem on q-spectrum distinguishing all 4-point spaces: The proof appears to proceed by case analysis on configurations, but it inherits the unverified encoding assumption from the matrix definition; without an explicit check that the eigenvalues separate distances in a manner guaranteed by the construction (rather than by enumeration alone), the result remains dependent on the matrix choice.

    Authors: For the finite case of at most 4 points, the proof consists of a complete case analysis: we classify all non-isometric metric spaces on 4 points, compute their q-spectra explicitly as functions of q, and verify that distinct spaces yield distinct spectra. This enumeration constitutes the explicit check that the eigenvalues distinguish the spaces under the given construction. The result is therefore not merely dependent on the matrix choice but verified by direct computation for this small cardinality. revision: no

  3. Referee: [Theorem on transition q-spectrum for rational independence] Theorem on transition q-spectrum for rationally independent distances: The separation is claimed for spaces whose distance multisets are Q-linearly independent, yet the argument does not include a verification that the transition matrix eigenvalues detect this independence independently of the matrix parameterization; this is load-bearing for the theorem's scope.

    Authors: The proof for the transition q-spectrum under rational independence proceeds by considering the characteristic polynomial of the transition matrix. Under the assumption that the distances are linearly independent over Q, the coefficients involve distinct monomials, ensuring that the eigenvalues uniquely recover the distance multiset. This algebraic verification directly ties the detection to the independence condition and the matrix parameterization. We can expand this explanation if the current presentation is deemed insufficiently clear. revision: partial

Circularity Check

0 steps flagged

No significant circularity; definitions and distinction theorems are independent

full rationale

The paper introduces the q-spectrum and transition q-spectrum as explicit constructions from similarity matrices on finite metric spaces, notes their relation to graph spectra in the q→0 limit, and then states theorems asserting that these invariants distinguish all spaces on ≤4 points (for q-spectrum) and spaces with rationally independent distance multisets plus all spaces on ≤3 points (for transition q-spectrum). No equation or claim reduces a distinction result to a definitional identity, a fitted parameter renamed as a prediction, or a self-citation chain; the distinguishing power is presented as a theorem to be proved rather than an input. The construction is self-contained and externally falsifiable by direct computation on metric spaces.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central contribution consists of new definitions of similarity matrices and their spectra; the paper therefore adds these definitions rather than deriving them from prior axioms. No free parameters or invented entities are mentioned in the abstract.

axioms (1)
  • standard math Standard properties of finite metric spaces and matrix spectra
    Invoked implicitly when defining similarity matrices and taking limits as q to 0.

pith-pipeline@v0.9.0 · 5653 in / 1258 out tokens · 47900 ms · 2026-05-23T03:55:00.584443+00:00 · methodology

discussion (0)

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