Spectral Coupling Matrix: A New Descriptor for Transport-Information Geometry

Read full text Summarise with VIRAT
Authors: Frederic David Blum
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Spectral Coupling Matrix: A New Descriptor for Transport-Information Geometry Frédéric David Blum Catalyst AI Research, Tel Aviv freddavidblum@catalystais.com | ORCID: 0009-0009-2487-2974 Date: March 22, 2026 Abstract We report the discovery and numerical characterization of the Spectral Coupling Matrix (SCM), a new mathematical object that decomposes the coupling between optimal transport and information divergence on a mode-by-mode spectral basis. Starting from a conjectured local relation between the Ricci curvature, the Hessian of the Brenier potential, and the Kullback-Leibler divergence (the "DRT conjecture"), systematic falsification on the 2-sphere S² and the flat torus T² revealed: (i) the conjecture is false in its original form (wrong sign, no universal constant), but (ii) a robust spectral anti-correlation exists between the heat-filtered Laplacian of the Kantorovich potential and the KL density. This anti-correlation is independent of Ricci curvature (persists on the flat torus where Ric ≡ 0), specific to structured optimal transport (disappears for random transport), and strongest in the non-perturbative regime. The scalar integral of this anti-correlation (the "STK kernel") collapses to a linear combination of W₂² and D_KL (R² = 0.989), losing essentially all unique information. In contrast, the SCM — which preserves the spectral decomposition — reveals: (a) a universal anti-alignment of transport and information in the first spectral bands (normalized coupling ~ −0.93 with 96% sign consistency), (b) 11/12 spectral bands contain >10% information not reducible to (W₂, D_KL), and (c) discrimination scores 460× between pairs indistinguishable by classical metrics. The SCM constitutes a new spectral fingerprint of the transport-information coupling at each spatial scale. 1. Introduction and Motivation The relationship between optimal transport (Monge-Kantorovich theory) and information geometry (KL divergence, Fisher metric) is a central theme in modern mathematics. The Lott-Villani-Sturm theory establishes that lower bounds on Ricci curvature imply convexity of entropy along Wasserstein geodesics — a global, macroscopic implication. We asked a different question: is there a local, microscopic relation between the transport potential and the information divergence? Specifically, the DRT conjecture (generated by the Catalyst AI ideational extraction pipeline) proposed: ⟨Ric_g, F_h[∇²φ]⟩_g ≈ C · D_KL(μ₀ ∥ μ₁) where φ is the Brenier potential, F_h = exp(−hΔ) is the heat filter, and C is a constant depending on (M, g). Rather than attempting a proof, we adopted a falsification-first methodology: implement the conjecture numerically, test it, and let the data speak. 2. Phase 1: Falsification on S² 2.1 Setup We discretized S² using a Fibonacci lattice (N = 800 points), computed geodesic distances, solved entropic optimal transport via Sinkhorn (regularization ε_OT = 0.005), projected the Kantorovich potential onto real spherical harmonics (l_max = 15), and applied the spectral operator: e^{−hΔ}[Δφ] = Σ_{l,m} [−l(l+1)] · exp(−h · l(l+1)) · a_{lm} · Y_{lm} 2.2 Results The conjecture is falsified on three counts: Wrong sign: All correlations between e^{−hΔ}[Δφ] and D_KL density are negative (best: r = −0.94). Wrong regime: The strongest anti-correlation occurs at ε = 2.0 (non-perturbative), not at ε → 0 as predicted. No universal constant: C depends on h, ε, and N (40% variation between N = 400 and N = 800). Additionally, Stokes' theorem guarantees ∫_{S²} e^{−hΔ}[Δφ] dVol → 0, making any global proportionality with D_KL > 0 impossible. 3. Phase 2: The Flat Torus Test 3.1 Rationale On S², Ric = g, making it impossible to distinguish the role of curvature from the metric itself. The flat torus T² = [0, 2π)² has Ric ≡ 0, providing the definitive test. 3.2 Method On T², the spectral decomposition is exact via FFT: Eigenvalues: λ_{m,n} = m² + n² Eigenfunctions: exp(i(mθ₁ + nθ₂)) Heat-filtered Laplacian: computed exactly in Fourier space 3.3 Key Results Property S² (Ric = g) T² (Ric = 0) Best |r| 0.94 0.975 ∫ e^{−hΔ}[Δφ] dVol ~10⁻⁴ (numerical) 0 (exact) Anti-correlation persists? Yes Yes (stronger) Conclusion: The anti-correlation is independent of Ricci curvature. The mechanism is purely spectral. 3.4 Control Tests Structured vs random transport: Structured r = −0.956, Random r = −0.369 ± 0.015. Difference: 38 standard deviations. Variance vs W₂: Var ∝ W₂^{2.35} (r = 0.998). The "tighter regulation" hypothesis is falsified; variance increases with transport cost. 4. The STK Scalar Kernel (Negative Result) We defined the Spectral Transport Kernel: K_h(μ₀, μ₁) = −∫_M e^{−hΔ}Δφ · log(dμ₀/dμ₁)(x) dVol(x) Testing on 12 measures on T²: Not positive definite (11/12 eigenvalues of Gram matrix negative) R² = 0.989 with α·W₂² + β·D_KL: only 1.1% unique information The integration destroys the spectral structure Lesson: The interesting information lives in the spectral profile, not in its integral. 5. The Spectral Coupling Matrix (Positive Result) 5.1 Definition Given measures μ₀, μ₁ on a compact Riemannian manifold (M, g) with Laplacian Δ, let φ be the Kantorovich potential for the transport μ₀ → μ₁. Define: Spectral coefficients: â_{m,n} = Fourier coefficients of e^{−hΔ}[Δφ] (transport spectrum) b̂_{m,n} = Fourier coefficients of log(dμ₀/dμ₁) (information spectrum) Mode-to-mode coupling: C_{m,n} = conj(â_{m,n}) · b̂_{m,n} Band-aggregated coupling (grouping by eigenvalue λ_k = m² + n²): S_k(μ₀, μ₁) = Σ_{(m,n) ∈ band_k} Re(conj(â_{m,n}) · b̂_{m,n}) Normalized spectral coupling (spectral cosine): C̃_k(μ₀, μ₁) = S_k / √(Σ|â|² · Σ|b̂|²) (within band k) The vector (C̃₁, C̃₂, ..., C̃_K) is the spectral fingerprint of the transport-information coupling. 5.2 Properties Discovered P1: Universal anti-alignment in low frequencies Band λ Mean C̃ Std Sign consistency 1 1 −0.931 0.249 95.6% 2 2 −0.927 0.326 95.6% 3 4 −0.938 0.161 95.6% 4 5 −0.823 0.287 100% 5 8 −0.696 0.486 82.2% 6 9 −0.541 0.642 73.3% In the first 4 spectral bands, the transport and information spectra are anti-aligned at ~93%, universally across all tested pairs. P2: Spectral concentration 80% of the total coupling energy is contained in the first ~4 bands, for all pairs tested. This concentration is an invariant of the underlying manifold topology. P3: Information beyond W₂ and D_KL Per-band regression against (W₂², D_KL) shows: Band 1: R² = 0.87 → 13% unique information Band 2: R² = 0.28 → 72% unique information Bands 3-10: R² < 0.16 → >84% unique information 11 out of 12 bands contain >10% information not reducible to the classical metrics. P4: Discrimination power Pairs that are nearly identical in (W₂, D_KL) space (distance < 2%) can differ by factors of 460 in their SCM profiles. The SCM sees geometric structure that classical metrics miss. P5: Frequency scaling |C̃| ∝ λ^{−0.107} — the coupling is quasi-uniform across frequencies, with a very mild low-frequency bias. All spatial scales contribute. 6. Interpretation The SCM reveals that optimal transport, when moving mass coherently, produces a potential whose spectral content is systematically anti-aligned with the spectral content of the information gain. This anti-alignment: Is not a consequence of Ricci curvature (persists on flat torus) Is not captured by W₂ or D_KL individually or jointly Is specific to optimal transport (random transport gives |r| ~ 0.37) Is concentrated in low-frequency modes but present at all scales Is universal across different measure pairs (96% sign consistency) We hypothesize this reflects a spectral complementarity principle: the optimal transport plan, by minimizing Wasserstein cost, necessarily produces a potential whose large-scale spectral content opposes the information gradient. Intuitively, efficient transport "works against" information accumulation at each spectral scale. 7. Relation to Existing Work Lott-Villani-Sturm: Global implication (Ric ≥ K → entropy convexity). The SCM is local and does not require curvature bounds. Otto calculus: Connects Wasserstein geometry to diffusion PDE. The SCM adds the spectral decomposition of this connection. Benamou-Brenier: Dynamic formulation of OT. The SCM could potentially be extended to the time-resolved coupling along geodesics. Wasserstein kernels (Bachoc et al.): Kernels on probability measures using W₂. The SCM provides a richer descriptor that captures spectral structure. 8. Code Availability All code is available at: github.com/davidangularme Three Python scripts: drt_phase1_sphere.py — Falsification on S² (spherical harmonics, Sinkhorn OT) drt_phase2_torus.py — Flat torus test, variance/efficiency, structured vs random control scm_analysis.py — Full SCM computation, universality, PCA, discrimination tests Dependencies: numpy, scipy, POT (Python Optimal Transport), matplotlib 9. Conclusion Starting from a false conjecture (the DRT equation), systematic numerical falsification led to the discovery of a genuine mathematical structure: the Spectral Coupling Matrix. The SCM decomposes the coupling between optimal transport and information divergence at each spectral scale, revealing universal anti-alignment in low frequencies that is invisible to classical metrics. It constitutes a new descriptor in transport-information geometry, with potential applications in kernel methods on probability spaces, generative model evaluation, and the spectral theory of optimal transport. The key methodological insight: integration destroys structure. The scalar kernel (STK) collapsed to W₂ + D_KL because it summed over modes. Keeping the spectral decomposition intact — the SCM — preserved the unique information. Acknowledgments: The initial DRT conjecture was generated by the Catalyst AI ideational extraction pipeline using a falsification-first prompt methodology. All

Fields of Study & Subjects

Geometric Analysis and Curvature Flows Advanced Differential Geometry Research Topological and Geometric Data Analysis

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
22 March 2026
Language
EN
Document type
Preprint