Date: 2026-05-28 Status: Internal synthesis — not for publication. This is the real insight.
We built a Tension-Graph Laplacian that shows 112× conservation in music, works for parsing, MoE routing, format analysis, and fails for the Ising model.
We stumbled onto something much more fundamental than "music has conservation." We found the spectral fingerprint of coherent structure — and its absence is equally informative.
Every system we studied has two things:
- A transition graph (states → states with probabilities)
- Attributes on states (tension, tension, cosine similarity, entropy)
When we combine them as W[i,j] = P(i→j) × similarity(i,j) and look at L = D - W, we're not just "building a Laplacian." We're constructing a compatibility operator that measures how well the dynamics (P) align with the structure (similarity).
The eigenvectors of this operator decompose the system into modes ranked by how well dynamics and structure agree:
- Low eigenvalues (near 0): modes where dynamics and structure are ALIGNED — the system "wants" to be here
- High eigenvalues: modes where dynamics and structure CONFLICT — the system resists these directions
The conservation ratio CR(k) = ∇²f|_{φ_k} measures how smooth an attribute is along each mode. Low CR = the attribute is naturally conserved in that mode = the system is well-adapted to that attribute in that direction.
Spectral graph theory studies the connectivity of a single graph. We're studying the spectral alignment between two structures:
- The dynamics (transition probabilities) — how the system actually moves
- The geometry (attribute similarity) — how the system "should" move if it were optimizing for smoothness
The Tension-Graph Laplacian is literally the product of these two structures. Its eigenvectors are the directions where dynamics and geometry are most compatible.
This is NOT standard spectral graph theory. It's closer to:
- Information geometry: the Fisher information metric measures how parameter changes affect distributions. Our Laplacian measures how state transitions respect attribute geometry.
- Optimal transport: the Wasserstein distance respects both probability and geometry. Our construction is related — we're finding the cheapest directions for transport on the attribute-weighted graph.
- Sheaf cohomology: we proved L_F = (δ⁰)†δ⁰. The conservation condition (low gradient variance) IS the cohomological condition for a global section.
Conjecture (Spectral Alignment Principle): For any Markov chain with a Lipschitz attribute on its state space, the product W = P ⊙ K (transition × kernel) yields a Laplacian whose eigenvectors decompose the system into a spectrum from "dynamics-geometry aligned" (λ → 0) to "dynamics-geometry opposed" (λ → max). Any attribute that is approximately conserved by the dynamics will concentrate its energy in the low-λ eigenvectors.
This is STRONGER than what we proved. Our theorems say "if conservation holds, it shows up in eigenvectors." The real insight is the CONVERSE: "the eigenvector spectrum TELLS YOU which attributes the dynamics conserves, and to what degree."
The Ising Hamiltonian is symmetric — it has no preferred direction. The spin-spin interaction is the same in all directions. So the "attribute" (magnetization) has no natural geometry to align with the dynamics. The Laplacian is roughly isotropic, and no eigenvector is special.
Music works because tonal harmony has EXTREMELY anisotropic structure. The circle of fifths creates a preferred direction. The PLR group creates algebraic constraints. The tension metric has massive variance across chord pairs. There's something for the Laplacian to "grab onto."
The condition for conservation detection is NOT that the system has conservation. It's that the system has ANISOTROPIC structure that the dynamics partially respects.
This predicts:
- Protein folding WILL show conservation (folded states are anisotropic — hydrophobic core vs surface)
- Language WILL show weak conservation (syntax has some anisotropy, but less than music)
- Random walks on symmetric graphs WILL NOT show conservation (no anisotropy)
- Social networks WILL show conservation in communities (homophily creates anisotropy)
The Tension-Graph Laplacian is a structural debugger. Given any system:
- Build the dynamics (transition graph from runtime behavior)
- Build the geometry (attribute similarity from domain knowledge)
- Compute L = D - W
- The eigenvectors tell you where the system is "healthy" (aligned) vs "broken" (conflicted)
This works for:
- Compilers: token transitions × character similarity → detect obfuscation
- MoE routing: expert transitions × output similarity → detect load imbalance
- Format parsers: chunk transitions × entropy similarity → detect corruption
- APIs: endpoint transitions × response similarity → detect misuse
- Neural networks: layer transitions × activation similarity → detect mode collapse
- Git history: file transitions × change similarity → detect architectural drift
1. Temporal Conservation Dynamics
We compute conservation as a single number. But conservation CHANGES over time. The time series CR(t) is itself a signal:
- Rising conservation: system settling into equilibrium
- Falling conservation: system entering transition (modulation, phase change, bug)
- Oscillating conservation: periodic structure (verse-chorus, day-night, sprint cycles)
The DERIVATIVE of conservation is more informative than conservation itself. This is our modulation detector, generalized.
2. Multi-Scale Conservation
Every system has conservation at multiple scales:
- Micro: individual transitions (local smoothness)
- Meso: phrases/sequences (medium-range structure)
- Macro: entire corpora (global structure)
We've only looked at one scale. A multi-resolution wavelet-Laplacian would reveal the scale structure of conservation. This could detect hierarchical structure automatically — without knowing the hierarchy in advance.
3. Conservation as a Training Signal
If conservation measures how well a system "makes sense," then we can TRAIN systems to maximize conservation. This is:
- A regularizer for neural networks (conserve activations along spectral directions)
- A curriculum for language models (learn low-conservation modes first)
- A breeding objective for agents (select for high-conservation behavior)
- A design principle for APIs (endpoints should conserve response structure)
This is the biggest unexplored idea: conservation as an objective function, not just a diagnostic.
-
Conservation Gradient Descent — optimize any system's parameters to maximize conservation along its Laplacian eigenvectors. Start with a simple neural network. Show it finds better representations.
-
Multi-Scale Wavelet Laplacian — build Laplacians at multiple scales, compute conservation at each. Apply to music (should recover phrase structure automatically) and code (should recover function/class boundaries).
-
Temporal Conservation Tracker (production) — not just sliding window, but proper Kalman-filter-based tracking of conservation over time. Detect phase transitions in real-time.
-
Conservation Transfer Learning — compute conservation on a well-understood system, transfer the Laplacian to a new system. If they share structure, the transferred Laplacian should still detect conservation. This is "pre-training" for structural analysis.
-
The Inverse Problem — given a conservation signal, reconstruct the transition graph. This is the tomography problem: observe the shadows (conservation), reconstruct the object (dynamics). If this works, we can infer system structure from conservation measurements alone.
The Tension-Graph Laplacian sits at the intersection of:
Spectral Graph Theory (connectivity → eigenvalues)
×
Information Geometry (distributions → metric tensor)
×
Sheaf Cohomology (local → global consistency)
×
Optimal Transport (geometry + probability → distance)
The conservation condition (low gradient variance in eigenbasis) is equivalent to:
- Spectral: energy concentration in low-frequency modes
- Information-geometric: the dynamics follow geodesics of the attribute metric
- Sheaf-theoretic: the attribute is a global section (cohomologically trivial)
- Transport-theoretic: the Wasserstein distance between consecutive states is small
These are all the SAME condition, viewed through different mathematical lenses. The Tension-Graph Laplacian is the object that unifies them.
This is why it keeps working across domains — not because conservation is universal (it's not, Ising showed us that), but because the MATHEMATICAL STRUCTURE is universal. Any system with dynamics + geometry has a Laplacian. The question is whether the dynamics respect the geometry. The Laplacian answers this question automatically.
What we have: a powerful diagnostic tool with strong empirical backing in music, promising early results in 5 other domains, solid mathematical foundations, and honest negative results.
What we don't have: a proof that this works for anything beyond music. The Neyman-Pearson control is strong but only for SYNTHETIC data. We need real out-of-sample validation.
What's most likely to be wrong: the specific number 112×. This almost certainly overestimates the effect size. On larger, more diverse corpora, I'd expect 10-30×. Still large, but not nearly as dramatic.
What's most likely to be RIGHT: the qualitative pattern. Eigenvectors of the dynamics-geometry Laplacian reveal structure that's invisible in the raw data. This is mathematically guaranteed by the Rayleigh quotient and Cheeger inequality. The question is always "how much structure," not "whether there is structure."
What would kill the program: if real-world systems (actual large-scale codebases, actual MoE models in production) show no conservation signal. This is the experiment we haven't done. Everything so far is synthetic or small-scale.
What would confirm the program: the Conservation Gradient Descent experiment. If optimizing for conservation improves neural network representations, that's not just a diagnostic — it's a new training paradigm.