{
 "artifacts": null,
 "category": "dynamics",
 "description": "Treat groups of neural-network states or experts as metastable sectors and estimate both sector imbalance and inter-sector connectivity from minibatch routing or trajectory transitions. At balanced sector usage, the effective two-sector spectral splitting becomes a direct estimate of connectivity: a large splitting indicates that the sectors are still strongly communicating, whereas a small splitting indicates genuine specialization or incipient collapse into disconnected modes.",
 "formulas_latex": [
  "$$\\mathcal{V}_{\\mathrm{eff}}=\\begin{pmatrix}a\u0026c_{12}\\\\c_{21}\u0026b\\end{pmatrix},\\qquad \\Delta=\\lambda_+-\\lambda_-=\\sqrt{(a-b)^2+4c_{12}c_{21}}.$$",
  "$$a=b\\ \\Longrightarrow\\ \\Delta=2\\sqrt{c_{12}c_{21}},\\qquad c_{ij}\\propto e^{-\\sigma L^{d-1}}.$$",
  "$$P_{ij}=\\frac{n_{ij}+\\alpha}{\\sum_{j'}(n_{ij'}+\\alpha)},\\qquad g=1-|\\lambda_2(P)|,\\qquad I=\\max_i\\left|u_i-\\frac{1}{K}\\right|.$$",
  "$$\\mathcal{L}_{\\mathrm{coex}}=\\mathcal{L}_{\\mathrm{task}}+\\beta\\,[\\max(0,g-g_{\\mathrm{high}})]^2+\\gamma\\,[\\max(0,I-I_{\\mathrm{max}})]^2.$$"
 ],
 "id": 2709,
 "implementation": "1. Integration point: apply this to a two- or multi-expert mixture-of-experts model, or to a classifier with two deliberately separated representation clusters. For each minibatch, record the sector identity of every token: the selected expert, nearest prototype, predicted class mode, or a discretized region of activation space. Maintain exponentially decayed transition counts $n_{ij}$ between the sector at the previous training step and the current sector for matched examples or adjacent hidden states. 2. Pseudocode: initialize $n_{ij}=0$; at every update, for each tracked item with old sector $i$ and new sector $j$, set $n_{ij}\\leftarrow \\rho n_{ij}+(1-\\rho)\\mathbf{1}[i\\to j]$; form the smoothed transition matrix $P_{ij}=(n_{ij}+\\alpha)/\\sum_k(n_{ik}+\\alpha)$; estimate the second eigenvalue with power iteration after projecting out the stationary vector; compute $g=1-|\\lambda_2(P)|$ and sector usage imbalance $I$; only add the connectivity penalty or change router temperature when $I\u003cI_{\\max}$ and $g\u003eg_{\\mathrm{high}}$. A practical controller increases router temperature or adds small assignment noise when balanced sectors communicate too strongly, and decreases noise when $g\u003cg_{\\mathrm{low}}$ to prevent irreversible expert fragmentation. 3. Computed versus estimated: the two-sector eigenvalue formula is the paper-derived mechanism; $P$, $g$, and $I$ are empirical estimates and require no Hessian computation. Use bootstrap windows to attach confidence intervals to $g$. 4. First cheap experiment: train a two-expert top-1 MoE MLP on MNIST or CIFAR-10 with identical experts, comparing standard load-balancing loss against the spectral controller. Sweep router temperature and noise. The falsifiable prediction is that at equal expert usage ($I\u003c0.05$), the measured two-sector gap should vary monotonically with empirical cross-expert transition probability; in a symmetric two-state chain with switching probability $q$, $g\\approx 2q$. A specialization transition should occur when $g$ falls below a reproducible threshold, and the controller should reduce high-gap balanced-but-nonspecialized training without increasing usage imbalance. Test whether the observed gap agrees with $2q$ within 20% and whether loss or accuracy changes sharply near the chosen $g_{\\mathrm{low}}$ or $g_{\\mathrm{high}}$ rather than only improving smoothly.",
 "math_summary": "The paper reduces two dominant sectors to a nonnegative effective operator $\\mathcal{V}_{\\mathrm{eff}}=\\begin{pmatrix}a\u0026c_{12}\\\\c_{21}\u0026b\\end{pmatrix}$, where $a,b$ are within-sector dominant weights and $c_{12},c_{21}\\geq 0$ are inter-sector communication amplitudes. Its eigenvalue splitting is $\\Delta=\\lambda_+-\\lambda_-=\\sqrt{(a-b)^2+4c_{12}c_{21}}$. The imbalance coordinate is $a-b$; at balance, $a=b$, the splitting reduces to $\\Delta=2\\sqrt{c_{12}c_{21}}$, so the observed gap isolates connectivity rather than sector preference. For a short-range system with interface tension $\\sigma\u003e0$ and transverse scale $L$, the mechanism predicts $c_{ij}\\propto\\exp[-\\sigma L^{d-1}]$, where $d$ is spatial dimension. In a neural implementation, let $P$ be an empirical row-stochastic transition matrix between $K$ sectors, with eigenvalues $1=\\lambda_1\\geq |\\lambda_2|\\geq\\cdots$; the effective connectivity is measured by the Markov gap $g=1-|\\lambda_2|$, while sector imbalance is $I=\\max_s|u_s-1/K|$ for usage frequencies $u_s$. Small $g$ at small $I$ is the neural analogue of balanced coexistence with weak communication.",
 "math_tags": [
  "statistical-mechanics",
  "spectral-theory",
  "linear-algebra",
  "dynamical-systems"
 ],
 "ml_areas": [
  "moe-routing",
  "training-dynamics",
  "regularization"
 ],
 "paper": {
  "arxiv_id": "2608.23757",
  "arxiv_url": "https://arxiv.org/abs/2608.23757",
  "summary_what_math_gives_to_ml": "The paper provides a spectral decomposition of finite-size phase coexistence into two independently measurable coordinates: sector imbalance determines where coexistence is balanced, while inter-sector connectivity determines whether competing sectors remain mixed or become effectively reducible. In a two-sector effective operator, the dominant-state splitting is controlled by both imbalance and off-diagonal connectivity; at balance, the splitting directly measures connectivity, which can decay exponentially with interface cost. This mechanism can transfer to mixture-of-experts routing or multimodal training by monitoring whether balanced sectors are genuinely separated or merely rapidly communicating, then using the measured spectral gap to trigger anti-collapse or specialization schedules.",
  "title": "Weak irreducibility as a spectral criterion for phase coexistence",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 7,
  "usefulness": 6
 },
 "solves": [
  "stability",
  "generalization",
  "accuracy"
 ],
 "title": "Spectral Coexistence Monitor for Expert Collapse",
 "url": "https://synthcore.org/idea/2709/spectral-coexistence-monitor-for-expert-collapse",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}