{
 "artifacts": null,
 "category": "dynamics",
 "description": "Add a deliberately nonconservative, antisymmetric parameter-space force to ordinary gradient descent, with its amplitude controlled by an empirically estimated stability margin. The force should move parameters around elongated loss valleys instead of repeatedly descending and stopping along the same local gradient direction, while damping preserves convergence. The method directly tests whether nonzero circulation can improve traversal of flat or ill-conditioned regions without destabilizing training.",
 "formulas_latex": [
  "$$\\mathbf F=-\\nabla V_{\\rm P}+\\mathbf F_{\\rm NCSF},\\qquad \\oint_C\\mathbf F_{\\rm NCSF}\\cdot d\\boldsymbol{\\phi}\\ne0.$$",
  "$$\\theta_{k+1}=\\theta_k-\\eta\\left[I-\\gamma\\alpha J\\right]\\nabla L(\\theta_k),\\qquad J^T=-J.$$",
  "$$L(\\theta)=\\frac12\\theta^TH\\theta\\ \\Longrightarrow\\ A=I-\\eta\\left[I-\\gamma\\alpha J\\right]H,\\qquad \\text{stable iff }\\rho(A)\u003c1.$$",
  "$$W_C=\\oint_C\\gamma\\alpha J\\nabla L(\\theta)\\cdot d\\theta.$$"
 ],
 "id": 2806,
 "implementation": "1. Integration point: implement this as an optimizer wrapper around any PyTorch model. Partition the flattened parameter vector into blocks of size 2 or 4, and define a block-diagonal skew matrix \\(J\\); for every 2D block use \\(J_b=\\begin{pmatrix}0\u0026-1\\\\1\u00260\\end{pmatrix}\\). Do not materialize a full matrix: rotate each parameter-gradient pair in place. 2. Pseudocode: compute the minibatch gradient \\(g\\); for each pair \\((g_i,g_{i+1})\\), compute \\(r_i=(-g_{i+1},g_i)\\); update \\(\\theta\\leftarrow\\theta-\\eta(g-\\gamma\\alpha r)\\). Every \\(K\\) steps estimate the local Jacobian/Hessian action with two Hessian-vector products or finite differences, form a small block approximation to \\(A=I-\\eta(I-\\gamma\\alpha J)H\\), and reduce \\(\\alpha\\) or \\(\\eta\\) if its estimated spectral radius exceeds 0.98. Use gradient clipping only as a fallback, not as the primary stability mechanism. 3. Computed from the paper: the force decomposition and the required nonzero circulation criterion. Estimated empirically: Hessian-vector products, the spectral radius, and loop work. Record checkpoints along a small rectangular loop in two selected parameter directions and numerically integrate \\(F\\cdot d\\theta\\); use the update direction as the force. 4. First cheap experiment: train a 2-layer MLP on MNIST or CIFAR-10 with SGD, Adam, and this optimizer at matched learning-rate budgets. Use a two-dimensional quadratic benchmark with known \\(H=\\operatorname{diag}(1,\\kappa)\\) first, sweeping \\(\\eta\\) and \\(\\alpha\\). The quantitative prediction is that divergence begins at the measured boundary \\(\\rho(I-\\eta(I-\\gamma\\alpha J)H)=1\\), within 20% of the observed boundary. On neural loss surfaces, nonzero loop work should scale approximately linearly with \\(\\alpha\\) for small \\(\\alpha\\), while setting \\(\\alpha=0\\) must recover the baseline exactly. Compare time to reach a fixed loss and verify that any speedup disappears when loop work is statistically indistinguishable from zero.",
 "math_summary": "The paper decomposes a force on collective coordinates \\(\\boldsymbol{\\phi}\\) as \\(\\mathbf F=-\\nabla V_{\\rm P}+\\mathbf F_{\\rm NCSF}\\), where \\(V_{\\rm P}\\) is a scalar ponderomotive potential and \\(\\mathbf F_{\\rm NCSF}\\) is the nonconservative component, satisfying \\(\\oint_C\\mathbf F_{\\rm NCSF}\\cdot d\\boldsymbol{\\phi}\\ne0\\) for some closed loop \\(C\\). Transfer this to parameters \\(\\theta\\) by using \\(\\dot\\theta=-\\nabla L(\\theta)+\\alpha J\\nabla L(\\theta)\\), where \\(L\\) is the training loss, \\(\\alpha\\ge0\\) is the circulation strength, and \\(J^T=-J\\) is a fixed or blockwise-learned skew-symmetric matrix. For a quadratic local model \\(L(\\theta)=\\frac12\\theta^TH\\theta\\), with symmetric positive-semidefinite Hessian \\(H\\), the discrete update is \\(\\theta_{k+1}=A\\theta_k\\), \\(A=I-\\eta(I-\\alpha J)H\\), where \\(\\eta\\) is the learning rate. Linear stability requires \\(\\rho(A)\u003c1\\), with \\(\\rho\\) the spectral radius. The nonconservative character is tested by loop work \\(W_C=\\oint_C\\alpha J\\nabla L(\\theta)\\cdot d\\theta\\), which should be nonzero when \\(J\\) varies across blocks or when the loss Hessian and \\(J\\) do not commute. A damping factor \\(\\gamma\\in(0,1]\\) can replace the update by \\(\\theta_{k+1}=\\theta_k+\\eta[-\\nabla L(\\theta_k)+\\gamma\\alpha J\\nabla L(\\theta_k)]\\), limiting rotational energy injection.",
 "math_tags": [
  "dynamical-systems",
  "control-theory",
  "linear-algebra",
  "differential-geometry"
 ],
 "ml_areas": [
  "optimizer",
  "training-dynamics",
  "scheduler"
 ],
 "paper": {
  "arxiv_id": "2608.29122",
  "arxiv_url": "https://arxiv.org/abs/2608.29122",
  "summary_what_math_gives_to_ml": "The paper identifies a transferable mechanism: periodic driving plus dissipation produces a static force with a genuinely nonconservative component, characterized by nonzero circulation around closed loops in collective-coordinate space. In neural-network training, this suggests augmenting gradient descent with a controlled rotational force in parameter space, rather than assuming every useful update is the gradient of a scalar objective. The rotational component can transport parameters along loss valleys and between basins while a symmetric dissipative component controls stability. The key falsifiable signature is nonzero measured loop work together with a sharp stability boundary determined by the spectral radius of the resulting update map.",
  "title": "Light-induced nonconservative static forces in many-body systems",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 8,
  "usefulness": 6
 },
 "solves": [
  "stability",
  "speedup"
 ],
 "title": "Dissipative Circulation Optimizer",
 "url": "https://synthcore.org/idea/2806/dissipative-circulation-optimizer",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}