{
 "artifacts": null,
 "category": "dynamics",
 "description": "Treat the number K of minibatches between expensive control updates as a review period: the controlled neural dynamics use parameters or decisions computed at time nK and hold them fixed until (n+1)K. Scan K, estimate first and second finite differences of validation loss or episodic return, and use the resulting nonmonotone-to-convex or concave phase diagram to select an update frequency rather than assuming that more frequent updates are always better.",
 "formulas_latex": [
  "$$V(h)=V(0)+hV'(0)+\\frac{h^{2}}{2}V''(0)+o(h^{2})$$",
  "$$\\widehat{J}'(K)=\\frac{J(K+\\Delta)-J(K-\\Delta)}{2\\Delta},\\qquad \\widehat{J}''(K)=\\frac{J(K+\\Delta)-2J(K)+J(K-\\Delta)}{\\Delta^{2}}$$",
  "$$K_{0}:=\\min\\left\\{K:\\widehat{J}'(k)\\geq 0\\ \\text{for all }k\\in\\{K,K+\\Delta,K+2\\Delta\\}\\right\\}$$",
  "$$u_t=\\pi\\left(s_{\\lfloor t/K\\rfloor K}\\right),\\qquad s_{t+1}=F(s_t,u_t,\\xi_t)$$"
 ],
 "id": 2828,
 "implementation": "Integrate this at a neural component whose decision is expensive or noisy: an Adam learning-rate controller, MoE routing logits, RL exploration temperature, target-network parameters, or a learned augmentation policy. Choose an integer review period K. At step t, recompute the control u only when t modulo K equals zero, then reuse u for the next K minibatches. For each candidate K, initialize identical weights and random seeds. At every step, if t modulo K equals zero, compute u=pi(s_t); then apply the update s_(t+1)=F(s_t,u,xi_t). Record a smoothed validation cost J(K) at a fixed optimizer-step budget and at a fixed wall-clock budget. Test K in {1,2,4,8,16,32,64} with at least five seeds. The paper supplies the review-period sensitivity mechanism and derivative or curvature interpretation; J(K), confidence intervals, and the transition location must be estimated empirically. First experiment: train a small Transformer on WikiText-2, comparing per-step against every-K-step recomputation of learning-rate and gradient-clipping controls, with ordinary Adam as baseline. The quantitative prediction is a small-K nonmonotonicity or curvature change: the estimated J'(K) should change sign or estimated J''(K) should change sign across the scan. This transition should persist across seeds and become sharper at larger batch sizes, where stochastic smoothing is weaker. Select K at the first stable curvature regime and test it on a second model.",
 "math_summary": "Let h\u003e0 be the review period, V(h) the optimal fluid-control value, and V'(h), V''(h) its first and second sensitivities. The transferable mechanism is that V'(h) can have either sign for small h, while beyond a problem-dependent threshold h0, V'(h)\u003e=0; V''(h) may then change sign across convex, linear, and concave regions. For neural training, let J(K) be a held-out cost after a fixed compute budget when a control vector u is recomputed every K minibatches and held constant between reviews. Estimate J'(K) and J''(K) using symmetric finite differences. Define K0 as the smallest K for which the derivative sign is stable over three neighboring values. The mechanism predicts that J(K) need not be monotone near K=1 and that curvature changes can identify a robust operating region.",
 "math_tags": [
  "dynamical-systems",
  "control-theory",
  "optimization",
  "numerical-analysis"
 ],
 "ml_areas": [
  "optimizer",
  "scheduler",
  "training-dynamics",
  "moe-routing",
  "rl"
 ],
 "paper": {
  "arxiv_id": "2608.29398",
  "arxiv_url": "https://arxiv.org/abs/2608.29398",
  "summary_what_math_gives_to_ml": "The paper provides a constructive sensitivity mechanism for controls that are frozen between discrete review epochs: the fluid value can be nonmonotone for short review periods and, after a threshold, develop monotone nondecreasing, convex or linear, and ultimately concave regimes. Its transferable asset is treating update frequency as a dynamical parameter and detecting qualitative transitions through first- and second-order value sensitivities. A neural analogue is to hold optimizer, routing, normalization, or target-network decisions fixed for K minibatches and measure training value as a function of the review period. The strongest test is whether loss or reward curves exhibit predicted derivative-sign and curvature transitions rather than merely a benchmark improvement.",
  "title": "Review-Period Sensitivity in Multiclass Queue Scheduling",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 4,
  "novelty": 6,
  "usefulness": 5
 },
 "solves": [
  "stability",
  "speedup",
  "accuracy"
 ],
 "title": "Review-Period Phase Diagram for Frozen Updates",
 "url": "https://synthcore.org/idea/2828/review-period-phase-diagram-for-frozen-updates",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}