Let \( A \) be an invertible normal hyperbolic operator on a complex Hilbert space \( H \). We obtain an exact spectral formula for the optimal uniform shadowability constant \( Shad(A) \), sharpening the general additive Green-operator estimates to a Euclidean combination of the stable and unstable spectral contributions. If both spectral components are nonempty, set \( \alpha(A)=1-\max\{|\lambda|:\lambda\in\sigma(A),\ |\lambda|1\}-1. \) We prove the exact formula \( Shad(A)=\left(\alpha(A)^{-2}+\beta(A)^{-2}\right)^{1/2}, \) with the corresponding one-sided formulas when only one spectral component is present. We further relate this exact constant to the robustness of hyperbolicity. Let \( r_{\mathrm{hyp}}(A) \) denote the operator-norm distance from \( A \) to the loss of hyperbolicity. We obtain the sharp universal ratio \( 1\le r_{\mathrm{hyp}}(A)Shad(A)\le\sqrt{2}. \) Both constants are optimal. In the mixed case, \( r_{\mathrm{hyp}}(A)Shad(A)=\left[1+\left(\frac{\min\{\alpha(A),\beta(A)\}}{\max\{\alpha(A),\beta(A)\}}\right)^2 \right]^{1/2}. \) Thus the normalized shadowability constant is determined exactly by the relative balance of the stable and unstable spectral gaps.
Let \(A\) be a bounded invertible hyperbolic operator on a complex Hilbert space. We compare its optimal two-sided shadowing constant \(S(A)\) with the hyperbolicity robustness radius \(r(A)\), the operator-norm distance to a perturbation whose spectrum meets the unit circle. We determine the optimal dimension-independent asymptotic coefficient: \[ \lim_{\varepsilon\downarrow0} \sup_{\substack{A\\0<r(A)\le\varepsilon}} S(A)r(A)^{3/2}=\frac{3^{3/4}}{\sqrt2}. \] The supremum ranges over bounded invertible hyperbolic operators on arbitrary nonzero complex Hilbert spaces. In the stable class, the coefficient is \(3^{3/4}/2\). Explicit finite-dimensional invertible matrix sequences attain both coefficients as limits. The universal upper estimate combines the exact dual norm of the Green operator with a weighted variational inequality and a Riccati equation. A complementary Fourier argument gives an explicit bound valid at every positive radius. The exponent \(\frac32\) remains optimal under fixed norm restrictions: it is attained in order by uniformly bounded matrices and by bilateral weighted shifts on \(\ell^2(Z)\) with fixed spectrum \(\{z:|z|=\frac12\}\), uniformly bounded norms and inverse norms, and an explicitly computable shadowing constant.
We study how radial gradient clipping affects the period-two orbit created by a supercritical flip bifurcation near a nondegenerate minimum with a simple maximal Hessian eigenvalue. Clipping leaves the local flip threshold unchanged. For a small clipping threshold \(c\), the stable smooth two-cycle first reaches the clipping boundary at a parameter distance \(Kc^2 + O(c^4)\) from the flip, with an explicit coefficient determined by the potential and the clipping norm. The two cycle points have opposite gradients and therefore reach the boundary simultaneously. A clipped two-cycle cannot have one point strictly inside the smooth region and the other strictly inside the clipped region. Every two-cycle with both points strictly clipped has an exact unit Floquet multiplier and is nonhyperbolic. A solvable planar model illustrates the transition from a stable smooth two-cycle to a neutral family of clipped two-cycles.
We construct two families of Anosov automorphisms on a fixed flat two-dimensional torus to compare topological entropy with the optimal local Lipschitz shadowing constant. In the first family, entropy tends to infinity while the shadowing constant tends to one. In the second, the maps are smoothly conjugate and have identical entropy, periodic-point counts, and eigenvalues, but their shadowing constants diverge. Thus neither entropy nor the full periodic-point counting sequence gives a finite upper bound for the shadowing constant. The constructions distinguish the effects of expansion strength and invariant-direction geometry.
Adaptive gradient clipping can improve training without establishing that its feedback signals are useful. We examine Dynamics-Aware Gradient Clipping (DAGC), a bounded threshold controller driven by clipping exposure and consecutive-gradient alignment. Its exposure term tightens clipping when clipping is already frequent. We show that an interior target-exposure equilibrium is unstable in a stationary mean-field model, even with exponential smoothing. We then compare the frozen controller with tuned clipping rules, an alignment ablation, and a predetermined threshold schedule on FashionMNIST and CIFAR-10, using validation-only calibration and twelve paired confirmation seeds in each of three conditions.These controls separate threshold scheduling from evidence for an online feedback mechanism; the analysis does not claim convergence of the coupled training dynamics.