Skip to content
Shadow Theory

Global attraction and nonlinear gain

A sharp scalar curvature estimate yields a strict sufficient all-input stability regime and a controlled contrast gain.

4 Sharp curvature and global attraction

Lemma 4.1 (Reduced edge curvature). For d∈C,r=∣d∣,s=W0(r2/(2κ)),w=e−sd\in\mathbb C,r=|d|,s=W_0(r^2/(2\kappa)),w=e^{-s}, the two real Hessian eigenvalues of κ(s+s2/2)\kappa(s+s^2/2) are

λtan=w,λrad=w1−s1+s.(18)\lambda_{\rm tan}=w,\qquad\lambda_{\rm rad}=w\frac{1-s}{1+s}.\tag{18}

Both are one at the origin by continuity. The sharp scalar lower bound is −χ-\chi, where

χ=(2−3)e−3=0.0474058943568….(19)\chi=(2-\sqrt3)e^{-\sqrt3}=0.0474058943568\ldots.\tag{19}

Proof. From ses=r2/(2κ)se^s=r^2/(2\kappa), ds/dr=2s/[r(1+s)]ds/dr=2s/[r(1+s)] for r>0r>0. Differentiate the edge gradient wdwd: its tangential eigenvalue is ww and its radial eigenvalue w+r dw/drw+r\,dw/dr is (18). For s≤1s\le1 both are nonnegative; for s>1s>1,

dds[e−ss−1s+1]=e−s3−s2(s+1)2.(20)\frac d{ds}\left[e^{-s}\frac{s-1}{s+1}\right]=e^{-s}\frac{3-s^2}{(s+1)^2}.\tag{20}

The magnitude peaks at s=3s=\sqrt3 with value χ\chi. Every s≥0s\ge0 is attained for some rr, proving scalar sharpness without assuming simultaneous worst directions on all graph edges.

Pull back the edge Hessians through z↦be⊤zz\mapsto b_e^\top z:

∇2Veff≥μI−χLR≥δI,LR=diag⁡(L,L),δ=μ−χ∥L∥.(21)\nabla^2 V_{\rm eff}\ge\mu I-\chi L_R\ge\delta I,\qquad L_R=\operatorname{diag}(L,L),\quad\delta=\mu-\chi\|L\|.\tag{21}

The spectral norm includes all multiplicities and collective modes.

Theorem 4.2 (Unique sustained state and global attraction). Under the positivity hypotheses and δ>0\delta>0, every constant complex ff has one globally asymptotically stable equilibrium, with ∥z∗∥≤∥ℓ∥/μ\|z_*\|\le\|\ell\|/\mu, m∗≤0m_*\le0.

Proof. The coercive VeffV_{\rm eff} is strongly convex by (21), so its unique critical point is the unique minimizer z∗z_*. The strictly convex memory slices restore m∗m_*. At equilibrium K(m∗)z∗=ℓK(m_*)z_*=\ell, giving the norm bound, and (15) gives the memory sign.

Along any bounded trajectory VV converges to a finite limit. Write g=γ∥Kz−ℓ∥2+η∥Fm∥2g=\gamma\|Kz-\ell\|^2+\eta\|F_m\|^2. An omega-limit point with g>0g>0 would have a ball where g≥ϵ>0g\ge\epsilon>0. Bounded speed on a compact neighborhood forces a positive residence time in that ball after each visit to a smaller concentric ball, unless the trajectory never leaves, in which case V→−∞V\to-\infty. Infinitely many disjoint residence intervals likewise force unbounded loss. Therefore all omega-limit points have g=0g=0; compactness and the unique equilibrium then imply convergence of the entire trajectory.

Moreover V(z,m)≥Veff(z)≥Veff(z∗)=V(z∗,m∗)V(z,m)\ge V_{\rm eff}(z)\ge V_{\rm eff}(z_*)=V(z_*,m_*), with equality only there. On a small sphere around the equilibrium the excess has a positive minimum. Initial states in a smaller sublevel cannot cross it by (10), proving Lyapunov stability. This direct finite-dimensional invariance argument [LaSalle 1976] gives global convergence, without a uniform global exponential or intertrajectory contraction estimate. Failure of the strict sufficient condition alone establishes neither instability nor multistability.

5 Complete local stability and a parameter bound

At equilibrium q∗=(ℜz∗,ℑz∗)⊤q_*=(\Re z_*,\Im z_*)^\top, set

H=(KRBB⊤Hm),B⋅e=em∗,ediag⁡(Be,Be)q∗,Hm=diag⁡(κ+12em∗,e∣be⊤z∗∣2).(22)H=\begin{pmatrix}K_R&B\\B^\top&H_m\end{pmatrix},\quad B_{\cdot e}=e^{m_{*,e}}\operatorname{diag}(B_e,B_e)q_*,\quad H_m=\operatorname{diag}\big(\kappa+\tfrac12 e^{m_{*,e}}|b_e^\top z_*|^2\big).\tag{22}

Differentiating stationary memory gives dm/dq=−Hm−1B⊤dm/dq=-H_m^{-1}B^\top, hence

S=KR−BHm−1B⊤=∇2Veff(z∗)≥δI.(23)S=K_R-BH_m^{-1}B^\top=\nabla^2 V_{\rm eff}(z_*)\ge\delta I.\tag{23}

Consequently H>0H>0. The constant-mobility Jacobian is

Afull=−DH,D=diag⁡(G,ηI),12(D+D⊤)=diag⁡(γI,ηI).(24)A_{\rm full}=-DH,\qquad D=\operatorname{diag}(G,\eta I),\qquad\tfrac12(D+D^\top)=\operatorname{diag}(\gamma I,\eta I).\tag{24}

Partial amplitude curvature alone cannot establish this joint-state result.

Proposition 5.1 (Linear rate and nonlinear local stability). With σ=min⁡{γ,η}\sigma=\min\{\gamma,\eta\}, define

θ0=∥ℓ∥∑e∥be∥4κμ,h0=min⁡{δ,κ}(1+θ0)2>0.(25)\theta_0=\frac{\|\ell\|\sqrt{\sum_e\|b_e\|^4}}{\kappa\mu},\qquad h_0=\frac{\min\{\delta,\kappa\}}{(1+\theta_0)^2}>0.\tag{25}

Then H≥h0IH\ge h_0 I, and for ∥v∥H=(v⊤Hv)1/2\|v\|_H=(v^\top Hv)^{1/2},

∥v(t)∥H≤e−σh0t∥v(0)∥H,max⁡λ∈spec⁡(Afull)ℜλ≤−σh0.(26)\|v(t)\|_H\le e^{-\sigma h_0 t}\|v(0)\|_H,\qquad\max_{\lambda\in\operatorname{spec}(A_{\rm full})}\Re\lambda\le-\sigma h_0.\tag{26}

Every positive rate strictly below σh0\sigma h_0 holds for nonlinear evolution in a sufficiently small neighborhood in the equilibrium Hessian norm.

Proof. Since m∗≤0m_*\le0, ∥B⋅e∥≤∥be∥2∥z∗∥\|B_{\cdot e}\|\le\|b_e\|^2\|z_*\|, hence ∥Hm−1B⊤∥≤κ−1∥B∥≤θ0\|H_m^{-1}B^\top\|\le\kappa^{-1}\|B\|\le\theta_0. Completing the square gives

(ξ,ν)⊤H(ξ,ν)=ξ⊤Sξ+(ν+Hm−1B⊤ξ)⊤Hm(ν+Hm−1B⊤ξ).(27)(\xi,\nu)^\top H(\xi,\nu)=\xi^\top S\xi+(\nu+H_m^{-1}B^\top\xi)^\top H_m(\nu+H_m^{-1}B^\top\xi).\tag{27}

The inverse shear has norm at most 1+θ01+\theta_0. Combine (23) and Hm≥κIH_m\ge\kappa I to get H≥h0IH\ge h_0I. For Ev=v⊤Hv/2E_v=v^\top Hv/2, the linearized system gives

E˙v=−(Hv)⊤diag⁡(γI,ηI)(Hv)≤−σh0v⊤Hv=−2σh0Ev.(28)\dot E_v=-(Hv)^\top\operatorname{diag}(\gamma I,\eta I)(Hv)\le-\sigma h_0 v^\top Hv=-2\sigma h_0E_v.\tag{28}

Integrate to obtain the norm and spectral bounds. The nonlinear vector field differs by O(∥v∥2)O(\|v\|^2), adding O(∥v∥3)O(\|v\|^3) to E˙v\dot E_v. Bound it by 2ϵEv2\epsilon E_v in a small invariant sublevel for any 0<ϵ<σh00<\epsilon<\sigma h_0.

The exact linear estimate, local nonlinear rate and numerical eigenvalues remain distinct. A fixed nonzero input fixes phase; at zero input positive μ,κ\mu,\kappa also remove zero decay rates despite the graph collective mode. For the path, ∥L∥=3,μ=2/5\|L\|=3,\mu=2/5. Since 3>5/3, 2−3<1/3\sqrt3>5/3,\ 2-\sqrt3<1/3 and the cubic Taylor polynomial for the exponential at 5/35/3 exceeds four, χ<1/12,δ>3/20\chi<1/12,\delta>3/20. With unit input ∣c∣=9/5>1|c|=\sqrt{9/5}>1, and ∑e∥be∥4=8<3\sqrt{\sum_e\|b_e\|^4}=\sqrt8<3, giving θ0<15/2, h0>3/1445\theta_0<15/2,\ h_0>3/1445, and

σh0>9/7225.(29)\sigma h_0>9/7225.\tag{29}

This is conservative: (25) evaluated numerically gives rate σh0≈0.00393374\sigma h_0\approx0.00393374, while the complete numerical spectral abscissa is about −0.24-0.24.

6 What the drive imposes and feedback changes

6.1 Collective control and symmetric contrast gain

With be⊤1=0b_e^\top\mathbf1=0, zˉ=n−11⊤z, fˉ=n−11⊤f\bar z=n^{-1}\mathbf1^\top z,\ \bar f=n^{-1}\mathbf1^\top f satisfy

zˉ(t)=fˉcμ+(zˉ(0)−fˉcμ)e−cμt.(30)\bar z(t)=\frac{\bar f}{c\mu}+\left(\bar z(0)-\frac{\bar f}{c\mu}\right)e^{-c\mu t}.\tag{30}

A matched fixed carrier has the same mean. Uniform input and initial amplitude with zero memory keep all contrasts and memories zero. Thus sustained collective amplitude controls for excitation without adaptive edge organization.

For the path take

v=(1,0,−1)⊤/2,f=av,a>0.(31)v=(1,0,-1)^\top/\sqrt2,\qquad f=av,\quad a>0.\tag{31}

Signed reflection (swap endpoints and reverse sign of amplitude) preserves this preparation. Starting from zero gives the invariant sector z=uvz=uv and equal memories. At equilibrium m1=m2=−sm_1=m_2=-s; Lv=vLv=v and ∣be⊤z∣2=∣u∣2/2|b_e^\top z|^2=|u|^2/2 imply

a2∣c∣2=4κses(μ+e−s)2,u=a/cμ+e−s.(32)\frac{a^2}{|c|^2}=4\kappa s e^s(\mu+e^{-s})^2,\qquad u=\frac{a/c}{\mu+e^{-s}}.\tag{32}

For a>0a>0, s>0s>0. The right side starts at zero and diverges; its derivative has the sign of μ+e−s(1−s)/(1+s)\mu+e^{-s}(1-s)/(1+s), positive for μ>χ\mu>\chi by Lemma 4.1. Thus it specifies exactly one ss there. The fixed carrier gives u0=(a/c)/(μ+1)u_0=(a/c)/(\mu+1), and

G=∣u∣∣u0∣=μ+1μ+e−s>1.(33)G=\frac{|u|}{|u_0|}=\frac{\mu+1}{\mu+e^{-s}}>1.\tag{33}

The drive specifies the spatial mode and relative phase. Feedback produces nonzero equal memories, persistent softening and excess amplitude. For the worked parameters s=0.07303438894…, G=1.0529729143…s=0.07303438894\ldots,\ G=1.0529729143\ldots, about 5.30% gain. This is a fixed-preparation prediction. A one-port input retains mean and contrast, and can sustain unequal memories as shown below, without implying every asymmetric input does.

6.2 Equal stationary power and continuing work

At equilibrium Kz∗=ℓ, Fm=0Kz_*=\ell,\ F_m=0, so

P∗=ℜ(ℓ∗cℓ)=γ∥ℓ∥2=γ∥f∥2γ2+α2.(34)P_*=\Re(\ell^*c\ell)=\gamma\|\ell\|^2=\frac{\gamma\|f\|^2}{\gamma^2+\alpha^2}.\tag{34}

The same identity holds for the fixed carrier. Equal-norm inputs have equal stationary power in this contract while differing in amplitude, memory, energy and response.

Corollary 6.1 (Work required for sustained organization). Under Theorem 4.2 with f≠0f\ne0,

∫0TPin(t) dt=P∗T+o(T),P∗>0.(35)\int_0^T P_{\rm in}(t)\,dt=P_*T+o(T),\qquad P_*>0.\tag{35}

Dissipation has the same leading growth, while stored energy converges.

Proof. Full convergence gives Pin→P∗,F→F∗P_{\rm in}\to P_*,F\to F_*. Split the integral at a time after which the difference in power is below any ϵ>0\epsilon>0. The early finite integral divided by TT vanishes; the remaining average error is at most ϵ\epsilon. Integrate (8) and use bounded energy change to transfer the leading term to dissipation.

An unreplenished finite work supply cannot maintain the nonzero state forever. Constant forcing represents powered dissipation rather than autonomous oscillation sustained indefinitely by initial energy.

Published source: Sustained Adaptive Spectral States: Global Attraction, Nonlinear Gain, and Linked Coherent–Dissipative Response. Read the original PDF alongside the web treatment.