Detecting reorganization onset via an operator commutator: Kuramoto, Floquet, and discrete scale invariance Jill F. Rankin Independent Researcher 3426 Shady Valley Drive Austin, Texas 78748 June 11, 2026 Abstract In many coupled dynamical systems, reorganization begins well before the dominant order parameter signals it. We introduce a two-dimensional operator-based diagnostic (χ,η) built from a participation operator P and a rigidity operator M: χ tracks the effective dimension of P weighted by M, and η tracks the normalized Frobenius commutator ∥[P,M ]∥ F . The construction admits a variational characterization with Gibbs-like stationary states and explicit bounds. Across four network topologies and six system sizes, η peaked before K c in 125 of 127 valid realizations (73/75 in the topology-comparison study; 52/52 in the finite-size scaling study), with the two exceptions occurring at N = 12 where finite-size fluctuations are largest. In direct comparison on the same simulations, the η-peak occurs 0.31 coupling units earlier than the pairwise transfer entropy peak and with 5× lower seed-to-seed variance, robust across estimator hyperparameters. On a periodically kicked transverse-field Ising chain (N = 4), the (χ,η) trajectory lies in the (χ < 1,η > 0) quadrant for all 80 sampled drive strengths, separating the driven regime from the Kuramoto selection–relaxation quadrant. On a discrete- scale-invariant model spectrum (N = 36; five input ratios λ ∈ {1.15, 1.25, 1.40, 1.60, 1.85}), the diagnostic recovers λ with mean absolute error 0.31% and worst-case error 0.41%. We interpret η as detecting operator-level alignment between participation and rigidity; on the systems examined, this alignment precedes both the order-parameter signal and the pairwise information-transfer peak. 1 Introduction The order parameter signaling a collective transition typically appears only after substantial internal reorganization has already occurred. In synchronizing systems, individual oscillators begin to align well before the global coherence becomes detectable in the standard Kuramoto order parameter r [1, 2]. In equilibrium systems approaching a phase transition, configurations fluctuate cooperatively while the magnetization or density order remains undisturbed [3]. In systems exhibiting discrete scale invariance, log-periodic oscillations in observables reflect a recursive reorganization of the underlying spectrum [4]. Detecting reorganization before it manifests in the conventional order parameter is both operationally important — for forecasting tipping points in ecological and climate systems, and for active control of engineered oscillator networks — and methodologically distinctive: precursor diagnostics must respond to structural changes that the order parameter, by construction, has not yet registered. 1

Several lineages of precursor diagnostics have developed. The classical critical-slowing-down in- dicators — increasing variance, rising lag-1 autocorrelation, and prolonged recovery time after perturbation — exploit the divergence of relaxation timescales near bifurcation points [3, 5, 6]. Information-theoretic precursors identify shifts in the predictive or synergistic structure of multi- variate time series: synergy from partial information decomposition peaks in the disordered phase before symmetry-breaking transitions [7], and pairwise transfer entropy peaks near the synchroniza- tion threshold in Kuramoto networks and decreases on both sides [8]. Operator-spectral methods, including Koopman-operator generalizations of stochastic resilience [9] recast precursor detection as an eigenvalue computation on an infinite-dimensional functional space. Across these lineages, the precursor signal is typically a scalar quantity, and the construction is specific to the model class on which it is defined: a synergy indicator on Ising spins does not naturally extend to a Floquet-driven Hamiltonian; a Koopman estimator built for population dynamics does not naturally extend to a scale-invariant electronic spectrum. We introduce a precursor diagnostic that is two-dimensional rather than scalar, and that is de- fined by the same operator construction across systems with otherwise unrelated phenomenology. The construction rests on a pair of Hermitian, positive-semidefinite operators: a participation op- erator P encoding which degrees of freedom are dynamically active in the collective state, and a rigidity operator M encoding the structural cost — graph Laplacian, static Hamiltonian, or band-structure operator — that organizes the participating modes. From this pair we derive two diagnostics: χ, the ratio of effective dimensions D eff (P 1/2 MP 1/2 ) at two control-parameter values, tracking selection and dimensional redistribution; and η, the normalized Frobenius commutator ∥[P,M ]∥ F /(∥P∥ F ∥M∥ F ). Both χ and η are dimensionless; η satisfies 0 ≤ η ≤ √ 2 by the Frobe- nius norm bound on commutators, with η = 0 when P and M commute (share an eigenbasis) and maximal when they are maximally misaligned. The construction admits a free-energy-like varia- tional characterization whose stationary states are Gibbs-like in M, and four bounding properties (Section 2) establish that the (χ,η) pair lives on a well-defined diagnostic plane. Three empirical anchors validate the construction across qualitatively distinct dynamical settings. (i) In the Kuramoto model, the η-peak precedes the logistic synchronization threshold K c in 125 of 127 valid realizations (73/75 in the topology-comparison study across four network topologies at N = 12; 52/52 in the finite-size scaling study across six system sizes, N = 12 to 384). The precursor gap remains positive across all sizes and does not decay with N, holding at⟨K c −K η ⟩ = 0.61± 0.05 across the large-N regime (N ≥ 96) where K c estimation is unambiguous. A direct head-to-head comparison on the same simulations against pairwise transfer entropy shows that η peaks 0.31 in coupling units earlier than TE and with approximately five-times lower seed-to-seed variance, robust across TE estimator hyperparameters. (ii) The construction extends without modification to periodically driven (Floquet) systems, where the (χ,η) plane distinguishes selection-relaxation from sustained-coherence regimes through the behavior of η under continued driving. (iii) In model spectra with engineered discrete scale invariance (E n = E 0 λ n ), the operator diagnostic recovers the input log-periodic ratio λ to within 0.3% across λ ∈ [1.15, 1.85] via collapse of η(logμ) under rescaling. We interpret η as detecting operator-level alignment between participation and rigidity. On the systems examined, this alignment precedes both the order-parameter signal and the peak in pairwise transfer entropy, and exhibits substantially lower seed-to-seed variance than the latter. The remainder of the paper is organized as follows. Section 2 defines the operators, diagnostics, and four mathematical properties. Section 3 presents the Kuramoto results: ensemble statistics across topologies and sizes, the slow-K-ramp temporal precursor experiment, and the head-to- 2

head comparison against transfer entropy with robustness checks. Section 4 presents the Floquet anchor. Section 5 presents the discrete-scale-invariance anchor. Section 6 discusses limitations, positions the construction against adjacent operator-theoretic lineages (Mori–Zwanzig projection- operator formalism, generalized inverse participation ratios, Laplacian-eigenvector synchronization diagnostics, and Frobenius commutator measures of quantum asymmetry), and outlines directions for application to physical systems spanning many orders of magnitude in characteristic frequency. 2 Methods 2.1 Operators and diagnostics Let H be a finite-dimensional Hilbert space with dimH = N. The framework is defined by a pair of operators on H. Participation operator. P is Hermitian, positive semidefinite, and normalized to TrP = 1. It plays the role of a density matrix for the collective state: its eigenvalue distribution encodes which degrees of freedom are dynamically active. Rigidity operator. M is Hermitian and fixed throughout any control-parameter sweep (the fixed-M convention; see below). It encodes structural cost — the energy or coupling weight each configuration would incur if active. The construction admits M ⪰ 0 as an additional hypothesis; when M ⪰ 0, D eff has a strict participation-ratio interpretation (Proposition 1). For indefinite Hermitian M — which arises in the Floquet and DSI anchors — D eff remains well-defined but should be read as a generalized signed-spectral ratio rather than a mode count. The Kuramoto anchor (§3) uses M = L, the graph Laplacian, which is PSD. The Floquet (§4) and DSI (§5) anchors use indefinite Hermitian M; in those settings η remains a normalized commutator bounded by √ 2, the Gibbs variational structure of §2 is preserved (e −M/T is PSD for any Hermitian M by spectral functional calculus), and D eff remains a well-defined real number. In every application below we use the fixed-M convention: M is held constant across the control- parameter sweep, and only P evolves. (The alternative — letting M co-vary with the state — forces η ≡ 0 by construction and is therefore uninformative; see §B.3.) From the pair (P,M ) we construct two diagnostics. The effective dimension is D eff (P,M ) = [Tr(MP )] 2 Tr[(MP ) 2 ] .(1) The trace identities Tr(MP ) = Tr(A) and Tr[(MP ) 2 ] = Tr(A 2 ) follow from the cyclic property of the trace and hold whether or not M is positive semidefinite. The identity Tr[(MP ) 2 ] = Tr(A 2 ) follows from cyclicity: Tr(P 1/2 MPMP 1/2 ) = Tr(P 1/2 P 1/2 MPM ) = Tr(PMPM ) = Tr[(MP ) 2 ]. D eff is defined whenever Tr[(MP ) 2 ] > 0, equivalently whenever M does not annihilate every vector in the range of P; this condition holds throughout the empirical anchors of §§3–5. When Tr(A) = 0 the effective dimension is defined as 0 by convention; this degenerate case does not arise in the empirical anchors. When M ⪰ 0, D eff is the participation ratio of the eigenvalues of A = P 1/2 MP 1/2 : for A supported on a single mode, D eff = 1; for r equal nonzero eigenvalues, D eff = r. Eq. (1) generalizes standard inverse participation ratios to the operator pair (P,M ). The dimension-change ratio is χ = D eff (P after ,M )  D eff (P before ,M ),(2) 3

where “before” and “after” denote two values of the control parameter. χ < 1 indicates selection (effective dimension reduced); χ≈ 1 indicates redistribution without net change in dimensionality; χ > 1 indicates dimension expansion (out of scope here, deferred to future work). The commutator mismatch is η(P,M ) = ∥[P,M ]∥ F ∥P∥ F ∥M∥ F ,(3) where ∥·∥ F is the Frobenius norm and [P,M ] = PM − MP. η = 0 iff P and M commute; η > 0 quantifies the misalignment between participation and rigidity. 2.2 Variational characterization Define the action functional A eff [P ;M,T ] = Tr(MP ) − T S[P ],(4) where S[P ] =−Tr(P logP ) is the von Neumann entropy and T > 0 is a positive parameter playing the role of temperature. Stationarity δA eff /δP = 0 under TrP = 1 yields P ∗ (M,T ) = e −M/T Z(M,T ) , Z = Tre −M/T .(5) At the stationary point, [P ∗ ,M ] = 0 exactly, so η(P ∗ ,M ) = 0. Conversely, η(P,M ) = 0 if and only if P shares an eigenbasis with M — a necessary but not sufficient condition for P to coincide with the Gibbs state P ∗ of Eq. (5), since any density diagonal in M’s eigenbasis (not only the Gibbs- weighted one) satisfies [P,M ] = 0. Thus η > 0 detects basis misalignment between participation and rigidity; this includes departures from variational equilibrium but is not synonymous with them. 2.3 Mathematical properties We establish four properties of the construction. Proposition 1 (Dimension bounds under PSD M). For any P ⪰ 0 with TrP = 1 and any M ⪰ 0, 1 ≤ D eff (P,M ) ≤ rank  P 1/2 MP 1/2  . For nonnegative {λ i }, ( P i λ i ) 2

P i λ 2 i

• 2 P i<j λ i λ j ≥ P i λ 2 i ; the upper bound follows from Cauchy–Schwarz on the r nonzero terms. For indefinite Hermitian M, A has eigenvalues of both signs and D eff = [TrA] 2 /Tr(A 2 ) remains real-valued and nonnegative, but is not constrained to the interval [1, rankA]; in the empirical anchors with indefinite M (§§4–5) we report D eff and χ as bare numerical quantities without invoking the strict participation-ratio reading. Verified numerically: 0 violations of the PSD-M bounds across 2000 random (P,M ) pairs. Proposition 2 (Commutator bounds). For Hermitian P and M, 0 ≤ η(P,M ) ≤ √

The lower bound is tight whenever [P,M ] = 0. The upper bound follows from the Böttcher–Wenzel inequality ∥[A,B]∥ F ≤ √ 2∥A∥ F ∥B∥ F for normal matrices [10]. Verified numerically: 0 violations across 5000 random (P,M ) pairs. 4

Proposition 3 (Stationary states are Gibbs). The state P ∗ = e −M/T /Z in Eq. (5) is the unique stationary point ofA eff under TrP = 1, and satisfies [P ∗ ,M ] = 0. Functional differentiation of Eq. (4) with respect to P and Lagrange multiplier λ for the trace constraint gives logP + 1 + M/T + λ = 0, hence P = e −M/T−λ−1 , which fixes λ by normalization. Commutativity follows because P ∗ is a spectral function of M. Proposition 4 (Lindblad invariance). Under any Lindblad (GKLS) dynamics [11, 12], ̇ P =−i[H,P ] + X k  L k PL † k − 1 2 {L † k L k ,P}  .(6) both TrP and positivity are preserved. The diagnostic η is well-defined whenever ∥P∥ F

0 and ∥M∥ F 0, which holds for any non-zero state with M ̸= 0. The ratio χ additionally requires D eff (P before ,M ) > 0 at the reference state; this holds throughout the empirical anchors where the reference is the uniform density P ref = I/N with M = L ̸= 0. We verify this numerically for representative dephasing channels using fourth-order Runge–Kutta integration; explicit Euler is unstable for dt≥ 0.05. 2.4 Empirical estimators Kuramoto. N phase oscillators{θ i } on a graph with adjacency matrix A and Laplacian L = D−A evolve as ̇ θ i = ω i

• K P j A ij sin(θ j − θ i ), with ω i ∼N (0, 1) subject to P i ω i = 0. After a burn-in transient (typically T burn = 50 time units), the participation operator is constructed from time- averaged coherences, P ij =

e i(θ i −θ j )  t ,(7) where for each time t the matrix with entries e i(θ i (t)−θ j (t)) is the rank-1 outer product v(t)v(t) † with v i (t) = e iθ i (t) , hence is positive semidefinite; the time average (and subsequent Hermitian- symmetrization) preserves positive semidefiniteness. We divide by N to enforce TrP = 1. The rigid- ity operator is the graph Laplacian, M = L, held fixed throughout the K-sweep. The order parame- ter is computed using the|⟨r⟩| t convention (modulus of the time-averaged complex order parameter, not the time-average of the modulus); this removes the finite-N baseline that contaminates the lat- ter. The logistic synchronization threshold K c is identified by fitting r(K) = r max /(1+e −(K−K c )/w ), excluding K = 0, with K c constrained to the swept coupling range. In the finite-size scaling sweep (§3.3), seeds whose fit reaches the upper boundary K max within tolerance 0.05 are flagged as clipped and excluded from the ensemble statistics; one seed at N = 12 is excluded on this criterion, giving 52 valid realizations. Floquet. A periodically driven Hamiltonian H(t +T ) = H(t) is integrated over one period to give the Floquet operator U F . Quasi-energies and Floquet states are extracted by Schur decomposition rather than direct diagonalization, since the latter fails at the degenerate quasi-energies induced by discrete symmetries. The participation operator is the time-averaged density matrix over one period; M is the static (undriven) part of H. Model DSI spectrum. A Hamiltonian with explicit log-periodic spectrum, E n = E 0 λ n , is con- structed diagonally. P (μ) is a Gaussian-weighted projector centered at chemical potential μ with relative width σ rel = 0.025; M is a fixed random Hermitian operator of unit Frobenius norm. The 5

control parameter μ is swept logarithmically. The DSI ratio λ is recovered from η(logμ) by mini- mizing the root-mean-square deviation between curves rescaled by candidate ratios λ test relative to a reference run. 2.5 Summary of(P,M) instantiations Table 1: Summary of the participation operator P, rigidity operator M, and control parameter for each empirical anchor. DomainP (participation)M (rigidity)Control KuramotoTime-averaged phase-coherence matrix, P ij =⟨e i(θ i −θ j ) ⟩ t /N Graph Laplacian L = D− A Coupling K Floquet (kicked TFIM) Block-dephased projection of thermal ref- erence P before ; see Appendix B.1 Ising rigidity H z

−J P i σ z i σ z i+1 Drive h DSI (model spectrum) Gaussian-weighted projector, P nn (μ) ∝ exp[−(E n − μ) 2 /(2σ 2 )] Fixed random Hermitian M, ∥M∥ F

√ N μ(log sweep) 2.6 Software and reproducibility All simulations were performed in Python with NumPy and SciPy. Source code, random seeds, and saved data files are provided in the supplementary material. Key implementation choices: (i) fourth- order Runge–Kutta for any Lindblad evolution; (ii) Schur decomposition (scipy.linalg.schur) for Floquet operators with potential degeneracies; (iii) logistic fit excluding K = 0 when estimating K c . 3 Kuramoto results We test the framework on the Kuramoto model of coupled phase oscillators, the canonical setting for synchronization transitions in coupled dynamical systems. This section presents results at four levels of empirical pressure: (i) the basic precursor result at fixed network size and topology; (ii) robustness across network topologies and a 32-fold range of system sizes; (iii) the temporal precursor under a slow-ramp protocol; and (iv) a direct head-to-head comparison against pairwise transfer entropy on the same simulation data. 3.1 Setup and protocol The Kuramoto dynamics on a graph with adjacency matrix A and Laplacian L = D− A read ̇ θ i = ω i

• K X j A ij sin(θ j − θ i ),(8) with intrinsic frequencies ω i ∼ N (0, 1) centered so that P i ω i = 0. We integrate Eq. (8) with time step dt = 0.025, allow a transient T burn that depends on system size, then time-average the phase-coherence matrix P ij = ⟨e i(θ i −θ j ) ⟩ t over a measurement window of length T meas . We enforce 6

TrP = 1 by dividing the matrix by N, and use the modulus-of-average convention r = |⟨e iθ ⟩ t | for the order parameter, which removes the finite-N baseline that contaminates the alternative average- of-modulus form. The rigidity operator is the graph Laplacian, M = L, held fixed throughout the K-sweep. For each realization we identify two characteristic couplings: the η-peak location K η = arg max K η(K) and the logistic synchronization threshold K c obtained from a three-parameter fit r(K) = r max /(1 + e −(K−K c )/w ) to the measured r values, excluding K = 0. Throughout §§3–4, ± values denote stan- dard deviations (SD) across ensemble realizations unless explicitly labeled as standard errors of the mean (SEM). 3.2 Steady-state K-sweep at fixed network size We first establish the precursor result at fixed network size N = 12 on Erdős–Rényi networks with mean degree d = 4. Across 20 ensemble realizations (independent networks, frequencies, and initial conditions), Figure 1 shows the per-seed η(K), χ(K), and r(K) traces with their ensemble means. In every realization, η(K) rises from zero at K = 0, peaks at a characteristic coupling K η , and decays toward zero as r approaches saturation. The ensemble mean K η = 0.29± 0.10 precedes K c = 0.65± 0.36 by a precursor gap ⟨K c − K η ⟩ = 0.36± 0.36, positive in 19 of 20 realizations (one-sample z = 4.49 against the null of zero mean lead). The remaining realization had the gap within K-sampling resolution of zero. The wide spread in K c relative to K η at this small system size is consistent with the finite-size noise that the N-scaling analysis in §3.3 subsequently shows to contract substantially as N grows. The dimension-change diagnostic χ falls monotonically from unity at K = 0 toward a plateau at K≳ 0.6, consistent with selection rather than dimensional expansion: the system reorganizes onto a smaller effective subspace as it synchronizes. 7

Figure 1: Basic precursor result on the Kuramoto model. N = 12 oscillators on Erdős–Rényi networks with mean degree d = 4, across 20 ensemble realizations. (a) The commutator mismatch η(K) rises sharply from zero, peaks at ⟨K η ⟩ = 0.29± 0.10, and decays as the system synchronizes. Per-seed traces (light red); ensemble mean and standard-deviation band (dark red, shaded). (b) The effective-dimension ratio χ(K) = D eff (K)/D eff (0) falls monotonically from unity toward a saturating plateau, indicating dimensional selection rather than expansion. (c) The order parameter r(K) rises through the logistic threshold ⟨K c ⟩ = 0.65± 0.36; the dashed (red) and dotted (black) vertical lines mark ⟨K η ⟩ and ⟨K c ⟩ respectively, and the shaded gold band marks the ensemble-mean precursor gap. (d) Distribution of precursor gaps K c −K η across the 20 realizations: positive in 19, with mean 0.36± 0.36 and one-sample z = 4.49 against the null of zero mean lead. The wide gap-distribution at N = 12 contracts with system size (Fig. 2). 3.3 Topology and finite-size robustness To test that the precursor result is not specific to ER networks at N = 12, we run the same protocol on five conditions sampling four topology classes: ER at N = 12 and N = 24, Watts–Strogatz at N = 12 (rewiring probability 0.1), Barabási–Albert at N = 12 (m = 2), and random-regular at N = 12 (d = 4). All graphs use mean degree d = 4 where applicable. With 15 realizations per condition, the η-peak precedes K c in 73 of 75 cases (97.3%). Two realizations in the topology-comparison study did not show a leading η peak: both occurred in the N = 12 Erdős–Rényi (ER) condition, where finite-size fluctuations produce the widest gap 8

variance in the ensemble (see Table 6, N = 12 row, SEM = 0.08). No non-leading realizations were observed in the finite-size scaling sweep, which sampled the same N = 12 condition with 19 valid seeds and recorded a positive lead in all 19. The discrepancy is therefore consistent with stochastic variation at small system size rather than a topology-dependent failure mode. We next turn to finite-size scaling, holding topology fixed to ER so that only N varies. We then test finite-size scaling by holding the mean degree fixed at d = 4 and varying N ∈{12, 24, 48, 96, 192, 384} on ER networks (Figure 2). Because per-realization compute scales as N 2 , the number of seeds decreases with N (20, 12, 8, 6, 4, 3 respectively) (one seed excluded at N = 12 where the logistic fit reached the sweep boundary), giving 52 valid realizations in total. The lead is positive in every valid realization at every size: 52/52 pooled across the N-scaling sweep. The precursor gap is 0.38± 0.08 (SEM) at N = 12. Across the large-N regime N ≥ 96, where the logistic K c estimate is unambiguous (no boundary clipping), the gap is constant at ⟨K c − K η ⟩ = 0.607± 0.045 (SEM) (χ 2 /dof = 0.35 against a constant). A power-law or logarithmic decay fits substantially worse (χ 2 /dof = 2.3). At small N (12–48) the gap shows larger scatter and sensitivity to the K c -estimation protocol; we therefore base the large-N statement on the N ≥ 96 points. We do not fit a parametric saturating form: with six size points, the asymptotic value and approach exponent are not jointly constrained. What the data establish is that the gap does not shrink across the 32-fold range in N examined, and that the positive lead in 52 of 52 valid realizations holds across all six values of N. Figure 2: Finite-size scaling of the Kuramoto precursor result. Erdős–Rényi networks at fixed mean degree d = 4, with N ∈ {12, 24, 48, 96, 192, 384} and per-size ensembles of 20, 12, 8, 6, 4, 3 realizations respectively. (a) Logistic synchronization threshold⟨K c ⟩ (blue circles) and commutator- peak coupling⟨K η ⟩ (red stars) versus N, with error bars showing ensemble standard deviation. Both decrease with N but K η decreases faster, opening the precursor gap. (b) Precursor gap ⟨K c − K η ⟩ versus N with error bars showing ensemble standard error of the mean (SEM). The gap is positive in all 52 valid realizations and does not shrink with N; across N ≥ 96 (unshaded), where K c estimation is unambiguous, it is constant at 0.607± 0.045 (SEM; red band). The small-N region (shaded) shows larger scatter and K c -protocol sensitivity. No parametric saturating fit is applied. (c) Realization standard deviations σ(K c ) (blue) and σ(K η ) (red) versus N on log–log axes, showing the contraction of finite-size noise with system size. 3.4 Slow-K-ramp temporal precursor The K-sweep is a steady-state protocol: at each K, the system is equilibrated before measurement. To test whether the precursor signal survives in real-time dynamics — where the coupling itself evolves — we run a slow-ramp experiment with N = 24, K(t) = K max (t/T ramp ), K max = 1.5, 9

and T ramp = 800 time units. Phases are pre-equilibrated at K = 0 for T pre = 80 to erase initial- condition memory, then evolved under the ramp. We compute sliding-window η(t), r(t), and χ(t) with a window of 40 time units, sampled every 1 time unit. For each realization we identify two onset times: t peak η , the time at which the operator misalignment η(t) is maximal, and t half r , the time at which r(t) first reaches half of its asymptotic value. Figure 3 shows the ensemble-mean trajectories on both time and K(t) axes with the detector times marked. Across 8 ensemble realizations, t peak η precedes t half r in all 8, with mean temporal lead⟨∆t⟩ = 152± 72 time units and corresponding K-space lead ⟨∆K⟩ = 0.29± 0.13. Two features of the slow-ramp result warrant comment. First, the K-space lead ⟨∆K⟩ = 0.29 is smaller than the steady-state large-N value 0.61 from the K-sweep. Two effects contribute: the slow-ramp uses the half-asymptote threshold of r rather than the logistic midpoint K c (the half- asymptote lies at lower K), and at any finite ramp rate the system slightly lags steady-state. Second, the η signal during the ramp sits on a finite-window measurement baseline of ∼ 0.12 and rises only to ∼ 0.13 at the peak before decaying to ∼ 0.04 in the synchronized regime. The relative bump is modest at N = 24 and would likely become cleaner at larger system sizes (window-baseline noise scales as 1/ √ T window ). The temporal lead is nonetheless recoverable in every realization at this size. 3.5 Head-to-head against transfer entropy The preceding sections establish that K η < K c in the steady state and that this precedence carries over to real-time dynamics. They do not establish that η is a more sensitive precursor than existing information-theoretic alternatives. The most direct competitor for the synchronization-onset case is pairwise transfer entropy [8], which has been shown to peak near the Kuramoto transition and decay on both sides [13, 14]. We compute both η and pairwise TE on the same simulation runs: N = 24, ER networks at d = 4, K ∈ [0, 2.5] on 30 values, T meas = 200 time units, 8 ensemble realizations. TE is computed via symbolic phase binning with n bins = 4, lag τ = 1, averaged over 60 randomly selected ordered pairs of oscillators per K value; the same set of pairs is used across all estimator configurations within a given (seed,K). We then locate the TE peak K TE = arg max K TE(K) for each seed. Figure 4 shows the four panels. The η(K) curves (panel a) cluster tightly around a common peak at K η = 0.23± 0.06; the TE(K) curves (panel b) show substantially wider seed-to-seed scatter, with K TE = 0.54± 0.31. Panel (c) shows the temporal sequence on normalized scales: η peaks first, TE peaks second, and the order parameter r rises through the logistic threshold K c = 1.05± 0.66 last. Panel (d) shows the per-seed scatter of (K η ,K TE ): in 7 of 8 realizations K TE

K η strictly, and in the one remaining realization (the seed with the lowest K c ) the two coincide. Table 2 summarizes the comparison. Three quantitative claims follow. (1) η peaks earlier. ⟨K TE − K η ⟩ = 0.31 in coupling units. In every realization, the operator- misalignment peak precedes or coincides with the information-transfer peak. (2) η is more reproducible. σ(K η ) = 0.060 versus σ(K TE ) = 0.313, a factor of 5.2. The coefficient of variation σ/μ is 0.27 for η versus 0.58 for TE. 10

Figure 3: Slow-K-ramp temporal precursor experiment. N = 24 oscillators on Erdős–Rényi net- works with mean degree d = 4. The coupling is ramped linearly from K = 0 to K max = 1.5 over T ramp = 800 time units, following a pre-equilibration at K = 0. Sliding-window diagnos- tics over a 40-time-unit window. (a) Ensemble-mean η(t) (red) and r(t) (black, dashed) versus time, with standard-deviation bands shaded. Vertical lines mark ⟨t peak η ⟩ (red, dotted) and ⟨t half r ⟩ (black, dotted); the shaded gold region marks the mean temporal lead ⟨∆t⟩ = 152± 72 time units. (b) Same data with abscissa reparameterized as K(t) to show the corresponding K-space lead ⟨∆K⟩ = 0.29± 0.13. (c) Distribution of temporal leads ∆t = t half r −t peak η across 8 ensemble realiza- tions: all 8 positive. (d) Per-seed scatter of slow-ramp K peak η vs K half r , compared to the steady-state K-sweep reference at N = 24 (blue star with error bars). All ramp points lie above the no-lead diagonal. 11

(3) Both lead K c . η leads K c by 0.82± 0.65, TE leads by 0.51± 0.52, both positive in all 8 realizations. The standard deviations on these lead values are inflated by two slow-synchronization seeds where K c approaches our K max cutoff; the σ-ratio statistic in claim (2), which depends only on K η and K TE and not on K c , is unaffected and is the more robust quantitative summary. Table 2: Head-to-head comparison of η and transfer entropy (TE) as precursors of the Kuramoto synchronization transition. Values are means ± standard deviation across n seed = 8 realizations (N = 24 oscillators on Erdős–Rényi networks with mean degree 4, K max = 2.5). TE computed via symbolic phase binning (n bins = 4, lag τ = 1). QuantityηTERatio (TE/η) Peak coupling ⟨K peak ⟩0.23± 0.06 0.54± 0.312.4× Standard deviation σ(K peak )0.0600.3135.2× Coefficient of variation σ/μ0.270.582.2× Lead relative to K c (mean)0.82± 0.65 0.51± 0.52n/a Positive lead (fraction of seeds)8/88/8n/a K TE

K η per seed (strict)——7/8 K TE = K η per seed——1/8 3.6 Robustness of the comparison to TE estimator choice A potential concern is that the TE-peak location depends on the discretization parameters chosen for the symbolic estimator. We test three alternative configurations on the same simulation data: (n bins ,τ )∈{(3, 1), (5, 1), (4, 2)}, with the same sampled pairs per (seed,K). Appendix A (Table 5, Figure 7) reports the results: across the 32 (seed, configuration) entries, only 2 changed. The configuration mean⟨K TE ⟩ varies by less than±0.02 across the four configurations; the seed-to-seed spread σ(K TE ) remains in the range [0.313, 0.319]; and the strict-inequality count K TE K η is 7 of 8 in every configuration. The σ-ratio relative to σ(K η ) = 0.060 therefore ranges from 5.22× to 5.32×. The empirical claim that η is the more reproducible precursor on this benchmark is robust to the estimator choice within the TE family. 4 Floquet anchor: regime distinction under periodic driving The (χ,η) construction extends without modification to periodically driven systems. We demon- strate this on the periodically kicked transverse-field Ising chain (N = 4, periodic boundary con- ditions), sweeping drive strength h ∈ [0, 2.5] with the rigidity operator fixed at M = H z = −J P i σ z i σ z i+1 . In §3, “before” and “after” refer to two values of the coupling K along a steady- state sweep; here they refer to states before and after the drive is applied at a fixed h. The “before” state is the thermal Gibbs density of H z at temperature T th = 1.5, for which [P before ,H z ] = 0 exactly and η before = 0 by construction. The “after” state is the Floquet-diagonal projection of P before —the steady state a small dephasing in the Floquet basis would produce—extracted via Schur decomposition (Appendix B). For every h > 0 sampled, the trajectory in the (χ,η) plane sits with χ < 1 and η > 0: the sustained- coherence quadrant. This is geometrically distinct from the strongly synchronized Kuramoto state at K ≫ K c , which has χ low and η small—the selection-relaxation quadrant—and any scalar 12

0.00.51.01.52.02.5 K (coupling) 0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 ( K ) (a) (K) for 8 seeds ensemble mean mean ± std K= 0.23 ± 0.06 0.00.51.01.52.02.5 K (coupling) 0.00 0.01 0.02 0.03 0.04 0.05 TE( K ) (bits) (b) TE(K) for 8 seeds ensemble mean mean ± std K TE = 0.54 ± 0.31 0.00.51.01.52.02.5 K (coupling) 0.0 0.2 0.4 0.6 0.8 1.0 normalized value KK TE K c (c) Normalized , TE, r (ensemble means) (norm.) TE (norm.) r (order param.) 0.00.20.40.60.81.01.21.4 K (peak coupling) 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 K TE (peak coupling) K TE

K : 7/8 seeds K TE = K : 1/8 (lowest-K c seed) seed 6 (K c = 1.59) (d) Per-seed peak locations y = x Figure 4: Head-to-head comparison of η and pairwise transfer entropy (TE) as precursors of the Kuramoto synchronization transition. Both diagnostics are computed from the same simulation data (N = 24 oscillators on Erdős–Rényi networks with mean degree 4, T meas = 200 time units, n seeds = 8, K ∈ [0, 2.5]). TE is computed with symbolic phase binning (n bins = 4, lag τ = 1 on samples spaced 0.25 time units), averaged over 60 randomly sampled ordered pairs of oscillators per coupling value. (a) η(K) for each seed (thin lines) and ensemble mean ± std (thick line, shaded band). All eight curves peak in a narrow window around K η = 0.23± 0.06. (b) TE(K) for the same seeds, converted to bits. Substantially wider seed-to-seed scatter, with K TE = 0.54± 0.31. (c) Normalized ensemble means show the temporal sequence: η peaks first, then TE, then the order parameter r rises through K c ≃ 1.05. (d) Per-seed peak locations. In 7/8 realizations K TE K η strictly; the one seed on the diagonal is the realization with the lowest K c , where both precursor diagnostics fire simultaneously at the very early transition. The outlier at (K η ,K TE ) = (0.17, 1.29) is the slow-transition seed (K c = 1.59). 13

precursor diagnostic collapses this separation. The fixed-M convention is essential: reassigning M to the stroboscopic Floquet Hamiltonian H F at each h places η ≡ 0 identically, since P after is diagonal in the Floquet basis by construction. Full details—setup, the (χ,η) trajectory, drive- strength dependence, and the Schur-decomposition rationale—are given in Appendix B. 5 Discrete-scale-invariance anchor: recovery of log-periodic struc- ture The Kuramoto and Floquet anchors establish that the (χ,η) construction applies to equilibrating and driven Hamiltonian systems respectively. The discrete-scale-invariance (DSI) anchor tests a different question: when a system carries hidden log-periodic structure in its spectrum, does the operator diagnostic recover that structure quantitatively? This anchor’s role is methodological validation: given a controlled input spectrum with known log-periodic structure, we check that the same (χ,η) construction used in §§3–4 reads out the input ratio with quantitative accuracy, validating its sensitivity to hidden scale structure without modification. The DSI anchor does not claim to detect emergent DSI from a physical model; that extension is left to future work (caveat (i), §5). 5.1 Setup Physical instances of DSI include the Efimov tower in three-body atomic physics [15, 16] and log- periodic oscillations in the magnetoresistance of certain topological materials under strong magnetic fields [4]. These systems share a recursive spectrum structure E n ∝ λ n over many decades of energy, with a characteristic ratio λ that is not directly registered by standard scalar order parameters. We construct a Hamiltonian with an explicit geometric spectrum, H = diag(E 0 , E 0 λ, E 0 λ 2 , ..., E 0 λ N−1 ),(9) with N = 36, E 0 = 0.05, and DSI ratio λ swept across five values λ ∈ {1.15, 1.25, 1.40, 1.60, 1.85}. The spectrum is log-periodic by construction: logE n+1 − logE n = logλ independent of n. The rigidity operator M is a fixed random Hermitian matrix (drawn once, seeded for reproducibil- ity), normalized so∥M∥ F

√ N. It is generically indefinite, so D eff is interpreted in the generalized sense described in §2. The participation operator is a Gaussian-weighted projector, P nn (μ) = 1 Z(μ) exp  − (E n − μ) 2 2σ 2  , σ = σ rel μ,(10) with σ rel = 0.025, diagonal in the energy eigenbasis and normalized so TrP = 1. The control parameter μ is swept logarithmically over the interior of the spectrum (μ ∈ [E 3 ,E N−4 ], omitting four boundary eigenvalues on each end) at 1200 sample values. 5.2 Diagnostic signature of DSI Figure 5(a,b) shows the diagnostics for the representative case λ = 1.40. The commutator mismatch η(logμ) oscillates with the eigenvalue spacing: the curve rises and falls each time the projector center crosses one of the levels E n . The effective-dimension ratio χ(logμ) shows the same structure as a 14

sequence of discrete drops; at each eigenvalue, χ falls sharply, indicating dimensional selection onto the Gaussian-broadened single-eigenstate manifold. The vertical gray lines mark the eigenvalues, and both diagnostics inherit the spectrum’s log-periodic spacing. Within each log-period, η has internal substructure — multiple local maxima as the projector transitions across the boundary between adjacent eigenstates — which makes naive period extraction by Fourier peak-finding or autocorrelation unreliable and motivates the universal-collapse approach we use below. Figure 5(c) overlays the normalized η(logμ/ logλ in ) curves for all five values of λ in . When the abscissa is rescaled by the input DSI ratio, the five curves collapse onto a single universal shape with no free parameter. The collapse is the central evidence that the operator diagnostic correctly inherits the spectrum’s log-periodicity: η(logμ) is a function of logμ/ logλ alone, modulo a λ- independent overall scale. Note on the collapse. For diagonal P (μ) with relative Gaussian width σ = σ rel μ, the shift μ → λμ maps P nn (μ) exactly to P n−1,n−1 (λμ). Exact periodicity η(λμ) = η(μ) would follow if M satisfied the shift symmetry M ij = M i+1,j+1 ; a generic random Hermitian M does not have this property, so the collapse is not an exact mathematical consequence of the operator definitions for a fixed M. For a random M with i.i.d. entries, the shift symmetry holds in distribution, and deviations for a fixed realization are O(1/ √ N ) by concentration. The empirical collapse in Fig. 6(c) and its robustness across 20 random M realizations in §5.4 are therefore consequences of this concentration, not theorems. A proof for fixed M would require either an explicit concentration bound or additional assumptions on M’s spectral statistics (e.g. shift-invariance of |M ij | 2 ). This is stated as a caveat rather than a gap: the DSI anchor’s claim is empirical — the diagnostic recovers log-periodic structure when that structure is present — and the concentration argument provides the mechanism, but the result should not be read as a derived exactness statement. 5.3 Quantitative recovery of λ To recover the DSI ratio from the diagnostic alone we use the universal-collapse principle in reverse: for each input λ in we ask which candidate λ test best collapses the rescaled η(logμ/ logλ test ) curve onto a fixed reference. We use the λ in = 1.40 run as the reference and search over candidate ratios λ test ∈ [1.05, 2.0] on a grid of 100 values, minimizing the root-mean-square deviation between the rescaled curve and the reference on a common abscissa. Specifically, both curves are evaluated on a common logarithmic abscissa grid with 500 uniformly spaced points in logμ/ logλ∈ [1, 9], linearly interpolated from the simulation samples, and normalized to unit maximum before computing the RMS deviation. Table 3 reports the recovered ratios. The mean absolute relative error is 0.31% across the five inputs; the worst-case error is 0.41%. Figure 5(d) plots recovered against input λ, with all five points lying on the identity line to within the marker size. Figure 5(f) shows the collapse-RMS landscape for input λ in = 1.60: a single deep, narrow minimum at λ test ≈ 1.61, with no spurious local minima in the search range. The recovery is unambiguous. 5.4 Robustness to the rigidity-operator realization The rigidity operator M used in §5.B is a single fixed random Hermitian matrix. To test whether the recovery accuracy depends sensitively on this choice, we repeat the full pipeline (five input λ values; 15

Figure 5: Discrete-scale-invariance anchor: recovery of log-periodic structure from the operator di- agnostic. (a) η(logμ) for the representative case λ in = 1.40; vertical gray lines mark the eigenvalues E n . (b) χ(logμ) for the same case, showing discrete sharp drops at each E n (dimensional selec- tion onto the Gaussian-broadened single-eigenstate manifold). (c) Universal collapse: normalized η(logμ/ logλ in ) for all five input values λ in ∈ {1.15, 1.25, 1.40, 1.60, 1.85} overlaid on a common rescaled abscissa. The curves collapse onto a single universal shape, demonstrating that η(logμ) is a function of logμ/ logλ alone. (d) Recovered λ from collapse-RMS minimization (blue circles) versus input λ. All five points lie on the identity line (dashed) within marker size. (e) Per-input relative error in recovered λ; mean absolute error 0.31%, worst case 0.41%. (f) Collapse-RMS land- scape for λ in = 1.60 as a function of candidate λ test . A single sharp minimum at λ test ≈ 1.61 with no spurious local minima recovers the input ratio unambiguously. Table 3: DSI ratio recovery via collapse-RMS minimization on η(logμ). Reference: λ in = 1.40. λ in λ recovered Relative error 1.151.146−0.35% 1.251.252+0.12% 1.401.396−0.32% 1.601.607+0.41% 1.851.856+0.33% Mean absolute error0.31% 16

collapse-RMS recovery against the λ = 1.40 reference) for 20 independently drawn M realizations, each constructed as (A + A † )/2 from a complex matrix A with N (0, 1) real and imaginary parts. Figure 6(a) shows the distribution of per-seed mean absolute recovery error across the 20 realizations: the mean is 0.25± 0.03%, with range [0.20%, 0.35%]. The original realization reported in Table 3 (mean error 0.31%) sits near the upper end of this distribution and is therefore representative, not anomalous. Figure 6(b) shows per-input-λ error scatter. For the two smallest inputs (λ = 1.15 and λ = 1.25), the recovered ratio is identical across all 20 realizations to within the collapse-test grid resolution (∆λ ≈ 0.01): the short log-period samples the spectrum densely enough that M- dependent noise averages out. For the two largest inputs (λ = 1.60 and λ = 1.85), three of twenty realizations produce outlier recoveries, but the worst-case relative error across the full 20× 5 grid of (realization, input) is 0.85%. The recovery is therefore robust to the choice of M at the precision relevant to the paper’s claims. 0.200.220.240.260.280.30 mean absolute recovery error (%) 0 1 2 3 4 5 count (a) Error across 20 rigidity operators mean=0.25% original=0.31% 1.151.251.401.601.85 input scaling ratio 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 relative recovery error (%) (b) Per-input recovery error scatter Figure 6: Robustness of the DSI recovery to the random rigidity-operator realization. The full pipeline of §5.B–C (five input λ values; collapse-RMS recovery against the λ = 1.40 reference) is repeated for 20 independently drawn M realizations. (a) Distribution of per-seed mean absolute recovery error. Across the 20 realizations the mean is 0.25% (blue solid line) with standard deviation 0.03%. The original M realization used in Table 3 (red dashed line at 0.31%) sits near the upper end of the distribution. (b) Per-input-λ relative error scatter across all 20 realizations. For the two smallest inputs (λ = 1.15, 1.25), the recovered ratio is identical across realizations to within the collapse-test grid resolution. For the two largest inputs (λ = 1.60, 1.85), three of twenty realizations produce outlier recoveries, but the worst-case relative error across the full 20× 5 grid is 0.85%. 5.5 What this anchor validates, and what it does not The DSI anchor establishes a specific and limited claim: when a system carries log-periodic structure in its spectrum, the (χ,η) diagnostic detects and quantitatively recovers that structure with sub- percent accuracy. Four honest caveats temper any broader interpretation. (i) Engineered, not derived. The log-periodic spectrum is imposed by construction, not derived from microscopic physics. This distinction is stated at the opening of this section: the anchor validates sensitivity to log-periodic structure, not discovery of emergent DSI. Demonstrating the latter on a real HfTe 5 band-structure calculation, on a renormalization-group flow with complex critical exponents, or on the Efimov tower [16], is the natural follow-up and is left to future work. 17

(ii) Random rigidity operator. The probe M is a fixed random Hermitian matrix rather than a physically motivated operator (e.g. a transport operator, a response function, or a band-structure observable). As shown in §5.4, the recovery accuracy is essentially M-independent across 20 re- alizations, so this choice is not load-bearing for the validation. A physically motivated M would, however, tie the demonstration more closely to specific materials applications. The universal collapse underpinning the λ-recovery procedure is an empirical consequence of con- centration for typical random matrices, not a theorem about any specific fixed M. The derivation is exact only for shift-invariant (circulant) M. (iii) Resolution-limited. The Gaussian width σ rel = 0.025 is narrow enough that P (μ) is well- localized on individual eigenvalues. The recovery accuracy degrades when σ rel becomes comparable to logλ (the eigenvalues smear into a continuum and the log-periodic structure of η(logμ) blurs out). We have not systematically explored sensitivity to this parameter. (iv) Not a comparative claim. The recovery comparison performed here is against the ground- truth input λ, not against an alternative DSI-detection method. A direct spectral analysis of the eigenvalues {E n } would trivially recover λ as well, and we make no claim that η is a more sensitive DSI detector than direct spectroscopy. The contribution of the DSI anchor is methodological — demonstrating that the same (χ,η) operator construction used for synchronization (§3) and driven dynamics (§4) extends cleanly to spectral DSI without modification — not comparative. Given these caveats, what the anchor provides is a methodological proof of principle: the operator- based diagnostic correctly reads out hidden log-periodicity, with mean recovery error 0.3% across a 1.6× range in λ. The framework passes its validation test. 6 Discussion We have introduced a two-dimensional operator-based diagnostic (χ,η) for detecting reorganization in coupled dynamical systems. The construction rests on a participation operator P and a fixed rigidity operator M, organized by a free-energy-like variational principle whose stationary states are Gibbs-like (§2). Three empirical anchors test the construction across qualitatively distinct domains: the Kuramoto model under steady-state and slow-ramp protocols, with a direct head- to-head against pairwise transfer entropy (§3); a periodically kicked transverse-field Ising chain in the Floquet steady state (§4); and a model spectrum with engineered discrete scale invariance (§5). The framework’s defining empirical claim — that the η-peak precedes the order-parameter signal on Kuramoto with a precursor gap that remains positive and does not decay across a 32-fold range in system size (holding at 0.61± 0.05 in the large-N regime) and substantially lower seed-to-seed variance than pairwise transfer entropy — holds across 73 of 75 ensemble realizations across four network topologies at N = 12, and across all 52 valid realizations in the six-size N-scaling sweep, remains stable for system sizes from N = 12 to N = 384, and is robust to the choice of TE estimator hyperparameters. We now position the construction against four adjacent lineages of operator-theoretic work that a reader from each subfield will reach for. None of these is a direct competitor; each is a foundational anchor whose techniques our construction reuses or whose ideas it develops in a different direction. 18

6.1 Adjacent lineages Mori–Zwanzig projection-operator formalism. The Mori–Zwanzig approach [17, 18] is the historical origin of using projection operators to organize coarse-grained dynamics. There, a projec- tor P separates the relevant subspace from the irrelevant one, and the off-diagonal couplings QLP generate memory kernels and noise via the Nakajima–Zwanzig equation. Our P shares the role of selecting “what participates,” but is used differently: rather than projecting equations of motion onto a slow manifold, we use P as a steady-state observable and combine it with a fixed reference M to generate diagnostic scalars. The Mori–Zwanzig literature describes how the projection is used; we describe how the projected state itself is diagnosed. Recent extensions to time-dependent Hamiltonians [19] bring the formalism closer to the Floquet setting we examined in §4 and would be a natural starting point for connecting the two formalisms. Generalized inverse participation ratios. The effective dimension D eff (P,M ) = [Tr(MP )] 2 /Tr[(MP ) 2 ] is the participation ratio of the eigenvalues of A = P 1/2 MP 1/2 . This generalizes the standard in- verse participation ratio [20, 21] for eigenstate localization to operator pairs: The standard IPR, IPR(ψ) = P i |ψ i | 4 , measures wavefunction amplitude concentration in a fixed basis. D eff (P,M ) generalizes the participation-ratio concept to operator pairs but does not reduce to IPR −1 for rank- one P in general: for P =|ψ⟩⟨ψ| and M diagonal in the localization basis, D eff

|⟨ψ|M|ψ⟩| 2 ⟨ψ|M 2 |ψ⟩ , which equals 1 when ψ is an eigenstate of M and otherwise measures the spread of ψ over M’s eigenstates — a distinct quantity from P i |ψ i | 4 . What is added by the present construction is the (P,M )-symmetric structure of D eff and its role as the leading term in the variational principle of §2; the connection to standard IPR is one of motivation and analogy, not algebraic identity. Laplacian-eigenvector diagnostics for synchronization. McGraw and Menzinger [22] intro- duced the Laplacian eigenvectors as a diagnostic for partial synchronization in oscillator networks, framing synchronization onset as “a series of quasi-independent transitions involving different nor- mal modes.” Their diagnostic is the participation of the oscillator state in each Laplacian eigen- mode, mode by mode. Our η = ∥[P,L]∥ F /(∥P∥ F ∥L∥ F ) collapses the same physics — alignment of the participating state with the Laplacian eigenbasis — into a single operator-norm scalar. The McGraw–Menzinger formalism is more fine-grained per mode; ours is more compact and admits cross-domain generalization (the Floquet and DSI anchors use the same scalar with a different M). The two approaches are complementary on Kuramoto specifically: a direct combination — McGraw–Menzinger per-mode decomposition alongside the scalar η — would provide both where and how strongly the operator misalignment lives. Frobenius commutator measures of quantum asymmetry. Yao and coauthors [23] use the Frobenius commutator∥[U (g),ρ]∥ F as a measure of quantum coherence and asymmetry with respect to a group action U (g). The mathematical object is the same as our η with P = ρ (the density matrix) and M = U (g) (a unitary symmetry generator). The interpretation is different: they measure static asymmetry under a fixed symmetry, while we sweep a control parameter and locate the commutator peak as a precursor. The underlying inequality ∥[A,B]∥ F ≤ √ 2 ∥A∥ F ∥B∥ F [10] provides the upper bound in both settings (our Proposition 2). 19

6.2 Limitations The present results establish superior reproducibility relative to pairwise transfer entropy on the Kuramoto benchmark. Several limitations bound the scope of this claim. First, it is not yet known whether the advantage of η over TE generalizes to other synchronization models, to higher-order information-theoretic measures, or to experimentally observed transitions where ground-truth K c is unavailable. Second, the η estimator carries a bin-width hyperparameter whose sensitivity at small N has not been systematically characterized; future work should establish whether the observed lead is robust across estimator configurations in the N ≤ 24 regime. Third, the DSI anchor validation was conducted on a single log-periodic template; generalization to other scaling symmetries remains an open question. 6.3 Outlook Three directions stand out for follow-up work, in order of methodological cost. Head-to-head against synergy and Koopman-based EWS. On the Kuramoto benchmark, computing the synergistic information component from partial information decomposition would provide the direct comparison against the Marinazzo synergy precursor [7] that the literature scan flagged as the closest information-theoretic competitor. The Koopman-operator EWS framework [9] is most naturally applied to the Floquet anchor and would extend the comparison there. Both are within reach with the simulation data already in hand. Materials-realistic anchors. For each of the three domains a physical realization is available. Real synchronization networks (cardiac myocytes, neural populations, power grids), real driven quantum systems (Floquet-engineered solids, cold-atom Floquet topological insulators), and real DSI materials (HfTe 5 at high magnetic field) provide test data that would push the framework beyond toy models. The principal methodological obstacle is the choice of M for each case, which our framework currently leaves to the practitioner. Dimension-expanding regime (χ > 1). The fourth quadrant of the (χ,η) plane, where the ef- fective dimension grows under the control-parameter sweep, was deliberately excluded from this pa- per’s scope. Such regimes appear naturally in dimension-expanding processes — biological growth, learning systems, and active matter undergoing morphogenesis — and a treatment of the (χ,η) diagnostic for these settings would complete the four-quadrant geometric organization. 7 Fixed-M convention: empirical optimality on rewiring graphs 7.1 Fixed-M versus floating/lagged M The fixed-M convention — holding the rigidity operator M constant throughout the control- parameter sweep — was introduced in §2 as a technical requirement: allowing M to co-vary with the instantaneous state forces η ≡ 0 by construction, since the participation operator P is then always diagonal in M’s eigenbasis (Appendix B). This makes the floating-M convention vacuous as a precursor diagnostic. The fixed-M convention avoids this degeneracy by anchoring M to a chosen 20

reference state and measuring how far P drifts from that reference as the system evolves. What was initially a constraint turns out to be the correct design choice for precursor detection, as we now establish empirically. Setup. We test three M conventions on the slow-ramp Kuramoto protocol of §3.4 (N = 24, Erdős– Rényi networks, K ramped linearly from 0 to 1.5 over 120 time units, 8 ensemble realizations per condition), extended to a stochastically rewiring graph. At each time step, each edge independently flips (present ↔ absent) with probability p rewire · dt, so that the graph Laplacian L(t) evolves during the ramp. We test four rewiring rates p rewire ∈ {0, 0.005, 0.02, 0.05} and five lag values τ ∈{0, 1, 3, 6, 10} time units. The three diagnostics are: • η fixed : M = L(t=0), the graph Laplacian at the start of the ramp. This is the paper’s construction throughout §§ 3–5. • η inst : M = L(t), the instantaneous Laplacian. For the static graphs of §3 this reduces to η fixed ; for rewiring graphs it tracks the current structure. • η lag : M = L(t− τ ), the Laplacian evaluated τ time units in the past. At τ = 0 this is η inst ; as τ →∞ it approaches η fixed . For each seed and condition we compute the precursor lead — the time elapsed between the η peak and the half-asymptote of r(t) — and report mean ± standard deviation across the 8 seeds. Results. Table 4 reports the lead and variance at τ = 3 across rewiring rates. Table 4: Precursor lead and seed-to-seed variance at τ = 3 time units, across four rewiring rates. Values are mean± standard deviation across n seeds = 8 realizations (N = 24, Erdős–Rényi, K ramp 0→ 1.5 over 120 time units). η fixed uses M = L(t=0); η lag uses M = L(t− 3). p rewire η fixed mean ±σ η lag (τ =3) mean ±σ σ lag /σ fixed 0.000+40.0± 0.0+40.0± 0.01.00× 0.005+37.4± 2.2+35.6± 2.81.32× 0.020+33.5± 4.2+28.9± 7.61.82× 0.050+28.5± 5.7+20.8± 3.90.68× Three findings follow. First, η fixed delivers a longer precursor lead than η lag at every rewiring rate, with the advantage growing monotonically from p = 0.005 to p = 0.02 (lead advantage 1.8 and 4.6 time units respectively). Second, η fixed has lower seed-to-seed variance than η lag at p ≤ 0.02, with the variance ratio σ lag /σ fixed reaching 1.82× at p = 0.02. Third, at the highest rewiring rate (p = 0.05), η lag achieves lower variance than η fixed (0.68×) but at a cost of 7.7 fewer time units of lead — not a favorable tradeoff for a precursor diagnostic whose primary figure of merit is early detection. The η inst convention (floating M) performs similarly to or worse than η lag at all rewiring rates. Even when P does not perfectly track M inst — as it cannot under finite coupling rates — the instantaneous Laplacian changes too rapidly for the misalignment to accumulate. The structural drift that constitutes the precursor signal requires a fixed reference to build against; a moving reference perpetually resets the baseline. 21

Interpretation. The physical reason is transparent. η fixed measures cumulative participation drift from the structural reference L 0 — how far P has traveled, in operator space, from the eigenbasis it was aligned with at the start of the ramp. As the system approaches the synchronization transition, P reorganizes substantially and this cumulative drift is large and detectable early. η lag (τ ) measures drift from a more recent reference L(t−τ ), which is smaller by construction: the system has moved less far from where it was τ time units ago than from where it started. Reducing τ toward zero further collapses the signal. The fixed-reference construction is therefore not a convenience — it is the natural choice for detecting the onset of a transition, because the onset is defined by departure from the pre-transition structural state, not by instantaneous rate of change. This result directly parallels the head-to-head comparison against transfer entropy in §3.5: just as η fixed outperforms TE by combining an earlier peak with lower seed-to-seed variance, it outperforms its own time-varying generalizations for the same reason — it measures structural history, not instantaneous state. The fixed-M convention was introduced to avoid a tautology. It turns out to be optimal. Data and Code Availability Code, random seeds, and raw numerical data used to generate every figure and table in this paper are available in the accompanying Zenodo archive at https://doi.org/10.5281/zenodo.20564191 A Robustness of the TE-peak location to estimator hyperparame- ters The pairwise transfer entropy used in the head-to-head comparison (§3.5) depends on two estimator hyperparameters: the number of phase bins n bins and the prediction lag τ. To verify that the comparison against η is not driven by a particular choice, we recompute TE on the same simulation data using four configurations. Trajectories, ensemble seeds, and the per-(seed,K) random pair samples are identical across configurations; only n bins and τ change. Table 5: Robustness of the transfer-entropy peak location to estimator hyperparameters. All values are means ± standard deviation across the same n seed = 8 Kuramoto realizations as Table 2. The peak location ⟨K TE ⟩ and its seed-to-seed spread σ(K TE ) are nearly identical across all four configurations, and the strict-inequality count K TE

K η is 7/8 in every case. The σ-ratio relative to η remains close to 5× throughout (σ(K η ) = 0.060, configuration-independent). Configuration⟨K TE ⟩ σ(K TE ) σ/μ K TE K η n bins = 4, τ = 1 (baseline)0.5390.3130.587/8 n bins = 3, τ = 10.5280.3150.607/8 n bins = 5, τ = 10.5390.3130.587/8 n bins = 4, τ = 20.5500.3190.587/8 Out of the 32 (seed, configuration) entries, only two changed under hyperparameter variation: seed 2 dropped from K TE = 0.517 to 0.431 with n bins = 3, and seed 5 rose from 0.690 to 0.776 with τ = 2. 22

0.00.51.01.52.02.5 K 0.00 0.01 0.02 0.03 0.04 TE(K) (relative units) (a) TE curves under estimator choices 4 bins, =1 3 bins, =1 5 bins, =1 4 bins, =2 K= 0.23 4 bins, =1 3 bins, =1 5 bins, =1 4 bins, =2 0.2 0.4 0.6 0.8 1.0 K TE (b) TE peak locations by configuration K± Figure 7: Transfer-entropy estimator robustness. (a) Ensemble-mean TE(K) for the four configu- rations of (n bins ,τ ). The curves differ in absolute magnitude (more bins yield larger nominal TE values; longer lag broadens the temporal window) but share peak location and shape. The dotted vertical line marks⟨K η ⟩ = 0.23. (b) Per-seed K TE values for each configuration (points jittered hor- izontally; horizontal bars indicate per-configuration means). The shaded red band shows ⟨K η ⟩± σ from Fig. 4. The K TE distribution is essentially configuration-invariant, and in every configuration most realizations sit well above the η band. The σ-ratio finding of Table 2 is preserved across all configurations: σ(K TE )/σ(K η ) ranges from 5.22× to 5.32×. B Floquet anchor: full details B.1 Setup We consider the periodically kicked transverse-field Ising chain on N = 4 sites with periodic bound- ary conditions. One Floquet period applies the Ising interaction followed by a transverse-field kick: U F (h) = e −ihτ x H x · e −iτ z H z ,(11) where H z = −J P i σ z i σ z i+1 is the Ising rigidity, H x

P i σ x i is the kick generator, J = 1, and τ x = τ z = 1. We sweep the drive strength h over [0, 2.5] with 80 samples and adopt the rigidity operator M = H z throughout. Note that H z is Hermitian but indefinite, so D eff should be read in the generalized sense described in §2. The “before” state is the thermal Gibbs density of the Ising rigidity at temperature T th = 1.5, P before = e −H z /T th /Z th . By construction [P before ,H z ] = 0, so η before = 0 exactly. The Floquet-dephased state is defined canonically by projectors onto full quasi-energy eigenspaces, P block after

X α Π α P before Π α , where Π α projects onto the full eigenspace of U F at distinct quasi-energy e iε α . In the numerical implementation we use a Schur decomposition to construct the dephased state; Appendix B.3a ver- ifies that the Schur-basis result is numerically indistinguishable from the block-projector definition for the present N = 4 benchmark. 23

The diagnostic computes χ = D eff (P after ,H z )/D eff (P before ,H z ) and η =∥[P after ,H z ]∥ F /(∥P after ∥ F ∥H z ∥ F ). B.2 Trajectory in the(χ,η) plane Figure 8 (left) shows (χ,η) as h is swept. At h = 0 the trajectory is at (1, 0): no drive, no deviation from the relaxed reference. As h grows the trajectory ascends into the upper-half plane and traces a loop through the sustained-coherence quadrant (χ < 1, η > 0), reaching η ≃ 0.21 near h ≃ 0.4 and oscillating with the resonant structure of the Floquet spectrum as h increases further. At the endpoint h = 2.5, (χ,η) = (0.526, 0.143). The drive-strength dependence (Figure 8 right) makes the resonant structure explicit. χ(h) and η(h) oscillate in approximate anti-phase: at h values where the system most strongly selects a sub- manifold (χ minimum), the operator misalignment is largest (η maximum); between resonances, (χ,η) relaxes toward the fixed-point quadrant. A linear-response-like power law η(h)∼ h 0.78 holds in the small-h window, fit over h∈ [0.05, 0.5] with ±0.05 sensitivity to endpoint choice, before the resonant features dominate. For every h > 0 sampled, the steady state sits with χ < 1 and η > 0 — the sustained-coherence quadrant. This separates the Floquet steady state from the strongly synchronized Kuramoto state (K ≫ K c ), which at large coupling has both χ low (dimension selected) and η small (aligned with the Laplacian) — the selection-relaxation quadrant. In (χ,η) language the two regimes are geometrically distinct, a separation that any scalar precursor diagnostic would collapse. Figure 8: Floquet anchor: (χ,η) trajectory for the periodically kicked N = 4 transverse-field Ising chain as drive strength h is swept from 0 to 2.5. (Left) Diagnostic plane. Fixed-M trajec- tory (circles, color-coded by h) ascends into the sustained-coherence quadrant (χ < 1, η > 0), reaching η ≃ 0.21 near h ≃ 0.4 and oscillating with resonances at higher h. At h = 2.5, (χ,η) = (0.526, 0.143). The floating-M trajectory (triangles) sits identically at η = 0, demon- strating that the floating-M convention is a tautology and motivating the fixed-M choice used throughout the paper. (Right) χ(h) (blue) and η(h) (red) under both conventions (solid: fixed- M; dashed: floating-M). The two diagnostics oscillate in approximate anti-phase under fixed-M, reflecting resonant features of the Floquet spectrum. 24

B.3 Convention dependence: why fixed-M A natural alternative to the fixed-M convention is the floating-M convention, in which M is reas- signed to the effective stroboscopic Hamiltonian H F = (i/T period ) logU F , where T period = τ x +τ z = 2 and log denotes the matrix logarithm, at each value of h. Figure 8 shows both: the floating-M points (triangles in the left panel; dashed red line in the right panel) sit at η ≡ 0 for all h. This is a tautology: P after is diagonal in the Floquet basis by construction, so [P after ,H F ] = 0 identically. The floating-M convention is therefore vacuous as a precursor diagnostic; we adopt the fixed-M convention throughout this paper, with M taken as the system’s intrinsic rigidity operator (graph Laplacian for Kuramoto, H z for Floquet, fixed reference operator for model DSI spectra). The Schur decomposition replaces numpy.linalg.eig for diagonalizing U F because the kicked TFIM hasZ 2 symmetry that produces degenerate quasi-energies. At these degeneracies, numpy.linalg.eig returns non-orthogonal eigenvector matrices within the degenerate subspace, propagating numerical error of order 10 −3 into the projected P. scipy.linalg.schur returns a unitary Schur basis and preserves unitarity to numerical precision. In degenerate quasi-energy sectors, rank-one Floquet dephasing depends on the choice of basis within the degenerate subspace and is therefore not uniquely determined by the dynamics alone. The block-projector definition resolves this ambiguity; see Appendix B.3a. In practice, we implement rank-one dephasing in a Schur basis as a numerically stable convention, and the qualitative regime distinction — sustained-coherence vs. selection-relaxation — is robust to small perturbations of the quasi-energy degeneracies. B.3a Basis invariance under the block-projector definition We define the Floquet-dephased state using projectors onto full quasi-energy eigenspaces: P block after

X α Π α P before Π α ,(12) where Π α is the orthogonal projector onto the full eigenspace of U F with quasi-energy e iε α , summing over distinct quasi-energies. For non-degenerate quasi-energies, Π α =|φ α ⟩⟨φ α |, and Eq. (12) reduces to ordinary rank-one Floquet dephasing. For degenerate blocks of dimension d α

1, Π α projects onto the entire degenerate subspace and is independent of the basis chosen inside that subspace. Invariance. Let U α be any unitary rotation acting only within the α-th degenerate subspace. Since U α preserves the subspace, U α Π α U † α = Π α . Thus the projector Π α , and therefore the map P 7→ P α Π α P Π α , is unchanged by arbitrary basis rotations within degenerate quasi-energy sectors. The diagnostics χ and η computed from P block after are therefore basis-invariant. Perturbation bound for the Schur-basis approximation. Let P Schur after denote the state pro- duced by rank-one dephasing in a particular Schur basis, and let P block after be the block-projector state in Eq. (12). Define ∆P = P Schur after − P block after . Using the commutator bound ∥[A,B]∥ F ≤ 2∥A∥ F ∥B∥ F and the global bound 0≤ η ≤ √ 2, we obtain    η

P Schur after ,H z
− η

P block after ,H z
   ≤ 2∥∆P∥ F ∥P Schur after ∥ F + √ 2   ∥P block after ∥ F −∥P Schur after ∥ F   ∥P Schur after ∥ F .(13) 25

When the two dephased states have nearly equal Frobenius norm, the second term is negligible and the leading contribution is proportional to ∥∆P∥ F . Numerical check. For the N = 4 kicked TFIM parameter values used in §4, the maximum value of∥∆P∥ F /∥P Schur after ∥ F across all 80 sampled drive strengths is 1.21× 10 −9 , with median 3.98× 10 −15 . The Frobenius-norm correction term in Eq. (13) is at most 4.40×10 −15 , numerically zero at machine precision. Consequently, max h |∆η|≤ 2.42× 10 −9 , which is negligible compared with the η values reported in Fig. 8. The block-projector and Schur- basis constructions are therefore numerically indistinguishable for the present Floquet benchmark, and the qualitative regime distinction (χ < 1,η > 0) is unaffected by the basis convention. B.4 Limitations and scope This anchor is a proof of principle for cross-domain applicability, not a systematic study of Floquet precursor diagnostics. Three limitations apply. First, N = 4 is small; no systematic finite-size check has been performed. Second, sensitivity to the reference temperature T th and kick periods τ x , τ z has not been explored. Third, no head-to-head comparison against Koopman-operator early-warning indicators [9] — the natural competitor for driven systems — has been performed, though the simulation data are in hand and this comparison is in preparation. What the anchor establishes is narrow: the (χ,η) construction applies without modification to driven Hamiltonian dynamics; the resulting trajectory is geometrically distinct from the Kuramoto selection-relaxation regime; and the floating-M convention is vacuous. C Per-N finite-size scaling data Table 6 reports per-N ensemble statistics underlying Fig. 2. Gap uncertainties are SEM. Per-seed ⟨K η ⟩ and ⟨K c ⟩ values are shown graphically in Fig. 2a. One seed at N = 12 was excluded by the logistic-fit clipping criterion (§2.4). Table 6: Per-N summary for the finite-size scaling sweep on Erdős–Rényi networks at fixed mean degree d = 4. N Seeds Valid Gap mean (SEM) Gap > 0 Clipped 1220190.376± 0.08019/191 2412120.793± 0.17112/120 48880.861± 0.1618/80 96660.541± 0.1036/60 192440.656± 0.0904/40 384330.607± 0.0613/30 Total 5352 Positive gap in 52/52 valid realizations 26

References [1] Yoshiki Kuramoto. Self-entrainment of a population of coupled non-linear oscillators. In In- ternational Symposium on Mathematical Problems in Theoretical Physics, volume 39 of Lecture Notes in Physics, pages 420–422. Springer, Berlin, 1975. [2] Steven H. Strogatz. From Kuramoto to Crawford: exploring the onset of synchronization in populations of coupled oscillators. Physica D, 143:1–20, 2000. [3] Marten Scheffer, Jordi Bascompte, William A. Brock, Victor Brovkin, Stephen R. Carpenter, Vasilis Dakos, Hermann Held, Egbert H. van Nes, Max Rietkerk, and George Sugihara. Early- warning signals for critical transitions. Nature, 461:53–59, 2009. [4] Didier Sornette. Discrete scale invariance and complex dimensions. Physics Reports, 297:239– 270, 1998. [5] Egbert H. van Nes and Marten Scheffer. Slow recovery from perturbations as a generic indicator of a nearby catastrophic shift. The American Naturalist, 169:738–747, 2007. [6] Vasilis Dakos, Marten Scheffer, Egbert H. van Nes, Victor Brovkin, Vladimir Petoukhov, and Hermann Held. Slowing down as an early warning signal for abrupt climate change. Proceedings of the National Academy of Sciences, 105:14308–14312, 2008. [7] Daniele Marinazzo, Ludovico Angelini, Mario Pellicoro, and Sebastiano Stramaglia. Synergy as a warning sign of transitions: the case of the two-dimensional Ising model. Physical Review E, 99:040101, 2019. [8] Thomas Schreiber. Measuring information transfer. Physical Review Letters, 85:461–464, 2000. [9] Yuta Miyauchi, Masahiro Ikeda, and Yoshinobu Kawahara. Generalized stochastic resilience for early warning signals based on Koopman operator. Nonlinear Dynamics, 114(4):246, 2026. [10] Albrecht Böttcher and David Wenzel. The Frobenius norm and the commutator. Linear Algebra and its Applications, 429(8–9):1864–1885, 2008. [11] Göran Lindblad. On the generators of quantum dynamical semigroups. Communications in Mathematical Physics, 48:119–130, 1976. [12] Vittorio Gorini, Andrzej Kossakowski, and E. C. G. Sudarshan. Completely positive dynamical semigroups of N-level systems. Journal of Mathematical Physics, 17(5):821–825, 1976. [13] Ricardo V. Ceguerra, Joseph T. Lizier, and Albert Y. Zomaya. Information storage and transfer in the synchronization process in locally-connected networks. In 2011 IEEE Symposium on Artificial Life (ALIFE), pages 54–61, Paris, France, 2011. IEEE. [14] Ali Seif and Mina Zarei. Synchronization, collective oscillations, and information flow in duplex networks. arXiv preprint, 2026. arXiv:2603.00313 [nlin.AO]. [15] V. Efimov. Energy levels arising from resonant two-body forces in a three-body system. Physics Letters B, 33:563–564, 1970. [16] P. Naidon and S. Endo. Efimov physics: a review. Rep. Prog. Phys., 80(5):056001, 2017. 27

[17] Hazime Mori. Transport, collective motion, and Brownian motion. Progress of Theoretical Physics, 33:423–455, 1965. [18] Robert Zwanzig. Nonlinear generalized Langevin equations. Journal of Statistical Physics, 9:215–220, 1973. [19] Michael te Vrugt and Raphael Wittkowski. Mori–Zwanzig projection operator formalism for far-from-equilibrium systems with time-dependent Hamiltonians. Physical Review E, 99:062118, 2019. [20] Franz Wegner. Inverse participation ratio in 2 +ε dimensions. Zeitschrift für Physik B, 36:209– 214, 1980. [21] Ferdinand Evers and Alexander D. Mirlin. Anderson transitions. Reviews of Modern Physics, 80:1355–1417, 2008. [22] Patrick N. McGraw and Michael Menzinger. Laplacian spectra as a diagnostic tool for network structure and dynamics. Physical Review E, 77:031102, 2008. [23] Yao Yao, Guo-Hui Dong, Xiao Xiao, and Chang-Pu Sun. Frobenius-norm-based measures of quantum coherence and asymmetry. Scientific Reports, 6:32010, 2016. 28