Contribution
The paper extends the generalized stability theory of Part I (LIT-931) to operators that change in time. Its main result is that time dependence and non-normality together can produce asymptotic instability in a system that is stable at each instant, and that this is generic under random variation. Non-normality is necessary for it. The paper presents the first result through simple examples and a bound, and the genericity result through cited theorems on stochastic systems and numerical calculation.
Key insight
In Part I, non-normal growth is transient because the operator stays the same: the growing combination of modes rotates into a decaying one and the slowest mode wins. If the operator changes before that happens, the new operator can take the grown perturbation and grow it again. “Transient growth instigated at the starting time is continued with further growth when the switch to the second operator takes place, forestalling the decay that would occur in the autonomous case” (§4). Repeated, that becomes exponential growth, though no snapshot of the system is unstable.
Assumptions
- Linear, time-dependent dynamics. du/dt = A(t)u (eq. 1), with propagator Φ(t, t₀) given by the time-ordered exponential (eqs. 3–5). In applications this is the tangent linear system about a trajectory.
- Lyapunov exponents exist. Guaranteed “for linear differential systems with continuous and bounded coefficients” (footnote 1). For an operator that is a time-independent mean plus a stochastic part, the multiplicative ergodic theorem gives exponents “almost surely in probability”, independent of the realization (§2).
- For the stochastic results: A(t) = Ā + εΣ_i ξ_i(t)B_i (eq. 27), with ξ_i independent, identically distributed, delta-correlated white noise (eq. 28), interpreted in the Stratonovich sense (eqs. 30–32).
- A norm, as in Part I. Non-normality is relative to it.
Key results
- No modes, but optimal perturbations and Lyapunov vectors (§2). The temporal eigenmodes of A(t) “are not defined”. The first Lyapunov exponent λ = lim sup ln‖Φ(t)‖/t (eq. 6) takes the place of the spectral abscissa, and reduces to it for a constant operator. For large t the SVD of the propagator has rank one to exponential accuracy (eq. 7). The optimal excitation of the first Lyapunov vector then depends only on t₀, and the first Lyapunov vector only on t. The argument is by the semigroup property and induction on orthogonal subspaces.
- Lorenz system (§3). At b = 8/3, σ = 10, r = 28 the exponents are
about 0.88, 0 and −14.55. Mean instantaneous growth rates along the
trajectory:
- numerical abscissa, about 4.2;
- optimal excitation of the Lyapunov vector, about 2.7;
- Lyapunov vector, about 0.88;
- the largest real part of the frozen-time eigenvalues, about 1.6. The numerical abscissa itself “exhibits frequent excursions of negative growth rate”. Convergence to the exponents is slow; convergence to the Lyapunov vector and its optimal excitation is fast, within about ten time units. The optimal perturbations lie up to 18° off the attractor plane.
- A bound on the Lyapunov exponent (eqs. 10–12). With r(t) = ‖u‖², dr/dt = u†(A + A†)u, so lim sup (1/t)∫λ_min(s)ds ≤ λ ≤ lim sup (1/t)∫λ_max(s)ds, where λ_min, λ_max are the extreme eigenvalues of (A(t) + A†(t))/2. “It is immediate from (12) that nonnormality of the evolution operator is a necessary condition for parametric asymptotic instability of such systems with stable eigenvalues since if the operator were at each instant asymptotically stable and normal it would necessarily also have negative Lyapunov exponents.”
- Switched oscillator (eqs. 13–17, Fig. 6). d/dt(x, v) = [[0, 1], [−ω²(t), −2γ]](x, v). The commutator (eq. 14) vanishes only for γ = 0 and ω ≡ 1, and no constant rescaling makes the operator normal when ω varies. Switching ω between ω₁ = 0.5 and ω₂ = 3 every T units, with γ = 0, gives a positive Lyapunov exponent except near integer multiples of half a natural period. “As T → 0, the Lyapunov exponent vanishes”: a finite switching period is needed. The paper states, and does not show, that the destabilization persists for γ > 0, when the operator is stable at each instant.
- Random neutral matrices (eqs. 18–26, Fig. 7). A_j = R_φj [[i, −2i cot θ], [0, −i]] R_φj†: eigenvalues ±i, trace zero, neutral at every instant, eigenvectors at angle θ. Each step T_j = e^{A_jδt} has det T_j = 1 and singular values (σ, 1/σ), σ ≈ 1 + 2 cot θ δt for small δt. A unit vector grows in one step when |cos φ_j| > 1/√(1 + σ²) (eq. 24), so with φ_j uniform it grows more often than it shrinks. The Lyapunov exponent is positive for every θ ≠ π/2 and increases with the non-normality. As δt → 0 it vanishes (eq. 26), because the trace is zero.
- Multiplicative white noise (§4, eqs. 27–35, Fig. 8). With the noise scaled by δt^{−1/2}, the destabilization survives the continuous limit. A system of the form (27) “has necessarily a positive Lyapunov exponent for noise of sufficient magnitude ε, provided the dimensionality of the system is greater or equal to 2 and the noise matrices, B_i, are neither skew-symmetric nor commute with the deterministic operator” (cited to Has’minskii 1980 and Colonius & Kliemann 1993). “Increasing the degree of nonnormality of the deterministic background state operator Ā reduces the noise amplitude, ε, required.” For the damped oscillator with Ā = [[0, 1], [−1, −2γ]] and B = [[0, 0], [−ε, 0]], the exponent falls with damping γ and rises with ε (Fig. 8). Stability is sample stability (an almost surely negative λ), not mean-square stability, since in stochastic systems μ₂ ≥ 2λ only (eq. 29).
- Three-layer baroclinic model (§5, eqs. 36–37, Fig. 9). Layer winds U_i = 10 + (3 − i)ΔU m s⁻¹ with independent white noise on each, damping r = 0.2 (5 days), global wavenumber 11. The noise needed for a positive exponent falls as the shear ΔU (the non-normality) rises. At ΔU = 10 m s⁻¹ the model “suggests existence of a positive Lyapunov exponent for rms temporal fluctuations of the order of 10%”, and about 1/5 day⁻¹ at 30%. A two-layer model shows no noise-induced instability, because its noise matrix commutes with the mean operator.
- Commuting noise cannot destabilize (§5). “When Ā and B_i commute, multiplicative noise cannot alter the stability properties of the mean operator”, because one change of basis diagonalizes both “and such a normal system cannot be destabilized by noise”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Modal analysis has no counterpart for time-dependent operators; optimal perturbations and Lyapunov vectors take its place | strong | §2, with Lyapunov 1907 and Oseledec 1968 cited |
| C2 | Non-normality is necessary for a system stable at each instant to be asymptotically unstable | strong | eq. 12, a one-line bound |
| C3 | A non-normal operator stable or neutral at each instant can be asymptotically unstable under time variation | strong | switched oscillator, Fig. 6; random neutral matrices, Fig. 7 |
| C4 | Multiplicative white noise of sufficient size destabilizes generically, unless the noise matrices commute with Ā or are skew-symmetric | moderate here | cited theorems (Has’minskii; Colonius & Kliemann), not proved in the paper |
| C5 | The variation needed falls as non-normality rises and rises with damping | moderate | numerical: Figs. 8 and 9 |
| C6 | Forecast error growth in the atmosphere at long range arises from the time dependence of the flow, not from intrinsic chaos | weak | a three-layer model and an argument (§§1, 5), stated as a suggestion |
Concepts
- Nonautonomous operator. A(t), time-dependent; the tangent linear operator along a trajectory is the standard case.
- Lyapunov exponent. λ = lim sup ln‖Φ(t)‖/t, the asymptotic growth rate; the role of the spectral abscissa.
- Lyapunov vector; its optimal excitation. The structure that asymptotically grows fastest, and the initial perturbation that excites it most.
- Parametric instability. Asymptotic instability produced by variation of a system’s parameters, though it is stable at each instant; the Mathieu equation is the classical case.
- Sample stability; mean-square stability. An almost surely negative Lyapunov exponent; a negative growth rate of the second moment. They differ for stochastic systems.
Connections
The sequel to Part I (LIT-931), reading the transient growth of that paper as the engine of instability under time variation. The stochastic results rest on Has’minskii (1980), Arnold & Kliemann (1983), Colonius & Kliemann (1993) and Furstenberg’s theorem on products of random matrices (1963), and the random-matrix example adapts Zel’dovich et al. (1984).
Bearing on the record
- CLAIM-186, “an argument under development is the nonautonomous case”. The paper supports the first half: for a time-dependent operator there are no modes, “even in principle”. It adds something CLAIM-186 does not say and should weigh. Part I’s comfort, that a non-normal system stable in every mode settles in the end after its burst, does not carry over. An operator that changes as it is worked on can grow disturbances without limit while being stable at every snapshot. The note in CLAIM-186 reads GST as qualifying settling with bursts. On Part II, time variation of a non-normal structure can remove settling altogether.
- The conditions are specific, and they cut both ways for CLAIM-186.
- Non-normality is necessary (C2). If a record’s dependency structure were normal in the relevant norm, changing it could not destabilize.
- Damping stabilizes, and the variation needed grows with damping (C5). In the reading across, anchoring is a candidate for damping.
- In the deterministic examples, a switching period short compared with the transient-growth time destabilizes nothing (T → 0, δt → 0). LIT-929’s Standing asks what carries over “to an operator that changes as fast as the state does”. Here, at least, change faster than growth is harmless, and change on the growth timescale is what destabilizes. Under white noise scaled as δt^{−1/2} the fast limit stays unstable, so the answer depends on how the change is modelled.
- What the paper does not support. It is about linear perturbation dynamics with a norm. Nothing in it is about arguments, and the bearing remains the assistant’s reading, as CLAIM-186 says.
- A new THEORY is filed, THEORY-325: time dependence can destabilize a non-normal system stable at every instant; non-normality is necessary; random variation destabilizes generically, needing less the more non-normal and less damped the mean operator. No existing THEORY held it.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- The genericity of noise-induced instability (C4) is cited, not proved, and the paper’s own evidence for it is numerical, in low-dimensional models.
- The claim that destabilization of the oscillator persists for γ > 0 is stated (“It can be shown”) without the calculation.
- The atmospheric conclusions (C6) rest on a three-layer quasigeostrophic model with white-noise winds. The authors call it “suggests”.
- Stratonovich versus Itô matters for multiplicative noise (eqs. 30–31); results for Itô noise could differ.
Open questions
- Coloured noise and slow variation. The paper’s two deterministic examples need a finite switching period; whether realistic, correlated variation sits nearer the switched or the white-noise case is not settled here.
- Whether the noise threshold can be estimated from properties of Ā alone, such as its numerical abscissa or pseudospectra, rather than by computing exponents.