Contribution
The paper collects the non-modal results of the preceding decade into one stability theory for time-independent linear operators. Modal theory becomes its t → ∞ limit. The finite-time, short-time and forced behaviour, which eigenvalues do not give when the operator is non-normal, comes from the singular value decomposition of the propagator and of the resolvent. The paper says it organises “recent advances into a systematic stability theory” (§5). Most of the results are earlier, many of them the authors’ own, and it is a synthesis with worked examples rather than a source of new theorems.
Key insight
Eigenvalues tell you what wins in the end, not what happens on the way. When the eigenvectors of A are not orthogonal, a combination of decaying modes can grow for a long time before the slowest mode takes over, and the perturbation that grows most is not a mode at all. The right object is the propagator e^{At} over the interval of interest. Its singular vectors are the optimal perturbations and its norm is the growth.
Assumptions
- Linear, autonomous dynamics. du/dt = Au (eq. 1), A independent of time. “Only the existence of a propagator is required for the following development, and stationarity is assumed for convenience only” (§2).
- Finite dimension. A is a discretized matrix. Footnote 2 notes that convergence of a discretization to the continuous problem in time is not assured.
- A chosen inner product. Growth is measured in a norm ‖u‖ = (u†Mu)^{1/2}, and the change of variables v = M^{1/2}u, D = M^{1/2}AM^{−1/2} reduces any such norm to the Euclidean. Non-normality is a property of A in that norm: “it is always possible to find a (usually unphysical) norm that renders the operator normal” (§2). The results hold for the chosen norm, which “should be made on physical grounds”, usually energy.
- For the stochastic results: additive white noise, delta-correlated in time, with spatial structure F (eqs. 23–24), and A asymptotically stable for a stationary state.
Key results
- Normality decides whether eigenvalues suffice (§2). A is normal if AA† = A†A. Then ‖e^{At}‖ = e^{λ^R_max(A)t}, with λ^R_max(A) = max Re λ(A) the spectral abscissa, and “the stability of a normal dynamical system is determined for all time by its eigenspectrum”.
- Finite-time growth (eqs. 4–6). The growth of u(0) over time t satisfies σ² = (e^{A†t}e^{At}u(0), u(0))/(u(0), u(0)). The maximum is λ_max(e^{A†t}e^{At}) = ‖e^{At}‖². With e^{At} = UΣV† (eq. 5), the columns of V ordered by growth are the optimal perturbations, and the maximum over all t is the global optimal. For every eigenvalue λ_i, σ²_min ≤ e^{(λ_i+λ_i*)t} ≤ σ²_max (eq. 6): the optimal does at least as well as the fastest mode.
- The t → ∞ limit (eqs. 7–9). The leading mode dominates, as modal theory says. But the unit initial condition that excites it most is the conjugate of its biorthogonal (the adjoint eigenvector), not the mode. In two dimensions, with eigenvectors at angle θ, the gain is csc θ (Fig. 1).
- The t → 0 limit (eq. 10). e^{A†t}e^{At} ≈ I + (A + A†)t, so the maximum instantaneous growth rate is the numerical abscissa a(A), the largest eigenvalue of (A + A†)/2. It can be positive “even though all normal modes of A are damped”.
- Forced response (eqs. 14–19). With resolvent R(ω) = (iωI − A)^{−1}, a normal system has ‖R(ω)‖ = 1/dist(iω, Λ(A)) (eq. 16), resonance at the least-damped mode. For non-normal A the optimal forcing is the leading right singular vector of R(ω), and ‖R(ω)‖ ≤ κ(E)/dist(iω, Λ(A)) (eq. 19), with κ(E) = ‖E‖‖E^{−1}‖ the condition number of the eigenvector matrix. The maximum response need not be at the least-damped mode’s frequency and “can be orders of magnitude greater”.
- Variance under white noise (eqs. 20–22). Forced by spatially uncorrelated white noise, ⟨u²⟩ = (1/2π)∫F(ω)dω with F(ω) = trace(R†(ω)R(ω)) ≥ Σ_i 1/|iω − λ_i|², “with equality only when A is normal” (eq. 22, cited to Ioannou 1995). A non-normal system maintains more variance than the sum over its poles.
- Stochastic optimals and EOFs (eqs. 23–31). The variance is F†B_tF with B_t = ∫0^t e^{A†s}e^{As}ds (eqs. 26–27). At stationarity B∞ and the covariance C_∞ solve the Lyapunov equations A†B_∞ + B_∞A = −I and AC_∞ + C_∞A† = −FF† (eq. 31). The leading eigenvectors of B_∞ (stochastic optimals: the forcings that maintain most variance) and of C_∞ (EOFs: the patterns that hold it) coincide with the modes only if A is normal and FF† = I.
- The 2 × 2 example (§4a, eqs. 32–43). A = [[−1, −cot θ], [0, −2]]:
eigenvalues −1 and −2, eigenvectors at angle θ, normal only at θ = π/2.
- Numerical abscissa a(A) = (−3 + csc θ)/2 for 0 < θ < π/2, so some perturbation grows at first when θ < π/9.244.
- Bound ‖e^{At}‖ ≤ κ(E)e^{λ^R_max t} = cot(θ/2)e^{−t} (eq. 40), which this operator does not attain. The optimal growth peaks near 8, at t ≈ 0.7, for θ = π/100; at θ = π/10 there is none (Fig. 4).
- Maintained variance under unit white noise ⟨‖u‖²⟩ = 3/4 + cot²θ/12 (eq. 43). The 3/4 is the equivalent normal response.
- Damped Eady model (§4b). Zonal wavenumber k = 2.5 (on the stable side of the modal short-wave cutoff at 2.39), l = π/5, damping r = 0.1 (e-folding 2.9 days). Asymptotically stable, yet the global optimal grows in energy by about 100 over 4 days (Fig. 8). The numerical abscissa has no short-wave cutoff (Fig. 7). Stochastically forced, the first stochastic optimal accounts for 32% of the maintained energy and the first EOF for 69%, and the system carries a heat flux of about 0.5 PW per W m⁻² of forcing at 30°.
- Convective instability (§4c). An Eady channel with damping zones at both ends (x_m = 20, r = 100) has only decaying modes. Its two-day optimal grows in energy by a factor of 111.6 (Fig. 12).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | For non-normal A, eigenvalues govern only the t → ∞ limit; finite-time growth is ‖e^{At}‖ and can exceed every modal growth | strong | linear algebra, eqs. 4–6 and 10, with standard SVD facts |
| C2 | A system with every mode damped can amplify perturbations by large factors before decaying | strong | eq. 36 and Figs. 3–4; Eady example, Fig. 8; convective example, Fig. 12 |
| C3 | The optimal excitation of a mode is its adjoint (biorthogonal), not the mode | strong | eqs. 8–9, the 2 × 2 calculation (eq. 39) |
| C4 | Under white-noise forcing a non-normal system maintains more variance than the normal system with the same spectrum | strong for the inequality, cited | eq. 22, cited to Ioannou 1995, not proved here; eq. 43 in the example |
| C5 | Atmospheric phenomena (cyclogenesis, forecast error growth, maintained turbulent variance, heat flux) are better described by this theory than by modal theory | moderate | worked models and cited literature, not a comparison with data in this paper |
Method
Discretize, choose the physical norm and transform to Euclidean variables, then take singular value decompositions:
- of the propagator e^{At} for the initial-value problem;
- of the resolvent R(ω) for harmonic forcing;
- of B_∞ and C_∞, from the Lyapunov equations, for stochastic forcing.
Concepts
- Normal operator. One that commutes with its adjoint in the chosen inner product, so its eigenvectors are orthogonal.
- Spectral abscissa λ^R_max(A). The largest real part of an eigenvalue; the t → ∞ growth rate.
- Numerical abscissa a(A). The largest eigenvalue of (A + A†)/2; the largest possible growth rate at t = 0.
- Optimal perturbation. The leading right singular vector of e^{At} for a chosen t. The global optimal maximizes over t as well.
- Biorthogonal (adjoint) vector. The vector orthogonal to all eigenvectors but one, which is the optimal excitation of that one.
- Equivalent normal response. The response of an operator with the same eigenvalues and orthogonal eigenvectors.
- Stochastic optimal; EOF. The forcing structures that maintain the most variance; the structures that hold it.
- Generalized stability analysis. The SVD analysis of the propagator at all t, which “subsumes” both asymptotic limits (§2).
Connections
It builds on Kelvin (1887) and Orr (1907), on Farrell’s optimal perturbation papers from 1982 onward, and on the shear-flow transient growth literature it cites in parallel (Butler & Farrell 1992; Reddy & Henningson 1993; Trefethen et al. 1993, LIT-932). It names pseudospectra as “a systematic basis” for analysing non-normality and leaves them there. Part II (LIT-929) extends the theory to time-dependent operators.
Bearing on the record
- CLAIM-186, “settling does not rule out bursts”. The note’s first sentence is what this paper shows, at the strength it shows it: “it can be stable in every mode and still amplify a disturbance by a large factor for a while before the disturbance decays.” That is C2.
- CLAIM-186, “directed, asymmetric coupling is a standard source of non-normality”. The paper does not say this. Its source of non-normality is energy drawn from a background shear flow. Its 2 × 2 example is upper triangular, which is a one-way coupling, and the non-normality grows as that coupling (−cot θ) grows relative to the difference of the decay rates. So the example fits the reading, but the general statement is the assistant’s, as CLAIM-186 itself says.
- The norm matters. Non-normality is relative to an inner product, and “it is always possible to find a (usually unphysical) norm that renders the operator normal”. Any reading across to an argument must say what its norm is. A measure of drift is a choice of norm, and whether a record’s dependency structure is non-normal is a question about that norm, not about the graph alone.
- CLAIM-186’s advice to measure drift over the short run after a disturbance, not only as an average. The paper supports the physics behind it: the t → 0 and intermediate-t growth differ in kind from the t → ∞ rate. What it adds is the forced case. Under steady stochastic disturbance a non-normal system holds more variance than its spectrum predicts (eq. 22), so a non-normal argument under continual small disturbance would show more standing drift, not only bursts.
- A new THEORY is filed, THEORY-323: the eigenvalues of a non-normal operator govern only the long-run limit, and finite-time growth and forced variance can exceed what they predict by large factors. No existing THEORY held it.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Linear and autonomous. Nonlinear equilibration and time-dependent operators are outside it (the latter is Part II).
- The atmospheric examples are idealized models (the Eady model with uniform or localized damping), not comparisons with observation.
- The stochastic results assume white noise and stationarity. Eq. 22 is cited, not proved.
- Non-normality is norm-relative, and the paper’s claims about the atmosphere rest on the energy norm being the physical one.
- Fig. 5’s caption labels the normal reference “θ = 0”; the text makes the normal case θ = π/2. A slip in the caption.
Open questions
- How much of the transient growth survives nonlinearity in each application. The paper parameterizes nonlinearity as noise and damping (§5) and does not test it.
- Which norm is physical when energy is not defined. That question is the one any reading across to a non-physical system must answer first.