Many engineered systems, including rotating machinery, combustors and multi-axis drives, have stiffness or damping that is randomly perturbed by their own environment. When that perturbation scales with the state ("multiplicative noise"), a system can look calm on average while its energy, the second moment, grows without bound through rare violent bursts. The noise level at which this happens is the moment stability threshold. In rotating and cyclic machinery the perturbation is rarely stationary. It is cyclostationary: its statistics, such as its variance and correlations, repeat with the machine's cycle. Current threshold calculations assume either a single scalar mode or stationary noise, so they cannot say whether a coupled multi-mode system near a shaft or blade-pass frequency is safe.
This note proposes a specific answer. For lightly damped multi-input multi-output (MIMO) systems, the second-moment threshold under cyclostationary multiplicative noise is set by a small, predictable set of numbers: the noise's cyclic spectrum, a frequency-domain description of how its correlations oscillate with the cycle, evaluated at sums and differences of the system's modal frequencies. Averaging the noise into a stationary equivalent discards exactly these numbers. Analyzing each mode separately discards the cross terms between modes. We will derive this reduced criterion with stochastic averaging, check it against exact thresholds, and show that a monitor needs to estimate only these few cyclic-spectrum entries from data, not a full lifted multichannel spectrum.
Core Claim
Consider an $n$-mode structure $\ddot q + 2Z\Omega\dot q + \Omega^2 q = \varepsilon\sum_k \xi_k(t)\,G_k\, q$. Here $\Omega=\mathrm{diag}(\omega_i)$ holds the modal frequencies, $Z$ holds small damping ratios, $G_k$ are coupling matrices through which noise channel $k$ perturbs the stiffness, and the noise vector $\xi(t)$ is cyclostationary with cycle frequency $\alpha_0$. The noise has a cyclic cross-spectrum $S^{\alpha}_{kl}(f)$, which is nonzero only at cycle frequencies $\alpha \in \{0, \pm\alpha_0, \pm 2\alpha_0,\dots\}$. In the standard stationary case, stochastic averaging gives a modal energy drift that depends only on $S^{0}$ evaluated at $0$ and $2\omega_i$. Our claim is that the cyclostationary case adds terms of the form $S^{\alpha}_{kl}\big((\omega_i-\omega_j)/2\big)$ whenever a cycle frequency nearly matches a modal combination, $\alpha \approx \omega_i+\omega_j$. These terms survive the averaging because the cyclic modulation cancels the fast rotation of the mode-pair product. The averaged second-moment equation then becomes an autonomous linear system in a frame rotating with the detuning $\delta=\alpha-\omega_i-\omega_j$, with threshold
$$ \varepsilon_{\mathrm{crit}}^{2} \;=\; \min\Big\{\varepsilon^2 : \lambda_{\max}\big(-D_\zeta + \varepsilon^{2}\, \Phi\big[\,S^{0}, \{S^{\alpha}\}_{\alpha\approx\omega_i\pm\omega_j};\ \delta\,\big]\big) = 0\Big\}. $$
In this expression $D_\zeta$ is the diagonal damping contribution to the energy equations, $\Phi$ is a matrix of size $O(n^2)$ assembled from those few spectral values and the $G_k$, and $\lambda_{\max}$ is the largest real part of its eigenvalues. Two consequences can be tested. First, at fixed total noise power, putting power into the cyclic component at a modal combination frequency can lower the threshold a great deal, and the stationarized model will not see this. Second, a cyclic cross-spectrum between two channels at $\omega_i+\omega_j$ can destabilize a pair of modes that each look stable in isolation, which a per-mode analysis will miss.
Relation to Prior Work
The scalar results motivate the proposal and supply the ground-truth technique. Ornstein-Uhlenbeck Process with Fluctuating Damping derives exact stability boundaries for a scalar system whose damping fluctuates under white, telegraph (two-state switching) and fractional noise, and shows that pure multiplicative noise destabilizes. We take its telegraph-noise idea and make the switching process periodically modulated and multichannel, because that is the case where exact moment equations still close. The two thermoacoustic papers, Rate-dependent tipping and early warning in a thermoacoustic system under extreme operating environment and Rate-dependent bifurcation dodging in a thermoacoustic system driven by colored noise, apply stochastic averaging to a single-mode oscillator under stationary colored noise and obtain an amplitude Fokker-Planck equation together with scaling laws for peak amplitude. Their averaging step is our baseline method. We change both of its assumptions, moving to several modes and to periodic noise statistics.
The closest existing work is multivariable moment analysis. On Second-Moment Stability of Discrete-Time Linear Systems With General Stochastic Dynamics gives necessary and sufficient Lyapunov conditions for $x_{k+1}=A(\xi_k)x_k$ with general stochastic $A$. A periodic Markov $A$ falls within its scope, but the result comes as parameter-dependent matrix inequalities that are exact and opaque, with no statement of which noise features matter. Moment stability of stochastic processes with applications to control systems bounds moment orders under state-dependent jumps and treats the noise as memoryless or Markovian. Our contribution is the opposite trade-off: an approximate but interpretable criterion that names the cyclic spectral entries that matter. We use exact discrete-time or lifted computations of this kind as the reference against which the criterion is judged.
The cyclostationary estimation work supplies the measurement machinery. Estimating Linear Dynamical Networks of Cyclostationary Processes and Exact Topology Learning in a Network of Cyclostationary Processes lift a period-$T$ process into a $T$-times larger stationary vector process. They then invert its spectral matrix to recover network structure, at a sample-complexity cost that grows with $pT$ for $p$ channels. These methods are passive and say nothing about stability. We argue that a threshold needs far less, only $O(n^2)$ cyclic entries at known frequencies. The bearing pipeline in An Optimized Maximum Second-Order Cyclostationary Blind Deconvolution and Bidirectional Long Short-Term Memory Network Model for Rolling Bearing Fault Diagnosis shows that real vibration records carry strong cyclostationary structure even at low signal-to-noise ratio. We use those records as realistic multiplicative drivers, not as a classification task.
Research Program
The first part is theory. We will extend the Khasminskii-style averaging used in the thermoacoustic work to an $n$-mode linear system. The noise will be represented as a sum of stationary components modulated by $e^{i\alpha t}$, and the assumption of a stationary ergodic fast process will be replaced by stationarity in the lifted sense. We will state conditions on the separation between correlation time, cycle period and modal frequencies under which the averaged moment equation holds to order $\varepsilon^2$. We will also characterize how the threshold depends on the detuning $\delta$, expecting a resonance tongue whose width is set by damping and noise bandwidth. Non-normal coupling, where the $G_k$ are not symmetric, is included explicitly, because the dossier identifies it as the uncharacterized case for vector moment boundaries.
The second part is exact reference computation. When each $\xi_k$ is a finite-state Markov switching process whose rates or amplitudes vary periodically, the mode-conditioned second moments $\mathbb{E}[x x^\top \mathbf{1}\{\text{mode}=m\}]$ obey a linear periodic differential equation. The largest Floquet multiplier of that equation, which measures growth over one cycle, then gives the exact threshold with no Monte Carlo error. This yields a ground truth with no sampling noise for any $n$ up to a few dozen states.
The third part is estimation. From an observed noise record we estimate only $S^{\alpha}_{kl}$ at the candidate combination frequencies, using standard cyclic periodogram estimators, and plug the estimates into $\Phi$. We will derive how the threshold-estimation error scales with record length. For a fixed set of spectral entries this should be the usual $1/\sqrt{N}$ rate of spectral estimators. The full lifted inverse spectrum, by contrast, should degrade as $pT$ grows.
The fourth part is a minimal feedback use. We will ask how much extra modal damping, added through velocity feedback, is needed to restore second-moment stability. Because the criterion is explicit, this becomes a small eigenvalue-constrained problem. We will compare its damping budget with the budgets the baselines prescribe.
Baselines
The naive reference point is the deterministic answer: the system is called stable whenever the noise-free plant is stable, which here is always. Every method has to beat this by predicting a finite threshold at all.
The strongest competitor on accuracy is stationarized stochastic averaging. This is the procedure of the thermoacoustic papers, extended to $n$ modes. The noise is replaced by its time-averaged stationary spectrum $S^0$, and the standard averaged moment equation is solved, including cross-mode terms from $S^0$. It is the strongest competitor because it is exactly what a careful practitioner would do today, and it is exact when every cyclic component at modal combinations vanishes. It receives the same noise records, the same spectral estimator and the same record lengths, and differs only in setting $S^{\alpha\neq 0}=0$. No reported result for multimode threshold prediction was found on Lacuna for this method; the cited papers report only scalar power-law scalings. We beat it if our relative error in $\varepsilon_{\mathrm{crit}}$ against the exact Floquet threshold is several times smaller in resonant configurations, and no worse in non-resonant ones.
The second baseline is decentralized per-mode analysis. Each mode receives its own scalar threshold, including its own cyclic term at $2\omega_i$, and the system threshold is the minimum over modes. This baseline isolates the MIMO ingredient. The material suggests decentralized single-channel treatments fail under cross-axis coupling (Multi-Robot Transfer Learning: A Dynamical System Perspective), but no threshold numbers are reported there.
The third baseline is the exact lifted criterion: the discrete-time second-moment test of On Second-Moment Stability of Discrete-Time Linear Systems With General Stochastic Dynamics, applied to the sampled, period-lifted system. When the true noise model is known it is the gold standard, so we do not try to beat it on accuracy. We compare against it on what it needs: a full generative noise model, and its computational cost as $n$ and $T$ grow.
The key ablation removes, one at a time, the cyclic terms at sum frequencies, the cyclic terms at difference frequencies, and the cross-channel cyclic terms. This shows which ingredient carries the accuracy gain.
Why not just…?
Why not just stationarize the noise and use ordinary stochastic averaging? This is the real threat. Whenever the cycle frequency sits far from every $\omega_i\pm\omega_j$, our theory predicts the cyclic terms average out, and the two methods should agree. The proposal stands or falls on resonant and near-resonant configurations. The detuning sweep in the experiments below maps the region where stationarization fails. If that region turns out to be narrower than the noise bandwidth makes plausible for real machines, we will report that the simpler method suffices, and the contribution shrinks to a diagnostic for when to worry.
Why not just lift the system to period $T$ and run the exact discrete-time moment test? If the noise's generative model is known, one should. Our experiments use this test as ground truth. In practice, though, one has a vibration record, not a Markov model. Fitting a full lifted model costs samples that grow with $pT$, which is the bottleneck noted for Estimating Linear Dynamical Networks of Cyclostationary Processes. The data-efficiency experiment compares threshold error against record length for the plug-in criterion and for a fitted lifted model, and that comparison settles the question.
Why not just run Monte Carlo? Near the threshold the second moment is dominated by rare bursts, and this intermittency makes sample-average energy estimates heavy-tailed and slow to converge. The thermoacoustic papers adopted averaging precisely to avoid brute-force simulation. We will still report Monte Carlo threshold estimates at matched compute so that this claim is measured here, not merely assumed.
Why not just design an $H_\infty$ or LQG controller and ignore moments? Both assume additive disturbances. Under multiplicative noise a system can be mean-stable while its variance diverges (Ornstein-Uhlenbeck Process with Fluctuating Damping), so neither design certifies the property we care about. We will check the damping that LQG-tuned loops deliver against the exact threshold.
Proposed Experiments
The first experiment uses synthetic resonant pairs. A two-mode system with non-normal coupling is driven by periodically modulated telegraph noise, and the cycle frequency is swept across $\omega_1+\omega_2$ while total noise power stays fixed. All methods are compared with the exact Floquet threshold. The decisive check is whether a dip in the true threshold appears near zero detuning and whether only the cyclic-aware criterion tracks it. We will also test whether a cross-channel cyclic spectrum destabilizes two modes that are individually stable. If it does, this refutes the per-mode baseline, and the ablation removing the cross-channel terms should break our criterion in the same way.
The second experiment scales up to a benchmark plant. Multiplicative cyclostationary noise is injected into the 8-state REA4 chemical reactor model from COMPleib, as used in Wasserstein distributionally robust regret-optimal control over infinite-horizon. For context, that paper reports regret 14.5% below $H_2$ and 89.7% below $H_\infty$ at radius $r=0.639$, solved in 5.8 seconds. Those figures concern additive-disturbance regret; no reported moment-threshold result was found on Lacuna for REA4. We report threshold error and wall-clock time for each method as $n$ grows through randomly coupled 4-, 8- and 16-mode systems.
The third experiment uses real cyclostationary drivers. Drive-end accelerometer records from the CWRU bearing dataset (1797 rpm, zero load, 0.007 to 0.021 inch faults) are normalized and used as $\xi_k(t)$ for a synthetic flexible-rotor model whose modal frequencies are placed near combinations of the fault characteristic frequencies. Ground truth here comes from long Monte Carlo runs with variance-reduced estimators. We measure threshold error against record length for the plug-in cyclic criterion, the stationarized baseline and a fitted lifted model. The cited diagnostic pipeline achieves 98.20% accuracy at −9 dB, which is evidence that these records carry exploitable cyclic structure; it is not a threshold result.
The proposal's central claim is falsified if, across the detuning sweeps, the stationarized criterion is within the Monte Carlo error bars of the exact threshold everywhere. It is also falsified if the plug-in criterion needs as many samples as the fully lifted model to reach the same threshold accuracy.
References
- Ornstein–Uhlenbeck process with fluctuating damping
- Rate-dependent tipping and early warning in a thermoacoustic system under extreme operating environment
- Rate-dependent bifurcation dodging in a thermoacoustic system driven by colored noise
- On Second-Moment Stability of Discrete-Time Linear Systems With General Stochastic Dynamics
- Moment stability of stochastic processes with applications to control systems
- Estimating Linear Dynamical Networks of Cyclostationary Processes
- Exact Topology Learning in a Network of Cyclostationary Processes
- An Optimized Maximum Second-Order Cyclostationary Blind Deconvolution and Bidirectional Long Short-Term Memory Network Model for Rolling Bearing Fault Diagnosis
- Multi-Robot Transfer Learning: A Dynamical System Perspective
- Wasserstein distributionally robust regret-optimal control over infinite-horizon