---
title: Colored Noise Coherency (CoNoCo)
url: https://www.emergentmind.com/topics/colored-noise-coherency-conoco
type: topic
---

# Colored Noise Coherency (CoNoCo)

Searching arXiv for the specified CoNoCo paper and closely related colored-noise coherency work.
arxiv_search(query="2512.15379 OR \"Remotely Detectable Robot Policy Watermarking\" OR \"Colored Noise Coherency\"", max_results=10)
arxiv_search(query="\"Effect of Colored Noise on Coupled Thermoacoustic Oscillators\" OR 2606.22665 OR \"Colored noise induces synchronization of limit cycle oscillators\"", max_results=10)
Colored Noise Coherency (CoNoCo) denotes, in current arXiv usage, both a specific watermarking strategy for stochastic robot control policies and a broader description of how temporal correlations in noise modulate coherence, regularity, synchronization, and detectability in dynamical systems. In robotics, CoNoCo is a keyed spectral watermark embedded into a policy’s stochastic exploration and recovered from remote observations through spectral coherency, with the explicit aim of surviving unknown sensing rates, unknown system dynamics, and observation noise [2512.15379]. In nonlinear dynamics and stochastic physics, the term is used more generally for colored-noise-induced organization of oscillations, phase relations, and spectral sharpness, including thermoacoustic oscillators, higher-order stochastic-resonance networks, and phase-reduced oscillators [2606.22665].

## 1. Definition and conceptual scope

Across the cited literature, CoNoCo is tied to the effect of non-white stochastic forcing on coherent system response. In the most general formulation, colored noise alters coherence because its power is not spectrally flat; instead, it selectively overlaps with phase-sensitive modes, least stable modes, or drive-relevant frequencies. For uncoupled limit-cycle oscillators, the effect is explicitly spectral: only noise components at the natural frequency and its harmonics contribute to the effective phase-difference diffusion, so synchronized and clustered states depend on the overlap between the colored-noise spectrum and the oscillator’s phase sensitivity harmonics [1402.1839]. In coupled thermoacoustic oscillators, CoNoCo is quantified by a coherence factor and is strongest when the noise memory matches the acoustic period and the system operates near a saddle-node threshold [2606.22665]. In higher-order bistable networks, CoNoCo-like behavior combines temporal matching and spatial synchronization, and classical colored noise suppresses stochastic resonance while shifting the optimal noise intensity [2603.06108]. In phase-response-curve estimation, colored noise improves finite-sample stimulus–response coherence by reshaping the stimulus spectrum while preserving total power [1409.1637].

This usage suggests that CoNoCo is not a single universal observable. Rather, it is a spectral-selective mechanism that appears through different diagnostics in different fields: complex coherency in remote watermark detection, coherence factor in thermoacoustics, spectral amplification and synchronization minima in stochastic resonance, and spike-triggered averages in phase reduction. The shared theme is that temporal correlation reorganizes the effective coupling between noise and system dynamics.

## 2. Remote robot policy watermarking

In robot learning, CoNoCo was introduced as “the first watermarking strategy designed for remote detection,” motivated by the fact that trained policies are valuable intellectual property while auditors often have access only to external observations such as video or motion capture [2512.15379]. The formal obstacle is the “Physical Observation Gap,” consisting of **C1: Synchronization Uncertainty**, where the policy runs at an internal rate \(f_\pi\) unknown to the auditor and remote sensors are asynchronous; **C2: System Dynamics**, where observers see the physical response rather than the actions; and **C3: Interference and Noise**, where nominal behavior, disturbances, and sensor noise obscure any embedded signal.

The policy model is the standard stochastic continuous-control form
\[
a_k=\mu_\theta(o_k)+\Sigma_\theta(o_k)\epsilon_k,
\]
with white Gaussian exploration noise \(\epsilon_k \sim \mathcal{N}(0,I)\). CoNoCo replaces that white noise with a normalized colored Gaussian noise sequence \(W_k\) whose spectral energy is concentrated in a secret band \(\mathcal{A}=[f_{\min},f_{\max}]\). The owner’s key is
\[
\mathcal{K}=\{S,\mathcal{A}\},
\]
where \(S\) is a PRNG seed and \(\mathcal{A}\) is a physical frequency band. For each action dimension, a pseudorandom white Gaussian sequence is filtered by a stable LTI band-pass filter \(H_{\mathcal{A}}\), producing
\[
W_{\mathrm{raw}}=H_{\mathcal{A}}*X,\qquad W=\frac{W_{\mathrm{raw}}}{\mathrm{Std}(W_{\mathrm{raw}})},
\]
and the watermarked policy uses
\[
\tilde a_k=\mu_\theta(o_k)+\Sigma_\theta(o_k)W_k.
\]

Two formal requirements structure the method. **W1: Marginal Distribution Preservation** requires
\[
p_{\pi_\theta}(a|o)=p_{\tilde\pi_\theta}(a|o),
\]
and CoNoCo proves this by showing that if \(W_{\mathrm{raw}}=H*X\) with \(X\sim\mathcal{N}(0,I)\), \(H\) stable LTI, and \(W=W_{\mathrm{raw}}/\mathrm{Std}(W_{\mathrm{raw}})\), then \(W_k\sim\mathcal{N}(0,I)\). **W2: Robust Detectability** requires that a keyed detector recover the watermark from remote “glimpse sequences”
\[
G_i=\mathcal{G}_{\mathrm{map}}(s(t_i))+\eta_i,
\]
despite asynchrony, unknown dynamics, and interference. The construction is explicitly frequency-domain because the watermark is intended to survive the transformation from action space to externally observed motion.

## 3. Spectral coherency, invariance, and remote detection

The central observable in robotic CoNoCo is complex coherency,
\[
C_{XY}(f)=\frac{S_{XY}(f)}{\sqrt{S_{XX}(f)S_{YY}(f)}},
\]
whose magnitude behaves like a frequency-specific correlation coefficient. The key theoretical statement is the **Theorem (Invariance under LTI)**: if \(Y\) is the output of an LTI system \(H_{\mathrm{sys}}\) with input \(X\), then, in the absence of noise and for \(S_{XX}(f)>0\), \(H_{\mathrm{sys}}(f)\neq 0\),
\[
|C_{XY}(f)|=1,
\]
regardless of \(H_{\mathrm{sys}}\) [2512.15379]. This is the mechanism by which coherency “sees through” unknown linear dynamics.

Because the policy clock is unknown, the detector searches a grid \(\mathcal{F}_{\mathrm{search}}\subset[f_{\pi,\mathrm{lb}},f_{\pi,\mathrm{ub}}]\). For each hypothesis \(s\), it regenerates the keyed base sequence, resamples it to the glimpse clock, computes per-dimension coherency with Welch’s method, averages over the secret band, and maximizes over \(s\):
\[
D(G)=\max_{s\in\mathcal{F}_{\mathrm{search}}}\left(\frac{1}{D}\sum_{d=1}^{D}\mathrm{mean}_{f\in\mathcal{A}}\left|C_{W'_{s,d},G_d}(f)\right|\right).
\]
To handle unknown start-time offsets, the method can be augmented with GCC-PHAT, which estimates delay within \(\mathcal{A}\) before coherency is computed.

Detectability is further related to spectral signal-to-interference ratio. Under constant \(\Sigma\) and LTI dynamics, with
\[
G=H_{\mathrm{sys}}*\tilde a+\eta,
\]
the magnitude-squared coherency satisfies
\[
|C_{W,G}(f)|^2=\frac{PS(f)}{PS(f)+PN(f)}=\frac{\mathrm{SINR}(f)}{\mathrm{SINR}(f)+1}.
\]
This formalizes two practical design rules already embedded in the method: increasing exploration scale raises watermark power, and choosing \(\mathcal{A}\) away from dominant nominal behavior improves detectability.

Empirically, CoNoCo “consistently achieves near-perfect detectability across tasks and modalities,” including remote motion capture and remote camera feed, while preserving reward and anonymity; detectability is assessed with ROC AUC, anonymity with \(1-\mathrm{AUC}(k')\), and reward preservation by comparing reward distributions of watermarked and original policies [2512.15379]. The reported glimpse-length sensitivity shows that AUC saturates around \(\sim 1000\) timesteps in most environments. The same experiments also delimit the method: severe occlusions, poor viewpoint geometry, rapid LTV dynamics, time-varying \(\Sigma_k\), and deterministic policies all reduce or complicate applicability.

## 4. Thermoacoustic CoNoCo

In coupled thermoacoustic oscillators, CoNoCo is not a watermark but a spectral coherence phenomenon induced by colored stochastic forcing. The system consists of two coupled horizontal Rijke tube thermoacoustic oscillators with time-delay and dissipative coupling, and the stochastic forcing is modeled as an Ornstein–Uhlenbeck process with independently controlled intensity and correlation time [2606.22665]. The noise satisfies
\[
\dot{\xi}(t)=-\frac{1}{\tau_c}\xi(t)+\sqrt{D}\,\tau_c\,\epsilon(t),
\]
with autocorrelation
\[
\langle \xi(t)\xi(t')\rangle=s^2\exp\!\left(-\frac{|t-t'|}{\tau_c}\right),\qquad s^2=\frac{D}{\tau_c},
\]
and low-pass power spectral density
\[
S_{\xi\xi}(\omega)=\frac{\Gamma}{2\pi}\frac{D}{1+\omega^2\tau_c^2}.
\]

Coherency is quantified by the coherence factor
\[
\beta=H_p\left(\frac{f_p}{\Delta f}\right),
\]
where \(H_p\) is the height of the dominant spectral peak, \(f_p\) the peak frequency, and \(\Delta f\) the half-power bandwidth from a Lorentzian fit. The quantity \(f_p/\Delta f\) is a quality factor, so \(\beta\) increases when a sharp, high peak emerges in the spectrum. In this setting, CoNoCo denotes how temporal correlation in the noise organizes oscillations and enhances their regularity near bifurcation thresholds.

Several parameter trends are explicit. Increasing noise intensity \(\sigma\) smooths the transition from limit-cycle oscillations to amplitude death and reduces the extent of amplitude-death regions, while correlation time has little effect on the topology of the bifurcation maps. By contrast, \(\beta\) is maximized near bifurcation thresholds when \(\tau_c/T_0\approx 1\): colored noise enhances coherency most when the correlation time is comparable to the acoustic period, and for both much shorter and much longer \(\tau_c\) the coherence is reduced. The paper further states that the coherence factor rises monotonically as the control parameter \(K\) approaches saddle-node thresholds from below, making \(\beta\) “a viable early warning indicator of impending thermoacoustic instability” [2606.22665].

A related dynamical point is that amplitude-death suppression remains robust under realistic colored noise, but stronger noise erodes amplitude-death regions and can re-excite anti-phase oscillations. The combination of robustness of amplitude-death topology and sensitivity of \(\beta\) to threshold proximity is one of the clearest examples of CoNoCo as a distinction between qualitative bifurcation structure and spectral regularity.

## 5. Synchronization, stochastic resonance, and phase reduction

A large part of the CoNoCo literature concerns oscillators and noise-driven switching. For common colored noise driving identical limit-cycle oscillators, the stationary phase-difference density
\[
f_0(\theta)=\frac{\nu'}{D[g(0)-g(\theta)]+\epsilon h(0)}
\]
is governed by Fourier coefficients
\[
g_l=\mathbf{Y}_{\mathrm{G},l}^\dagger \mathbf{P}_\xi(l\omega)\mathbf{Y}_{\mathrm{G},l},
\]
so only the noise spectrum at discrete frequencies \(l\omega\) matters [1402.1839]. This establishes a spectral-matching mechanism: colored noise can favor synchrony at \(\theta\approx 0\) or clustered states when higher harmonics dominate. In the Kuramoto model with Ornstein–Uhlenbeck or random-telegraph colored noise, linear stability analysis around incoherence bridges the white-noise and quenched-disorder limits, showing that coherence thresholds depend on the correlation time and on the specific stochastic process generating the colored noise [1003.1549].

In higher-order stochastic-resonance networks, the state variables obey
\[
\frac{dx_i}{dt}=x_i-x_i^3+A\sin(\Omega t)+G(\mathbf{x},x_i)+\zeta_i(t),
\]
with a convex combination of pairwise diffusive and \(2\)-simplex coupling and OU colored noise. The principal response measure is the spectral amplification factor
\[
S=4\left(\frac{|M_1|}{A}\right)^2.
\]
Classical colored noise suppresses stochastic resonance by reducing \(S_{\mathrm{peak}}\) and shifting \(D_{\mathrm{optimal}}\) to larger values, while increasing the triadic weight \(\alpha\) exacerbates that suppression. The same paper links the resonance maximum to the second minimum of the spatial synchronization measure
\[
R=E\!\left(\langle \mathrm{std}(x_i(t))\rangle\right),
\]
so CoNoCo-like behavior appears as a joint temporal–spatial effect rather than purely temporal locking [2603.06108].

Phase-reduction studies sharpen the single-oscillator interpretation. For oscillators driven by OU noise, a rigorous small-\(\tau\) reduction yields an equivalent white-noise SDE with a Wong–Zakai drift correction and a reduced phase equation of drift–diffusion type,
\[
d\theta=\left[1+b_\theta(\theta,0)\right]dt+B_\theta(\theta,0)dW_t.
\]
The result is a noise-induced frequency shift related to the variance and to the correlation time of colored noise, improving on earlier phase-only reductions that did not predict a nonzero asymptotic shift [1905.12994]. In neural phase-response-curve estimation, colored noise likewise alters coherence constructively: the colored-noise spike-triggered average
\[
S(t)=-\sigma^2\int_0^\infty Z'(T-t')K(t-t')\,dt'
\]
is a convolution with the noise autocorrelation, and the resulting PRC estimation error is non-monotonic in the OU time constant, yielding an optimal colored noise at intermediate \(\tau\) for a fixed sample size [1409.1637].

Taken together, these results make CoNoCo in oscillatory systems fundamentally spectral. Colored correlations can sharpen, suppress, or retune coherence depending on whether the relevant objective is phase alignment, resonance gain, threshold regularity, or inverse estimation accuracy.

## 6. Related generalizations, caveats, and limitations

Beyond oscillators, colored-noise coherency has been analyzed in reaction–diffusion systems, linear nonequilibrium systems, relativistic hydrodynamics, quantum environments, and topological time-series diagnostics. In reaction–diffusion systems with stochastic inflows, the central transfer relation is
\[
P_{\mathrm{coloured}}(\kappa,\omega)=P_{\mathrm{white}}(\kappa,\omega)\cdot P_{\mathrm{Correlations}}(\omega),
\]
so the colored-noise spectrum multiplicatively modulates the system’s white-noise transfer function across spatial modes [1810.12663]. In linear overdamped systems driven by colored noise, the “spreading matrix”
\[
B(s)=2\int_0^\infty e^{At}G(t+s)\,dt
\]
enters an \(s\)-dependent Lyapunov equation for the autocovariance, and the area enclosing rate
\[
\mathcal{A}=\frac{1}{2}\big[BA C-C A^\top B^\top\big]
\]
quantifies time-irreversibility induced by temporal correlations [2111.07359]. In causal baryon diffusion, the Gurtin–Pipkin model yields noise that is colored in both time and space, with a finite-speed front and a wake inside the causal cone rather than a space-time delta function [1404.4894].

Other studies emphasize that the spectrum alone need not determine coherency. In two-qubit systems subjected to \(1/f^\alpha\) classical colored noise, environments with the same power spectrum but different microscopic configurations produce opposite behaviors for entanglement and discord, including monotonic decay in some cases and sudden death with revivals in others [1212.1484]. In persistence-based dynamic state detection, colored noise of the form \(1/f^\alpha\) can drive false positives at lower signal-to-noise ratios for \(\alpha<0\), with the effect depending on both the color and the signal’s dominant frequency [2009.00494].

These results delimit any overly narrow reading of CoNoCo. A plausible implication is that CoNoCo is best understood as a family of spectral-structure effects rather than as a single invariant phenomenon. In the robotic watermarking setting, the aim is deliberate embedding and recovery of spectral coherence under unknown observation pipelines. In thermoacoustics and oscillator theory, the focus is threshold sensitivity, clustering, resonance, or phase diffusion. In quantum and topological settings, the same colored spectrum can be insufficient if higher-order or microscopic structure matters. The common denominator is not one metric but one principle: temporal correlation in noise reorganizes what aspects of a system become coherent, detectable, or regular, and it does so through the interaction between spectral color and system-specific transfer structure.

Source: https://www.emergentmind.com/topics/colored-noise-coherency-conoco