WorksPhysics and Astronomy
Frequency Comb as a Probe of Fluctuation Mechanisms in Nanomechanical Systems
Abstract
Frequency combs are indispensable in metrology. We show that a frequency comb also probes frequency fluctuations of the system. In our nanomechanical resonator, the comb arises from self-sustained vibrations of a resonantly driven nonlinear mode. The comb lineshapes reveal exponentially correlated (Ornstein–Uhlenbeck) frequency noise. Its correlation time agrees with the independently measured relaxation time of the vibrations. The noise leads to anomalous phase diffusion. Its mechanism is generic: like the comb itself, it stems from the strong nonlinearity of the self-sustained vibrations.
Introduction
Introduction — Frequency fluctuations in vibrational systems are of major interest for physics and applications. Because of time-translation symmetry, the effect of frequency fluctuations on self-sustained vibrations, from lasers to atomic, electronic, and mechanical clocks, accumulates over time in the vibration phase. This limits the precision of clocks or, equivalently, broadens the vibrational spectrum [1, 2, 3, 4, 5], the nonzero laser linewidth being an example [6, 7]. Understanding the mechanisms of frequency and the resulting phase fluctuations remains a major challenge that has been attracting much attention across various fields [8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18].
Frequency combs can provide insight into frequency fluctuations. They present equidistant peaks in the spectrum of a vibrational system and therefore play a fundamental role in metrology and sensing, a prominent example being optical atomic clocks [19, 20, 21, 22]. Combs often arise from the vibration nonlinearity. The spectra of the individual comb lines are then determined by the frequency fluctuations of the underlying generating system. Because there are many lines, their spectra provide a highly sensitive probe of the fluctuation mechanism.
In this paper, we show that the strong nonlinearity underlying the comb formation also shapes the frequency fluctuations. Accumulated in the phase, they give rise to pronounced anomalous phase diffusion. The frequency fluctuations primarily come from fluctuations of the amplitude of the self-sustained nonlinear vibrations. A major source of amplitude fluctuations is broadband thermal noise [15, 23, 24, 25, 26, 27, 28, 29]. When transferred to the vibration frequency, the noise is filtered since the amplitude typically has a relaxation time far longer than the vibration period. The resulting filtered white noise with exponential correlations is known as Ornstein–Uhlenbeck (OU) noise. Consequently, phase diffusion displays a characteristic time dependence, and the comb lines are strongly non-Lorentzian.
We perform high-resolution measurements of a frequency comb on a nanomechanical system. Its lowest-frequency eigenmode is resonantly driven into forced vibrations at the drive frequency . When the drive becomes sufficiently strong, these vibrations become unstable, and the mode starts oscillating about the state of forced vibrations [30] at a frequency much smaller than . In the frame rotating at , the self-sustained oscillations are strongly non-sinusoidal.
In the laboratory frame, these slow non-sinusoidal oscillations modulate the forced vibrations and produce equidistant peaks in the mode’s spectrum, a phononic frequency comb. We unambiguously trace the shapes and widths of all observed comb lines to OU frequency noise.
In a broader context, frequency combs present a highly sensitive means to explore spectral features. Thanks to the multitude of the comb lines, they are ideally suited to reveal the mechanisms of the underlying fluctuations. We expect this generic approach to equally apply to other types and implementations of frequency combs.
System and setup — The nanomechanical resonator investigated in this work is a doubly-clamped string as depicted in the false-color scanning electron micrograph of Figure 1(a). It is fabricated from strongly pre-stressed stoichiometric silicon nitride on a fused silica substrate, enabling large, dissipation-diluted quality factors, and features a width of 270 nm, a thickness of 100 nm and a length of 55 μm.
Figure 1. (a) False-color scanning electron micrograph of a doubly-clamped silicon nitride string resonator (green) and adjacent electrodes (gold). (b) Pink line depicts a theoretical trajectory toward the stable limit cycle in the rotating frame in the weak damping limit. The starting point is shown by an open circle. The stable limit cycle is shown by the blue trajectory. Gray lines show the contours of . The measured limit cycle is shown as black dots for a drive power of 0 dBm and a drive frequency detuned from resonance by 840 Hz. (c) Phononic frequency comb generated from a single driven nonlinear nanomechanical mode for the same drive parameters, including measured data (black) and global fit of comb lines with Ornstein–Uhlenbeck (OU) noise (red). (d)–(f) Close-ups of the lines along with the OU fit (see Supplemental Material [33] for close-ups of further comb lines).
The two flanking electrodes adjacent to the string enable dielectric drive through the Kelvin polarization force, while a heterodyne microwave-cavity-assisted detection scheme is used for the readout of the mechanical vibration [31, 32]. The cavity is driven on resonance at 3.6 GHz to exclude backaction effects while maximizing the detection sensitivity. Application of a DC voltage allows frequency tuning of the fundamental out-of-plane mode of the nanostring. We keep it fixed at 5 V to preclude coupling to other flexural modes, so that only a single mode is excited. All experiments are performed at room temperature (295 K) and in high vacuum at about mbar, using an H2LII lock-in amplifier (Zurich Instruments).
At weak drive powers, the nanomechanical resonator behaves as a harmonic oscillator, and its Lorentzian response yields the bare eigenfrequency MHz and the linewidth Hz, where is defined as the amplitude decay rate. The resonator is thus strongly underdamped in the laboratory frame, with a quality factor of approx. . As the drive power increases, the frequency response becomes asymmetric, indicating the onset of a Duffing nonlinearity. At even stronger drive, we enter the beyond-Duffing regime, where higher-order coefficients are needed to describe the response; Ref. [34] details how we accurately determine them. In the laboratory frame, the mode coordinate obeys the equation of motion
where comprises the linear and nonlinear damping as well as the resonantly induced friction force (RIFF) responsible for the comb generation [30, 35]. The quadratic and cubic nonlinearities are denoted by and , respectively, is the effective mass, and are the drive amplitude and frequency, and is the thermal noise.
We describe the dynamics in the frame rotating at by introducing the slow variables
where is the momentum in the laboratory frame, and and are the in-phase and quadrature components of the vibration in phase space. The rotating-wave approximation (RWA) averages over the fast oscillations at .
Within the RWA, the conservative dynamics are governed by the effective Hamiltonian
for a detuning . The nonlinear coefficient is determined primarily by the Duffing nonlinearity [34]. Higher-order nonlinear coefficients enter as additional terms in powers of . A detailed theoretical description is given in the Supplemental Material [33].
Figure 1(b) shows the dynamics in the quasienergy landscape for a detuning Hz and a drive power of 0 dBm. We plot the theoretical evolution (pink trace) from an initial condition (pink open circle) towards the stable limit cycle (blue trace). It includes the RIFF and the first three nonlinear coefficients of ; Ref. [33] details the model and lists all parameters. The resulting limit cycle is profoundly non-elliptical, which is a signature of non-sinusoidal vibration. We overlay the measured limit-cycle trajectory for the same parameters (black dots), calibrated as described in Ref. [34]. Theory and experiment are in excellent agreement.
Frequency comb with phase diffusion — Figure 1(c) shows the measured power spectral density for the same parameters as in Figure 1(b), featuring the frequency comb. In the spectrum, the non-sinusoidal vibrations in the rotating frame appear as equidistant peaks around the drive frequency, spaced by the limit-cycle frequency Hz.
This frequency comb is well-defined because the vibrations are underdamped also in the rotating frame, . Without noise, the motion along the limit cycle is then periodic with phase , and the comb lines are -peaks at [30]. The measured lines, however, have a finite width well above the 30 mHz spectral resolution of the experiment (see Figure 1(d)–(Figure 1f) and S3 for close-ups of the comb lines in Figure 1(c)).
In the presence of frequency noise , the phase obeys
For stationary Gaussian fluctuations, the power spectrum becomes
where is the comb line index, and are the Fourier components of the self-sustained vibrations. The function is the noise correlation function. It determines the lineshape of the comb lines and, through the factor , how the linewidths grow with comb line number.
N ature of the frequency noise—The most common model of frequency noise is white Gaussian noise, cf. [1, 10, 14, 24, 28]. It yields Lorentzian comb lines with widths growing as . Another frequently encountered type is noise, which can be modeled as Gaussian but not white. It has been observed with in nano-[10, 15, 24, 36] and micromechanical systems [37]. The effect of noise on the spectrum of a nonlinear resonator was theoretically investigated in Refs. [38, 39], and frequency fluctuations due to low-frequency fluctuators were recently reported in nanomechanical phononic systems [40, 41, 42].
A different model of noise that has been broadly used in various contexts, from the theory of rare events to stochastic resonance to quantum optics, cf. [43, 44, 45], is exponentially correlated noise described in the seminal paper of Uhlenbeck and Ornstein [46]. This OU noise is low-pass filtered white Gaussian noise, its power spectrum is Lorentzian. The three noise models are discussed in more detail in Ref. [33].
For our phononic frequency comb, the mechanism outlined in the introduction predicts OU noise: Thermal noise drives fluctuations of the quasienergy of the limit cycle, which relax over a finite time. Since depends on , these fluctuations translate into frequency fluctuations with the noise correlation function
where is the noise correlation time and is the noise strength. Inserting Eq. (6) into Eq. (5), the exponent becomes .
We fit the OU-determined lineshape to all comb lines simultaneously, with and as global fit parameters. The fit, included as a red line in Figure 1(c)–(Figure 1f) and S3, yields and . The fitted lineshapes (red) show excellent agreement with the measured comb lines (black). Fits with white or noise, in contrast, fail to describe the lineshapes [33]. This unambiguously identifies OU frequency noise as the source of phase diffusion. To the best of our knowledge, this is the first direct observation of OU frequency noise in a nanomechanical resonator.
Comb linewidths — The correct description of the lineshapes allows us to extract the linewidths of all comb lines. We define the linewidth of each line as its integral width, i.e., the spectral area divided by the peak height. For the theory, we evaluate Eq. (5) with the OU correlation function, Eq. (6), for the parameters obtained from the global fit. For the data, we use the same definition, but extracting the spectral area requires care because the tails of the comb lines merge with the noise floor. The End Matter describes how we approximate the tails to correctly estimate the noise floor. Figure 2 shows the resulting linewidths as a function of comb line number . Measured (black) and theoretical (red) linewidths coincide within the experimental uncertainty. They grow with , but much more slowly than the parabola expected for white Gaussian noise (gray dashed line). Scaled to the linewidth at , the parabola deviates strongly from the data for all higher comb lines.
Figure 2. Comb linewidths. Integral linewidths of the comb lineshapes versus comb line number , extracted from the measured spectra (black) and calculated from the fitted OU model (red). The two coincide within the experimental uncertainty. White Gaussian noise would instead lead to a quadratic growth (gray, , scaled to ).
Noise origin — Frequency fluctuations of a self-sustained oscillator have two sources [16]: noise acting directly on the phase , and noise acting on the quasienergy , which the nonlinearity converts into fluctuations of . Direct phase noise comes primarily from white frequency noise and thus yields Lorentzian comb lines with widths growing as , which is incompatible with our data. The strong nonlinearity, evident from the profoundly non-elliptical limit cycle in Figure 1(b) and the many comb lines in Figure 1(c), makes the conversion of quasienergy fluctuations a dominant mechanism. The correlation time of these fluctuations sets the correlation time of the resulting frequency noise and makes it of the OU type [28], as observed.
Limit-cycle formation — Our analysis attributes the correlation time of the frequency noise to the relaxation of the quasienergy toward the limit cycle. We measure this relaxation directly by recording how the system approaches the limit cycle after the drive is tuned across the supercritical Hopf bifurcation [30]. Since this measurement does not rely on noise, it provides an independent determination of . We start with the drive at 0 dBm at zero detuning, where the system rests at a stable fixed point in the rotating frame, and abruptly switch the drive frequency to Hz, where the previously stable fixed point becomes unstable and is replaced by a stable limit cycle. Both quadratures recorded over time are displayed as a phase space plot in Figure 3(a). We observe the system leaving the fixed point (gray), evolving towards (teal) and settling on the limit cycle (orange). To quantify the relaxation time of the system towards the limit cycle, we record the crossings of the trajectory through at the smaller of the two values per revolution, i.e., the one with , and plot the Poincaré map versus in Figure 3(b).
Figure 3. Formation of the limit cycle. (a) Phase space trajectory where the detuning of the drive (0 dBm) is abruptly switched from to 840 Hz. The system leaves the now unstable fixed point (gray), evolves towards the stable cycle (teal), and finally settles on it (orange). (b) Poincaré map versus of the crossings of the trajectory through with . The teal points show the approach to the limit cycle, and the orange points the noise-broadened stable limit cycle. The dashed line represents the line. (c) Step between consecutive crossings versus crossing number in the transient regime (teal). The exponential fit (red) yields the Floquet exponent corresponding to a relaxation time ms, in agreement with the OU correlation time from the comb lineshapes.
The crossings move along the diagonal (teal) towards a dense cluster (orange), which marks the limit cycle. The step between consecutive crossings is depicted in Figure 3(c). It decays exponentially with crossing number , and a fit (red line) yields the Floquet exponent .
The frequency of the established limit cycle is Hz. This results in a period of ms, giving the relaxation time ms. This agrees within its uncertainty with the correlation time ms of the OU noise obtained from the global lineshape fit (see [33] for more details of the measurements). This measurement was taken months after the data in Figure 1, which explains the slight change in the limit-cycle frequency and noise correlation time.
Conclusion — We use the spectra of the phononic frequency comb lines as a probe of the mechanism of frequency fluctuations of the source of the comb. The high sensitivity of the method stems from the multiplicity of the comb lines. In our system, the comb emerges as a result of nonsinusoidal self-sustained vibrations in the rotating frame of a resonantly driven single nanomechanical mode. We find that the lineshapes of all comb lines are determined by phase diffusion due to an exponentially correlated frequency noise. The noise has the same origin as the comb itself: both come from the strong nonlinearity of the self-sustained vibrations. In particular, this nonlinearity converts thermal-noise induced quasienergy fluctuations into exponentially correlated frequency fluctuations. An independent and direct measurement of the time of relaxation toward the state of self-sustained vibrations agrees with the noise correlation time, confirming our interpretation. The fluctuation mechanism is therefore generic. The results provide a new insight into the origin of frequency fluctuations that ultimately limit the precision of comb-based measurements and suggest new strategies for improving comb technology.
Acknowledgments
M.K., D.K.J.B., W.B., and E.M.W. gratefully acknowledge financial support from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) through Project-ID No. 425217212—SFB 1432 and under Germany’s Excellence Strategy—EXC-2111—390814868. The research is further supported by the Bavarian state government with funds from the Hightech Agenda Bavaria. M.I.D. gratefully acknowledges partial support from the Gordon and Betty Moore Foundation (Award No. GBMF12214) and the US Defense Advanced Research Projects Agency (Grant No. HR0011-23-2-004). D.K.J.B. acknowledges funding by the Carl-Zeiss-Foundation through the Carl-Zeiss-Foundation Center QPhoton.
Data Availability
Experimental data are available at 10.5281/zenodo.xxxxxxx.
End Matter
*Spectral resolution—*Resolving the lineshapes of the comb lines requires a spectral resolution well below their linewidths. We record the spectra with an HF2LI lock-in amplifier (Zurich Instruments), which demodulates the signal at the drive frequency, thereby shifting the comb to baseband. A sampling rate of 30 kSa/s then covers the full comb, and samples correspond to a record length of 35 s. The resulting resolution of 30 kSa/s/ mHz lies well below the linewidths of all comb lines, ensuring that the comb measurement is resolved. [Figure S3].
Figure S3. (a)–(e) Close-ups of the comb lines of the comb in Figure 1(c) not shown in Figure 1(d)–(Figure 1f) of the main text, including both the measured spectra (black) and global OU fit (red) for , , , , and . (f) The main tone (forced vibration at the drive frequency) is limited by the spectral resolution of 30 mHz.
Phase-space rotation—The detection chain imparts a constant phase offset on the signal, mainly through the propagation delay in the cables. We compensate for it by rotating the measured quadratures by rad in all phase-space plots [Figure 1(b) and Figure 3], determined by minimizing the distance between the measured points and the theoretical limit cycle. Such constant rotation leaves the shape of the trajectories unchanged and only aligns the measured frame with the theoretical one, in which the drive acts along [Eq. (3)].
Global lineshape fit—Least-squares fitting is not suited for power spectra, since the fluctuations of each spectral bin scale with its mean. For Gaussian noise in the time domain, the real and imaginary parts of each Fourier component are Gaussian. Since we perform no spectral averaging, the power in each bin follows a distribution, i.e., an exponential distribution [47]. While a Hann window was applied to the data, the resulting correlation between adjacent spectral bins has a negligible effect given our high frequency resolution relative to the broad lineshapes fitted here. We therefore maximize the Whittle likelihood [48, 49, 50], which models the measured power in bin as exponentially distributed with mean given by the model spectrum plus a constant noise floor. This amounts to maximizing
with respect to and , jointly for all comb lines. We include only bins more than 5 dB above the estimated noise floor, so that the noise floor does not bias the estimates of and .
Linewidth extraction—We extract the linewidth of each comb line in four steps. First, we determine the tails of the comb lines, which carry a substantial part of the peak area but disappear in the noise floor. Far from the line center, the OU lineshape decays as the inverse fourth power of the distance to the center (see Eq. (42) of the Supplemental Material [33]). Figure 4 plots the positive half of the measured data of two different comb lines ( and ) on a log-log scale in gray. The red solid and blue dashed lines represent the OU theoretical lineshape and the asymptotic tail, respectively. Where the OU lineshape and its asymptotic approximation agree to within 0.5 dB, we replace the data by the asymptotic tail. The resulting hybrid lineshape is included in Figure 4 in black. Second, we estimate the noise floor of each line as the median difference between the data (gray) and the asymptotic tails (blue) in the noise-dominated region of the spectrum where the noise floor of the experiment exceeds the asymptotic tail by more than an order of magnitude. The obtained noise floor is included in Figure 4 as a green dotted line. Third, we determine the peak height of each comb line, which is set by the oscillator’s strength in Eq. (5). The global
Figure 4. Log-log representation of comb lines for (a) and (b) . Positive half of the measured peak is plotted as a function of the frequency difference from the center frequency of the comb line (gray). The theoretical OU lineshape obtained from the global fit and its asymptotic tail are included as red solid and blue dashed line, respectively. The reconstructed hybrid OU lineshape is depicted in black. The resulting estimate of the noise floor is shown as a green dotted line.
fit uses normalized lines, so that and are its only free parameters. We therefore obtain the strengths from a second fit of Eq. (5), with and fixed to the values of the global fit. The noise floors and strengths are listed in Table S1 of the Supplemental Material [33]. Finally, we obtain the peak area as the sum of the noise-floor-corrected central part, integrated with the trapezoidal rule, and the analytically integrated tails. The integral linewidth is the ratio of the peak area to the peak height. The experimentally determined linewidths depicted in Figure 2 result from averaging over 100 measurements, and shown along with their standard deviation. The values of the linewidths of all comb lines predicted by the theory and extracted from the experiment are compared in Table S2 of the Supplemental Material [33].
| Noise floor (V/Hz) | (V) | |
| 3 | ||
| 2 | ||
| 1 | ||
| -1 | ||
| -2 | ||
| -3 | ||
| -4 | ||
| -5 |
Table S1. Noise floor and oscillator’s strength of each comb line .
| Theory (Hz) | Measured (Hz) | Std. dev. (Hz) | |
| 3 | 16.006 | 16.777 | 0.959 |
| 2 | 9.468 | 9.794 | 0.542 |
| 1 | 3.292 | 3.528 | 0.284 |
| -1 | 3.292 | 3.529 | 0.284 |
| -2 | 9.468 | 9.777 | 0.541 |
| -3 | 16.006 | 16.405 | 0.926 |
| -4 | 22.577 | 23.134 | 1.177 |
| -5 | 29.149 | 30.658 | 2.009 |
Table S2. Theoretical versus experimental linewidths (mean of 100 measurements and standard deviation).
SUPPLEMENTAL MATERIAL
THEORY
Fluctuations of action-angle variables
We consider a mode with natural vibration frequency , driven at a frequency with amplitude . In a frame rotating at the drive frequency, the mode is described by the quadratures and ,
where the rotating-wave-approximated Hamiltonian is
Here, denotes the detuning and are the nonlinear coefficients of the rotating-wave Hamiltonian. For a nonlinear potential, the coefficients can be related to the corresponding microscopic nonlinearities. For example, for a Duffing nonlinearity with coefficient , one finds . Here we include all three , and , which are necessary to describe the backbone curve of our resonator [34]. In general, the coefficients are renormalized by other nonlinearities, including nonlinear coupling to other modes [30, 34, 51]. The terms and describe dissipation. In the absence of the drive, they reduce to and . The associated fluctuating forces satisfy the fluctuation-dissipation relation [51]
where is the diffusion coefficient and .
Under resonant driving, nonlinear friction generates higher-order dissipative terms that can give rise to a limit cycle in the rotating frame. A simple phenomenological form of such a friction force is RIFF [35],
where the averaging is over one period . In the rotating frame, accounting for this nonlinear friction, the dissipative terms become
In the absence of noise, these nonlinear dissipative terms can prevent and from relaxing to a fixed point, instead giving rise to closed trajectories corresponding to a limit cycle.
For weak noise, the motion remains approximately periodic. A remarkable feature of these vibrations is that once the driving exceeds the critical value for the onset of the vibrations, they quickly become strongly non-sinusoidal, which underlies the onset of the frequency comb. The vibrations are conveniently described in terms of action-angle variables, and for a system with Hamiltonian [52]. In the absence of dissipation and noise, the action is conserved, while the phase increases linearly in time with the vibration frequency . The original variables can then be expressed as and , which are -periodic functions of . In the presence of dissipation and fluctuations, the action is no longer conserved, and the dynamics of the phase variable are modified such that
In the present system, the vibration frequency is large compared with both the decay rate and the diffusion rate [30]. Therefore, within a single vibration period, the action changes only slightly, while accumulates approximately . The long-term dynamics can thus be described by averaging Eq. (S7) over one vibration period, yielding [28]
Here, the overline indicates averaging of the periodic functions over for a given . The noise terms are given by and where the curly brackets denote an average of the periodic coefficients over , such that the correlators of the noise terms depend only on the action . In the absence of fluctuations, the stationary orbit is determined by the condition , with
where the integral is taken over the interior of the orbit . The corresponding vibration frequency of the limit cycle is , where . In the following, we assume that the noise is sufficiently weak such that fluctuations around the stationary value of the action remain small, allowing us to expand the dynamics around . The equations of motion can then be written as
where and . For a stable limit cycle, we require . The action variable therefore experiences random kicks from the fluctuating force while relaxing back to its stationary value at the rate . In contrast, increases at the rate while being influenced by fluctuations. These fluctuations have two distinct contributions. On the one hand, phase fluctuations act directly on the phase variable through . On the other hand, fluctuations in the action are converted into fluctuations of through the nonlinear dependence of the vibration frequency on the action. Specifically, the action-dependent frequency allows fluctuations in the action variable to propagate to the phase variable through the term . For a weakly nonlinear system, the contribution is typically small. In the strongly nonlinear system considered in this work, however, it can become significant. At the same time, the direct contribution is small for large-amplitude self-sustained vibrations. Therefore, in our system it is the contribution of the noise that comes from the amplitude fluctuations that becomes the dominating mechanism of the frequency noise.
In the following, we are interested in the effect of frequency fluctuations and the resulting phase diffusion on the spectral shape. For that we consider dynamics of the form
where
is assumed to be Gaussian distributed.
Gaussian phase noise and spectral broadening
We consider a general stationary Gaussian noise process . Its probability density functional can be written as
where the function and its inverse are related by
cf. [53]. Stationarity requires to depend only on the time difference. Since the probability functional is Gaussian, the noise correlator is
One way to characterize different noise processes is in terms of the resulting diffusion, cf. [54, 55]. For this purpose we define the phase difference , which undergoes a random walk . Its mean-square displacement is given by
which is therefore directly related to .
We are interested in the power spectral density of the resonator. In order to calculate it, we expand the complex quadrature in harmonics of the phase variable,
We remind that the coordinate and momentum of the oscillator are simply related to , namely, . The squared absolute values of the expansion coefficients determine the oscillator’s strength of the different comb lines, which are shifted from the drive frequency by multiples of the limit cycle frequency . For the simple case , the trajectories for are given by Jacobi elliptic functions. The double periodicity of these functions allows one to calculate the coefficients analytically [30, 35] (in those papers, was expressed as a function of , where corresponds to the value of at ). We can now write the power spectral density (PSD) of the resonator as
where we have used that .
For the stationary Gaussian noise process assumed here, the correlator can be related [53, 54, 55] to the mean square displacement of from Eq. (26):
Inserting this expression into the PSD yields
The function determines the lineshape of the th harmonic of the limit cycle. It is normalized such that . The width and shape of depend explicitly on the harmonic index and on the specific form of .
White noise
For white Gaussian noise, the correlator is
The parameter is the phase diffusion coefficient. From Eq. (26), the mean-square phase displacement grows linearly with time [1],
With Eq. (30) taken into account, this yields the PSD of the resonator
Each harmonic produces a Lorentzian line with a half-width .
1/f-alpha noise
We now compute the power spectral density (PSD) of the resonator for noise with in an appreciably broad frequency range. For this noise, the correlation function decays as a power law in time,
with the noise strength . Then, we find that the mean-square phase displacement is given by
Inserting Eq. (34) into the general expression for the PSD, Eq. (30), gives
Ornstein–Uhlenbeck noise
We now consider an Ornstein–Uhlenbeck process where the PSD of the noise follows a Lorentzian distribution and
The mean-square phase displacement of the phase difference evaluates to
To connect explicitly to the main text, we indicate that, for the Ornstein–Uhlenbeck noise, in Eq. (5) of the main text we have
For , we have , an anomalous diffusion corresponding to ballistic motion. For , the mean-square phase displacement approaches , corresponding to normal diffusion.
The oscillator PSD is found by inserting Eq. (37) into Eq. (30). The resulting integral can be reduced to the incomplete Gamma function
where , and is the lower incomplete Gamma function.
The parameter compares the correlation time of the noise with the decay time of the phase correlator for white noise of the same . For , the correlator decays while the phase displacement is still ballistic. Near its maximum, each peak then becomes a Gaussian,
with a half-width that grows linearly with . In contrast, for , the phase correlator decays slowly compared to the correlation time. For , the peaks are Lorentzian with half-width , as for white noise. This is consistent with the noise correlator in Eq. (37) approaching the white-noise correlator, Eq. (31), for .
In Figure S1, we compare the lineshape of the peak for Ornstein–Uhlenbeck noise (blue) and white noise (dark yellow), with the same and , i.e., in the intermediate regime . OU noise displays a narrower peak that falls off quickly for larger detunings from the center frequency. In this regime, the peaks from the Ornstein–Uhlenbeck type noise are significantly non-Lorentzian, and the half-width of the th peak is not proportional to . Sufficiently far away from the peak, where the absolute value of the detuning is large compared to and , we can use partial integration to approximate
Figure S1. Lineshape of the comb line near for Ornstein–Uhlenbeck noise (blue) with (i.e. ), compared with white noise assuming the same parameters (dark yellow).
Therefore, the tails of the peaks fall off as the reciprocal fourth power of the distance to the peak, in contrast to the tails of the Lorentzian peaks that fall off as the reciprocal square of this distance.
Parameter values of the effective Hamiltonian describing the experiment
The trajectories obtained in Figure 1(b) are the solution of the flow equations for the coordinates and as given in Eq. (8) and Eq. (9) neglecting the noise contributions and . The parameter values used to numerically integrate these trajectories are obtained directly from experimental data. The fundamental bare resonance frequency is MHz (where ), with a linear damping rate of Hz. The system is driven at a power of 0 dBm with a drive detuning of Hz.
As the phase space data are measured directly as voltage signals from the lock-in amplifier, the nonlinear coefficients in the flow equations correspond to the physical parameters scaled by mass and transduction, and we use the values V, s V, and s V. The RIFF damping is estimated as s V, where m V is the conversion factor between measured and physical units.
LINEWIDTH FIT DETAILS
Noise floor and peak height of comb lines
Table S1 summarizes the values of the noise floor and peak strength of each comb line required for the linewidth extraction described in the End Matter. Table S2 compares the theoretical linewidths of the OU comb lines with the experimentally obtained values along with their standard deviation. All experimental data are presented in units of voltage as obtained directly from the measurement instruments to avoid ambiguity regarding the scaling to physical units.
ADDITIONAL EXPERIMENTAL DATA
Signature of the frequency comb in the response curve
Figure S2 shows the frequency response at the drive power of 0 dBm used for the comb measurements in the main text. Its
Figure S2. Frequency response of the single nanomechanical mode at a drive power of 0 dBm. The characteristic dip in the otherwise nonlinear response heralds the formation of the frequency combs related to the vibrations in the rotating frame that cause a reduction in the amplitude of the fixed vibrations. The red dashed line marks the offset of the drive frequency from the bare resonance by 840 Hz used in the comb measurements of the main text.
overall shape is Duffing-like, with higher-order nonlinearities contributing at this drive [34]. Over a range of frequencies, the response amplitude exhibits a pronounced dip, as also reported in Ref. [30]. The dip marks the onset of the frequency comb: the lock-in amplifier records only the component at the drive frequency, and spectral weight is redistributed from this central peak to the comb lines. The dip thus identifies the drive frequencies at which the comb exists. At the deepest point, at Hz, coincides with the strongest comb, i.e., the one with the most and largest comb lines. We therefore use this offset throughout the main text.
Lineshape fits of the remaining comb lines
Figure 1(c) of the main text shows the full comb together with the global OU fit, and Figs. 1(d)–1(f) show close-ups of the lines , , and . For completeness, Fig. S3(a)–S3(e) shows close-ups of the remaining lines, , , , , and , together with the same global fit yielding rad s and ms.
The fit excellently reproduces all measured comb lines. Figure S3(f) displays the main tone directly at the drive frequency, i.e. the forced vibrations at . It is significantly narrower than the comb lines, since its phase is locked to the drive and thus not subject to phase diffusion. Its width is limited by the spectral resolution of 30 mHz.
Obtaining the frequency comb from the phase space trajectory
The comb can also be obtained from the measured phase-space trajectory. We perform a Fourier transform on the stationary part of the trajectory (orange) in Figure 3(a), after the system has settled on the limit cycle. The resulting spectrum shows comb lines spaced by Hz, and Figure S4 displays the lines to . Applying the same global lineshape fit as in the main text (orange lines) yields and ms. This correlation time agrees with the relaxation time 25(6) ms obtained from the Poincaré analysis of the same data.
Figure S4. Exemplary comb lines of the frequency comb obtained from the Fourier transformation of the phase space trajectory of the limit cycle in Figure 3(a) of the main text. The global OU fit (orange lines) yields and ms.
The two data sets that we present were recorded many months apart. Over this period, the limit-cycle frequency shifted from 618.9 Hz to 611.7 Hz, i.e., by only 1.2%, which demonstrates the long-term stability of the system. The associated shift in the quasienergy accounts for the modified phase diffusion coefficient .
Lineshapes of spectral peaks in the stable regime
As described in the main text, the primary noise responsible for phase diffusion is OU noise. However, in the regime right before the onset of the frequency comb, thermal white Gaussian noise is dominating. In this regime, the spectrum is described by two spectral peaks equidistantly spaced from the drive frequency [56]. In Figure S5, the Lorentzian fit of the spectral peak measured for a drive frequency detuning from the bare resonance frequency by 0 Hz and a drive power of 0 dBm is depicted. The fit yields a linewidth of 1.36 Hz. With increasing detuning frequency, the system approaches the Hopf bifurcation and the linewidth of the peaks decreases [30]. This is illustrated in Figure S5 for the peak at a drive frequency detuning from the bare resonance frequency by 70 Hz and a drive power of 0 dBm. The Lorentzian fit yields a reduced linewidth of 0.58 Hz.

Figure S5. Peak in the spectrum () of the driven nonlinear resonator in the stable regime (before limit cycle formation). (a) Spectral peak measured at a drive power of 0 dBm with the drive frequency detuning from the bare resonance by 0 Hz. Green line depicts Lorentzian fit. (b) Spectral peak measured at the same drive power with a drive frequency offset from the bare resonance of 70 Hz. Green line depicts Lorentzian fit yielding a narrower linewidth.
NOISE MODEL COMPARISON
White Gaussian noise
White Gaussian noise yields the comb spectrum of Eq. (33), with Lorentzian lines whose widths grow as . First, the measurement of the linewidths presented in Figure 2 of the main text does not yield the expected scaling. Second, Figure S6 shows the comb line along with the global OU fit (red) and a Lorentzian fit (light blue). Clearly, the Lorentzian fails to describe the measured lineshape.

Figure S6. Comparison of OU noise and white Gaussian noise as the origin of the phase diffusion broadening the comb lines. Data of the comb line is shown along with its global OU fit (red) and a Lorentzian fit according to Eq. (33) for white Gaussian noise (light blue).
1/f-alpha Gaussian noise
Gaussian noise with a spectrum has also been observed in micro- and nanomechanical systems [10, 37]. We therefore test whether it can explain the observed phase diffusion. The noise correlation function of noise can be described by Eq. (34). Figure S7 compares the measured , and comb lines with the theoretical expectation for noise. For this, we solve the integral in Eq. (5) for the noise correlation function Eq. (34) for a value of and a dimensionless exponent . We note that carries units of . The theory can be adjusted to agree with the comb lineshape. However, it does not describe the subsequent comb lines for or . We therefore conclude that Gaussian noise can be excluded as a potential source of the phase diffusion.
Figure S7. Comparison of comb lines with theory for Gaussian noise as the origin of the phase diffusion broadening the comb lines. (a)–(c) Comb lines with , , along with the theoretically predicted lineshape in the presence of noise with and (violet).
References
References
- [1]I. L. L. Berstein, On fluctuations in the neighborhood of periodic motion of an auto-oscillating system, C. R. (Dokl.) Acad. Sci. URSS 20, 11 (1938).
- [2]S. M. Rytov, Fluctuations in oscillating systems of the Thomson type, I. Sov. Phys. JETP 2, 217 (1956).
- [3]M. Lax, Classical noise. V. Noise in self-sustained oscillators, Phys. Rev. 160, 290 (1967).DOI
- [4]A. Demir, A. Mehrotra, and J. Roychowdhury, Phase noise in oscillators: a unifying theory and numerical methods for characterization, IEEE Trans. Circuits Syst. I Fundam. Theory Appl. 47, 655 (2000).DOI
- [5]Y. Zhang and M. I. Dykman, Spectral effects of dispersive mode coupling in driven mesoscopic systems, Phys. Rev. B 92, 165419 (2015).DOI
- [6]A. L. Schawlow and C. H. Townes, Infrared and optical masers, Phys. Rev. 112, 1940 (1958).DOI
- [7]M. P. van Exter, W. A. Hamel, and J. P. Woerdman, Nonuniform phase diffusion in a laser, Phys. Rev. A 43, 6241 (1991).
- [8]C. Henry, Theory of the linewidth of semiconductor lasers, IEEE J. Quantum Electron. 18, 259 (1982).DOI
- [9]A. Hajimiri and T. Lee, A general theory of phase noise in electrical oscillators, IEEE J. Solid-State Circuits 33, 179 (1998).DOI
- [10]K. Y. Fong, W. H. P. Pernice, and H. X. Tang, Frequency and phase noise of ultrahigh-Q silicon nitride nanomechanical resonators, Phys. Rev. B 85, 161410 (2012).DOI
- [11]W. Loh, S. Yegnanarayanan, R. J. Ram, and P. W. Juodawlkis, Unified theory of oscillator phase noise I: White noise, IEEE Trans. Microw. Theory Tech. 61, 2371 (2013).DOI
- [12]J. Moser, A. Eichler, J. Güttinger, M. I. Dykman, and A. Bachtold, Nanotube mechanical resonators with quality factors of up to 5 million, Nat. Nanotechnol. 9, 1007 (2014).DOI
- [13]A. D. Ludlow, M. M. Boyd, J. Ye, E. Peik, and P. O. Schmidt, Optical atomic clocks, Rev. Mod. Phys. 87, 637 (2015).arxiv.org/abs/1407.3493
- [14]M. Sansa, E. Sage, E. C. Bullard, M. Gély, T. Alava, E. Colinet, A. K. Naik, L. G. Villanueva, L. Durafourq, M. L. Roukes, G. Jourdan, and S. Hentz, Frequency fluctuations in silicon nanoresonators, Nat. Nanotechnol. 11, 552 (2016).DOI
- [15]O. Maillet, X. Zhou, R. R. Gazizulin, R. Ilic, J. M. Parpia, O. Bourgeois, A. D. Fefferman, and E. Collin, Measuring frequency fluctuations in nonlinear nanomechanical resonators, ACS Nano 12, 5753 (2018).DOI
- [16]A. Bachtold, J. Moser, and M. I. Dykman, Mesoscopic physics of nanomechanical systems, Rev. Mod. Phys. 94, 045005 (2022).DOI
- [17]K. Kanellopulos, F. Ladinig, S. Emminger, P. Martini, R. G. West, and S. Schmid, Comparative analysis of nanomechanical resonators: sensitivity, response time, and practical considerations in photothermal sensing, Microsyst. Nanoeng. 11, 28 (2025).DOI
- [18]V. Vachkula, D. Dolinina, A. Vladimirov, and G. Huyet, Coherence properties of Kerr frequency combs under noisy injection and optical feedback, Phys. Rev. A 114, 033502 (2026).
- [19]S. A. Diddams, L.-S. Ma, J. Ye, and J. L. Hall, Broadband optical frequency comb generation with a phase-modulated parametric oscillator, Opt. Lett. 24, 1747 (1999).pubmed.ncbi.nlm.nih.gov/18079922
- [20]T. Fortier and E. Baumann, 20 years of developments in optical frequency comb technology and applications, Commun. Phys. 2, 153 (2019).DOI
- [21]T. Udem, R. Holzwarth, and T. W. Hänsch, Optical frequency metrology, Nature 416, 233 (2002).
- [22]A. Bisht, P. Arora, and V. G. Achanta, Precision timekeeping with atomic clocks: evolution and future directions, Nat. Commun. 17, 5284 (2026).DOI
- [23]T. R. Albrecht, P. Grüter, D. Horne, and D. Rugar, Frequency modulation detection using high-Q cantilevers for enhanced force microscope sensitivity, J. Appl. Phys. 69, 668 (1991).DOI
- [24]A. N. Cleland and M. L. Roukes, Noise processes in nanomechanical resonators, J. Appl. Phys. 92, 2758 (2002).DOI
- [25]A. N. Cleland, Thermomechanical noise limits on parametric sensing with nanomechanical resonators, New J. Phys. 7, 235 (2005).DOI
- [26]L. Huang, S. M. Soskin, I. A. Khovanov, R. Mannella, K. Nin- ios, and H. B. Chan, Frequency stabilization and noise-induced spectral narrowing in resonators with zero dispersion, Nat. Commun. 10, 3930 (2019).
- [27]A. Demir, Understanding fundamental trade-offs in nanomechanical resonant sensors, J. Appl. Phys. 129, 044501 (2021).DOI
- [28]N. J. Miller, S. W. Shaw, and M. I. Dykman, Suppressing frequency fluctuations of self-sustained vibrations in underdamped nonlinear resonators, Phys. Rev. Appl. 15, 014024 (2021).DOI
- [29]S. Schmid, L. G. Villanueva, and M. L. Roukes, Measurements and noise, in Fundamentals of Nanomechanical Resonators (Springer International Publishing, 2023) Chap. 5, pp. 145–174, 2nd ed.DOI
- [30]J. S. Ochs, D. K. J. Boneß, G. Rastelli, M. Seitner, W. Belzig, M. I. Dykman, and E. M. Weig, Frequency comb from a single driven nonlinear nanomechanical mode, Phys. Rev. X 12, 041019 (2022).arxiv.org/abs/2207.04030
- [31]Q. Unterreithmeier, E. M. Weig, and J. P. Kotthaus, Universal transduction scheme for nanomechanical systems based on dielectric forces, Nature 458, 1001 (2009).DOI
- [32]T. Faust, P. Krenn, S. Manus, J. P. Kotthaus, and E. M. Weig, Microwave cavity-enhanced transduction for plug and play nanomechanics at room temperature, Nat. Commun. 3, 728 (2012).DOI
- [33]See Supplemental Material at [URL will be inserted by publisher] for detailed theoretical analysis, further information on the fitting procedures, and additional experimental data.
- [34]M. Kallergi, D. K. J. Boneß, M. Seitner, W. Belzig, and E. M. Weig, Extracting higher-order nonlinearities in nanomechanical resonators using the backbone relation, arXiv:2608.02008 (2026).arxiv.org/abs/2608.02008
- [35]M. I. Dykman, G. Rastelli, M. L. Roukes, and E. M. Weig, Resonantly induced friction and frequency combs in driven nanomechanical systems, Phys. Rev. Lett. 122, 254301 (2019).DOI
- [36]M. L. Roukes, 1/f frequency noise in mechanical resonators scales inversely with volume, not with quality factor, arXiv:2609.16668 (2026).arxiv.org/abs/2609.16668
- [37]F. Sun, X. Dong, J. Zou, M. I. Dykman, and H. B. Chan, Correlated anomalous phase diffusion of coupled phononic modes in a sideband-driven resonator, Nat. Commun. 7, 12694 (2016).DOI
- [38]A. Demir, Phase noise and timing jitter in oscillators with colored-noise sources, IEEE Trans. Circuits Syst. I Fundam. Theory Appl. 49, 1782 (2002).DOI
- [39]E. Kenig, M. C. Cross, L. G. Villanueva, R. B. Karabalin, M. H. Matheny, R. Lifshitz, and M. L. Roukes, Optimal operating points of oscillators using nonlinear resonators, Phys. Rev. E 86, 056207 (2012).DOI
- [40]M. Yuksel, M. P. Maksymowych, O. A. Hitchcock, F. M. Mayor, N. R. Lee, M. I. Dykman, A. H. Safavi-Naeini, and M. L. Roukes, Intrinsic phononic dressed states in a nanomechanical system, Nat. Phys. 22, 1646 (2026).arxiv.org/abs/2502.18587
- [41]M. P. Maksymowych, M. Yuksel, O. A. Hitchcock, N. R. Lee, F. M. Mayor, W. Jiang, M. L. Roukes, and A. H. Safavi-Naeini, Spectral diffusion of nanomechanical resonators due to single quantum defects, Phys. Rev. Appl. 24, 044066 (2025).DOI
- [42]A. Hitchcock, F. M. Mayor, W. Jiang, M. P. Maksymowych, S. Malik, and A. H. Safavi-Naeini, Correlated dephasing in a piezoelectrically transduced silicon phononic waveguide, Phys. Rev. Appl. 25, 024005 (2026).arxiv.org/abs/2502.16426
- [43]A. J. Bray and A. J. McKane, Instanton calculation of the escape rate for activation over a potential barrier driven by colored noise, Phys. Rev. Lett. 62, 493 (1989).pubmed.ncbi.nlm.nih.gov/10040249
- [44]L. Gammaitoni, P. H. Hänggi, P. Jung, and F. Marchesoni, Stochastic resonance, Rev. Mod. Phys. 70, 223 (1998).
- [45]A. Carmele, S. Parkins, and A. Knorr, Quantum-optical realization of an Ornstein-Uhlenbeck-type process via simultaneous action of white noise and feedback, Phys. Rev. A 102, 033712 (2020).DOI
- [46]G. E. Uhlenbeck and L. S. Ornstein, On the theory of the Brownian motion, Phys. Rev. 36, 823 (1930).DOI
- [47]P. J. Brockwell and R. A. Davis, Time Series: Theory and Methods (Springer New York, 1991).
- [48]P. Whittle, Estimation and information in stationary time series, Ark. Mat. 2, 423 (1953).DOI
- [49]I. J. Myung, Tutorial on maximum likelihood estimation, J. Math. Psychol. 47, 90 (2003).DOI
- [50]C. Roume, A guide to Whittle maximum likelihood estimator in MATLAB, Front. Netw. Physiol. 3, 1204757 (2023).DOI
- [51]M. I. Dykman and M. A. Krivog, Classical theory of nonlinear oscillator interacting with a medium, Phys. Status Solidi B 48, 497 (1971).DOI
- [52]L. D. Landau and E. M. Lifshitz, Mechanics: Volume 1 (Elsevier, 1982).
- [53]R. P. Feynman and A. R. Hibbs, Quantum Mechanics and Path Integrals (McGraw-Hill, New York, 1965).
- [54]P. W. Anderson and P. R. Weiss, Exchange narrowing in paramagnetic resonance, Rev. Mod. Phys. 25, 269 (1953); P. W. Anderson, Mathematical model for the narrowing of spectral lines by exchange or motion, J. Phys. Soc. Jpn. 9, 316 (1954).DOI
- [55]R. Kubo, Note on the stochastic theory of resonance absorption, J. Phys. Soc. Jpn. 9, 935 (1954); Stochastic theory of magnetic resonance, Nuovo Cimento 6, 1063 (1957).DOI
- [56]I. S. Huber, G. Rastelli, M. J. Seitner, J. Kölbl, W. Belzig, M. I. Dykman, and E. M. Weig, Spectral evidence of squeezing of a weakly damped driven nanomechanical mode, Phys. Rev. X 10, 021066 (2020).DOI








