WorksPhysics and Astronomy
Redefining the limits of real-time noise cancellation in optical fiber links
Abstract
A broad and growing array of applications rely on the faithful transmission of ultrastable optical signals over noisy paths, requiring cancellation of environmentally induced noise. A generally accepted limit constrains how well the path length noise can be suppressed in real time. Here, we show that this standard limit is not fundamental and can be improved upon. By considering the temporal correlations between the round-trip and one-way signals, we develop a new framework for optimizing the noise cancellation feedback signal for any spatial distribution of noise along the signal path. We use this framework to surpass the standard limit in two sets of experiments. First, we demonstrate noise cancellation in a deployed urban optical fiber, where we achieve noise suppression approximately 6 dB beyond the standard limit. Then, in a reconfigurable lab-based fiber-optic testbed, we show that, for certain spatial distributions of noise, suppression of well over 10 dB beyond the standard limit is readily achievable. With the use of digital signal processing to generate the correction signal, our new technique requires no new electro-optic hardware relative to the field-standard noise cancellation scheme. This will allow for widespread adoption of these improved limits in existing systems, with applications in optical clock distribution, optical clock comparisons for fundamental physics and geodesy, and quantum networking.
requires no new electro-optic hardware relative to the field-standard noise cancellation scheme. This will allow for widespread adoption of these improved limits in existing systems, with applications in optical clock distribution, optical clock comparisons for fundamental physics and geodesy, and quantum networking.
Transmission of optical signals through deployed fibers is a cornerstone of modern technology, quietly enabling nearly every facet of modern life. Remarkably, the same technology that gives the world high-speed internet can also be leveraged to detect earthquakes [1, 2, 3], measure the shape and structure of the earth [4, 5, 6], and shed light on the great mysteries of the universe, from dark matter [7] to changes in the fine structure constant [8]. However, while deployed fibers provide high-speed and low-loss optical transmission, they also couple strongly to environmental fluctuations, often causing difficulty for precision applications. Vibrations, temperature variations, and other effects can change the path length that transmitted light experiences, distorting optical signals. For applications such as optical clock comparisons [9, 10, 11, 12, 13, 14], ultrastable time and frequency transfer [15, 16, 17, 18, 19, 20, 21, 22, 23], and quantum networking [24, 25, 26, 27, 28, 29, 30, 31, 32], all of which rely on the faithful transmission of low-noise signals, these environmental distortions are a major impediment.
To deliver ultrastable optical signals from one location to another, real-time measurement and active compensation of path length fluctuations are needed. Following the seminal work of Refs. [33, 34], this is typically accomplished by constructing a Michelson-type interferometer with the fiber path as one arm, thereby measuring the total phase noise accumulated by light that has double-passed the fiber. A correction signal is then generated that is proportional to half of the round-trip phase noise and is applied to the outgoing light field to compensate for the phase fluctuations it will experience along the way. While this technique has proven highly effective at suppressing fiber noise, it cannot perfectly cancel all environmental fluctuations. For the typical scheme described here, there is a widely accepted limit on how well fiber noise can be suppressed in real time, which we will refer to as the ‘standard limit’ [35, 36]. For spatially uniform and uncorrelated noise, in the low Fourier frequency limit, this is
where is the angular Fourier (or noise) frequency, is the one-way propagation delay through the fiber, is the phase noise power spectral density (PSD) of the delivered light, and is the free-running, one-way phase noise PSD of the fiber. It is worth emphasizing that this standard limit dictates how well the excess fiber noise on a delivered signal can be suppressed in real time. Some applications of fiber links are able to utilize post-processing, the colocation of the two ends of the fiber, or the use of stable laser light at the remote end [37, 38, 39]. In these cases, with stabilized lasers at both ends of the fiber, the noise can be improved by 6 dB. However, the important cases of ultrastable frequency dissemination and certain phase-stable quantum networks [29, 30] require real-time, one-way noise cancellation. To the best
Results
Theoretical Framework
All real-time ultrastable laser delivery techniques have the same basic structure: light that has double-passed the fiber path and returned to the local end is used to estimate the phase noise that light will experience when traversing a single pass. This estimate is used to generate a correction signal, which is applied to an acousto-optic modulator (AOM) or other actuator, to pre-compensate the departing signal for the one-way noise it will experience on its way to the remote end. The fundamental challenge is then to determine how best to use the double-passed light to estimate the one-way noise. To first order, the total phase noise accumulated over the round trip should be halved. (This is the standard practice.) But we show that full optimization requires examination of the temporal correlation between the round-trip and one-way signals. Specifically, the correction signal should be proportional to a temporally shifted version of the round-trip signal, and the optimal temporal shift can be determined by the correlation function. In this section, we provide a pictorial derivation of the correlation function and discuss how it is used to provide improved cancellation of the one-way noise. A rigorous derivation is provided in Methods.
We begin with the simplified picture of Fig. 1a, which depicts a fiber path with three distinct noise events that occur simultaneously at different locations along the fiber. Light is sent through this fiber to the remote end, where some of the light is delivered and some is reflected back to the local end. Noise from each of the three events is carried on the outgoing light and is seen once at the remote end of the fiber, in proper order. The local end of the fiber will experience each event twice: first in reverse order on light that was already returning at the time of the events, and then in proper order on light that was outgoing at and subsequently reflected back at the remote end. The phase noise from these events is represented in Fig. 1b, where the arrows indicate the time of arrival of the fluctuations at either the local or remote end, and are color-coded to correspond to the three events. Thus, we see that, in addition to the total round-trip noise amplitude being double that of the one-way noise, the round-trip noise is temporally delayed. Since each event appears twice on the round-trip noise with different delays, it is the center of mass that is used to describe its effective delay. For example, when the noise is at the center of the fiber (green ‘event’), the relative delay between the one-way noise (at ) and the round-trip center

Figure 1. The ‘center of mass’ picture of fiber noise. (a) A simplified distribution of noise along a fiber path with three localized, simultaneous ‘events’ along an otherwise quiet fiber. (b) Phase noise of the one-way outgoing light, , and the round-trip light, . The latter signal has the three events encoded twice each: once on light on its outward journey and once on returning light. (c) Cross-correlation of the round-trip signal and the one-way light. Each of the three events appears twice in the cross-correlation. The first instance comes from the overlap of the one-way light with the return journey of the round-trip light, resulting in a spread in the cross-correlation from to . The second instance comes from the overlap between the one-way light and the outward path of the round-trip light; these overlaps occur at a delay of for all three events, resulting in a strong peak at . Rainbow shading extrapolates from the ‘three-event’ case above to a uniform noise distribution along the fiber. The center of mass of the uniform distribution cross-correlation function is located at , requiring a compensating temporal advance relative to the delay imposed by the standard feedback scheme.
of mass (at ) is . Since the round-trip center of mass is always at and the one-way noise varies from 0 to , the relative delay between the two signals can vary between 0 and , depending on whether the noise is closer to the local or remote end of the fiber. The optimized correction signal will counteract this delay by applying a corresponding temporal advance.
To generalize the delay for arbitrary noise distributions, we next consider the cross-correlation between the one-way and the round-trip noise. The cross-correlation function is plotted in Fig. 1c, created from the plots of Fig. 1b by simply sliding the round-trip signal by a shift of relative to the one-way signal and marking the delays where same-colored arrows overlap. In this Figure, the delay variable is negative when round-trip noise lags the one-way noise. The correlation function consists of a spike at as well as a spread from to . Importantly, the area of the function for the spread from to is always the same as that of the spike at . For noise at the center of the fiber, the center of mass of the correlation function is therefore at , as expected. For noise localized at the local end, the correlation function is centered near , whereas it is centered near for noise localized at the remote end. These observations can be generalized: the spatial center of mass along the fiber corresponds to the temporal center of mass of the correlation function. That is, once the center of mass of the noise along the fiber is known, so too is the temporal shift that will optimally cancel the noise.
If we parameterize the relative temporal shift between the round-trip and one-way signals as a fraction of the one-way transit time , we can write the phase noise of the (corrected) delivered light as
where positive corresponds to a temporal advance of the round-trip light. Here, and are the one-way and round-trip phase noise signals depicted in Fig. 1b. In the case of a uniform, spatially uncorrelated fiber noise distribution, this leads to a maximum suppression of the one-way noise of approximately
in the low frequency limit (see Methods). This function is optimized for , which corresponds to the expected temporal advance of . Notably, with we recover the standard limit (equation (1)), which is four times higher than the optimal case.
Of course, a true temporal advance of the round-trip noise in real time violates causality, and can therefore only be applied in post-processing [37]. However, an effective real-time advance can be achieved for noise frequencies where . This is achieved by recognizing that the round-trip signal received at the local end, in addition to containing the round-trip noise, also contains the correction signal applied at a time prior. This, in effect, stores the round-trip noise from an earlier time. It is by properly weighting and adding delayed copies of the correction signal that we may achieve the desired effective temporal shift. (This process assumes the noise does not change much after a round trip, imposing the condition.) Notably, the standard scheme also uses the delayed correction signal to provide a phase shift (corresponding to an effective temporal advance of ) at the local end of the fiber, such that there is zero temporal shift by the time light reaches the remote end. From this point of view, the standard noise-cancellation architecture is a special case of this more general noise-correction scheme.
Experimental Implementation
To implement this new feedback scheme, we need only slightly modify the standard layout for fiber noise cancellation. As shown in Fig. 2, the basic configuration is that of a Michelson interferometer. An ultrastable laser is split into a short ( cm), free-

Figure 2. Setup for optimized fiber noise cancellation. The round-trip phase noise is detected with a Michelson interferometer. The effective temporal advance is then applied digitally by adding appropriately weighted and delayed copies of the correction signal , which is stored in memory for at least . This temporally advanced signal is then filtered, amplified, and applied as a correction signal to the feedback acousto-optic modulator (AOM). ADC, analog-to-digital converter; DAC, digital-to-analog converter; VCO, voltage-controlled oscillator; PC, polarization controller.
space reference arm and a long, fiber-optic delivery arm containing the fiber link path; upon reflection, the two beams are combined on an in-loop detector to generate a heterodyne beat note for noise cancellation. In the long arm of the interferometer, an AOM at the local end applies the noise cancellation to both the outgoing and return light. The first-order output of an AOM at the remote end is used simply to identify light that has traveled the full length of the fiber with a frequency shift, thereby minimizing the impact of spurious reflections. Part of the light at the remote end is reflected back through the fiber link path, while the rest is used for ultrastable laser delivery. In our experiments, the local and remote ends are collocated, allowing for an out-of-loop heterodyne between a portion of this delivered light and light directly from the ultrastable laser to assess the performance of each feedback scheme tested. Up to this point, the hardware description is identical to a standard real-time fiber noise cancellation scheme.
The important difference that our new scheme introduces is in the way the correction signal is processed. The heterodyne signal from the in-loop detector is mixed down in frequency, digitized, and its phase is extracted. This phase contains not only the round-trip fiber noise, but also copies of the correction signal that were applied via the feedback AOM both at the beginning of the outward journey ( ago) and at the end of the return journey (presently). To this digital phase we add two corresponding terms, each proportional to : a negative copy of the feedback signal from ago and a positive copy of the most recent feedback signal applied. This method requires retaining a rolling record of the feedback signal applied up to ago, which we accomplish using a field-programmable gate array (FPGA). Weighting the present and -delayed copies of the correction signal in this manner gives the desired temporal shift of (see Methods for more details). This weighted signal is then conditioned by a digital servo to add frequency-dependent gain and converted back to an analog signal, which is used to drive a voltage-controlled oscillator (VCO) actuating the feedback AOM. Prior knowledge of the transit time is required for this scheme, but this is easily obtainable from interferometric fringes visible on the in-loop spectrum at the local end. Notably, the only additional requirement for this scheme, relative to the standard method, is the digital signal processing; we expect that adapting an existing fiber noise cancellation setup to take advantage of this new method should be relatively low-cost and simple. It is worth noting that a fully analog implementation of this temporal advance scheme is also possible (as outlined in Methods), but requires additional hardware and cannot be as readily adapted to different values as the digital implementation.
Deployed Fiber Demonstration
As a first demonstration of this new feedback scheme, we suppressed the noise in a 5.5 km fiber deployed in an urban environment. This fiber is part of the Boulder Research and Administrative Network (BRAN) in Boulder, Colorado, and stretches from the campus of the National Institute of Standards and Technology (NIST) to the campus of the University of Colorado (CU) Boulder, as shown in Figure 3a. To facilitate measurements at both ends of the fiber link path, the local and remote ends were co-located at NIST with a loop back at the CU Boulder side. This was achieved using two separate single mode fibers in a co-packaged fiber bundle.

Figure 3. Performance in deployed fiber. (a) Overview map of the section of the BRAN network fiber in Boulder, Colorado used in the experiment. Light is looped back such that both ends of the fiber are co-located at the ‘Local End’. One-way fiber length is approximately 5.5 km. Map data: Google ©2024 (b) Optical phase noise in the standard case (), the predicted phase noise limit for the standard case, and phase noise in the optimized case assuming a uniform spatial distribution (). The noise at offset frequencies below 5 Hz is limited by out-of-loop and non-common fiber paths. The free-running fiber noise is also shown for comparison.
Using our digital feedback control scheme, we proceeded to cancel the fiber noise with both the standard delay and with . Although the exact spatial noise distribution along this deployed fiber is not known, we selected under the assumption that the noise is approximately uniform. (In fact, if the noise is nonuniform but mirrored about the midpoint due to the loop back configuration, we would still expect an optimal of .) Figure 3b shows the results of this experiment. In combination with equation (3), the free-running noise (noise on the delivered light when no feedback is engaged) defines the limit on the residual fiber noise for a given value. We show that over a broad range of noise frequencies ( Hz), the new feedback scheme out-performs the standard scheme, which itself is at its theoretical limit. It is worth noting that there is an evident trade-off in bandwidth (discussed more in Methods) as increases, as seen by the reduction in the frequency of the servo peak from 6 kHz for to 3 kHz for . However, we find that this is a worthwhile concession, since the frequency range over which we see the most improved performance also contains the highest free-running noise — meaning the reduction is
greatest where it is needed most. This improvement can also be appreciated by examining the integrated phase noise. For , integrating the noise from 10 kHz to 1 Hz (1 kHz to 1 Hz) results in 153.7 mrad (36.2 mrad), whereas for , integration over the same frequency range gives 100.6 mrad (25.1 mrad). We also implemented the fully analog scheme over this fiber link, with both phase noise and modified Allan deviation results given in Methods.
Localized Noise Demonstration
Whereas the previous section demonstrated the utility of this scheme in a deployed fiber, in this section we verify that the optimal temporal advance coincides with the spatial location of the noise, and we exploit this property to suppress localized noise more than 10 dB beyond the standard limit. As discussed previously, this new feedback scheme is adaptable to any noise distribution simply by tuning . One class of distributions we can consider is that in which the noise is predominantly localized in one region of the fiber. In this case, it is possible to achieve far better suppression than the standard technique provides. (In principle, perfect cancellation is possible if the noise is entirely in one location. See Methods for more details.) Such a localized noise distribution may arise in an urban link containing an aerial section or in an undersea fiber passing over a geologically active site, for example.
To model a range of localized noise distributions, we built a lab-based fiber testbed to take the place of the ‘Fiber Link Path’ in the Fig. 2 diagram. This testbed is an approximately 8.8 km fiber comprised of three long and relatively quiet sections of spooled fiber, plus one short and relatively noisy fiber section that is routed around the lab. In this way, virtually all of the noise is localized within this short section, equivalent to approximately 1% of the overall fiber length. By changing the order of the component fibers, we can shift where this localized noise is, effectively engineering various noise distributions within our fiber link path. We parameterize the fractional location of the noisy section along the total length in terms of the dimensionless quantity , where () describes a fiber with all of the noise at the local (remote) end.
With this testbed system, we engineered six different fiber noise distributions with the noise localized at different points along the 8.8 km fiber length. Optimal suppression of a localized noise distribution is achieved when . Therefore, for each of these distributions, we performed noise cancellation using and measured the out-of-loop phase noise of the light delivered to the remote end in each case. Figure 4a-Figure 4c present a selection of these phase noise results focusing on the relative performance of and . When all of the noise is at the local end of the fiber, gives the optimal performance, since is approximately zero; however, as the noise is moved further along the fiber (increasing ), the performance of the standard case degrades while the case continues to provide optimal noise suppression. Note that the finite length of the noisy section of fiber has led us to approximate the values of the first and last fiber configurations as 0.01 and 0.99, respectively.

Figure 4. Optimizing performance for spatially localized noise. (a-c) Optical phase noise for fiber configurations where the noise is predominantly localized at the fractional location along the fiber path. Data is plotted for the standard () case and for the optimized () case. (d) Comparison of performance for different values as a function of noise location. Phase noise is integrated from 10–100 Hz (shaded region in (a-c)) for the cases of , where gives optimal noise suppression. Since the noisy section of fiber has finite length, values of 0 and 1 cannot be exactly realized. Based on the fiber lengths used, the first and last configurations are taken to have actual values of 0.01 and 0.99. Theory curves are derived from (43).
To consolidate these results and the results from the other combinations of and we explored, we compare the integrated phase noise over 10-100 Hz for each case, shown in Figure 4d. This integration range was chosen because it is the peak of the free-running noise PSD (making suppression here most critical) and it is also where the fundamental limits are most consistently the current limitation on performance (allowing for meaningful comparison with theoretical calculations). The theory curves in Figure 4d are based on equation (43) in Methods. For every localized noise distribution tested, provides the best noise suppression in this frequency range. In fact, for the cases with the noise in the second half of the fiber (i.e. ), the optimized feedback enhances noise suppression by approximately 13 dB compared to the standard case, with further suppression likely possible. With such large suppression of the localized noise, the small amount of noise from the other ‘quiet’ fiber sections will ultimately dominate, once the localized noise is suppressed beyond that lower level. The delay is only ever perfectly optimized for a single spatial location.
Discussion
We have developed a new, intuitive way to conceptualize the correlations between round-trip and one-way noise along a path. By adding a temporal shift through the variable , we introduce a new parameter by which to optimally exploit these correlations in real time. The two experiments presented demonstrate the utility and the potential of this ‘center of mass’ framework for optimally suppressing fiber noise for ultrastable laser delivery. The localized noise experiment supports the assertion that the spatial center of mass of noise along the fiber corresponds to the temporal center of mass of the correlation function. Put differently, localized fluctuations are optimally suppressed when , which we show can lead to over 10 dB enhancement in noise suppression relative to the standard limit. The deployed fiber experiment leverages this new understanding to realize a significant improvement in noise suppression in a realistic noise environment, achieving approximately 6 dB enhancement in noise cancellation relative to the standard () case. It is interesting to note that since the center of mass of the correlation function must always lie between and 0, any will provide enhanced suppression for virtually all possible fiber noise distributions. (Only in the case of noise localized entirely at the very beginning of the fiber path would the standard provide optimal suppression.) That is to say, any temporal advance () will provide at least some improvement over the standard scheme for nearly all situations. The optimal temporal advance, of course, corresponds to the noise distribution center of mass. Since the signal processing hardware is the only significant change to the traditional fiber noise cancellation architecture, we expect that the (at least) fourfold performance enhancement presented here can be readily achieved with only minor modification to existing systems.
Using digital signal processing, can be easily and quickly tuned to any value, making this feedback system extremely adaptable. To optimize , the location of the center of mass of the noise distribution is required; this can come from prior knowledge of the fiber path (e.g., the known location of an aerial section), or can be measured with, for example, distributed acoustic sensing (DAS) [40, 41]. With DAS, real-time data on the fiber noise could allow dynamic optimization of the noise suppression. This could be of particular advantage in stabilizing paths with noise distributions that change based on tides, rush hour, weather, or any number of external variables. Finally, we note that the framework and feedback schemes developed here need not be limited to optical fiber systems. While optical fiber links represent a ubiquitous and important platform for noise cancellation, these ideas can also be directly applied to free-space optical and microwave systems.
Supplementary information
Not applicable.
Acknowledgments
We thank Scott Diddams, Andrew Ludlow, and their respective groups for equipment and access to the BRAN fiber network. We thank Roger Brown, Fabrizio Giorgetta, and Krister Shalm for their helpful comments on the manuscript.
Declarations
Funding: This research was supported by an appointment to the Intelligence Community Postdoctoral Research Fellowship Program at the National Institute of Standards and Technology administered by Oak Ridge Institute for Science and Education (ORISE) through an interagency agreement between the U.S. Department of Energy and the Office of the Director of National Intelligence (ODNI), and by NIST.
Conflict of interest/Competing interests: None.
Ethics approval and consent to participate: Not applicable.
Data availability: Data available upon reasonable request.
Materials availability: Not applicable.
Code availability: Not applicable.
Methods
Digital Servo
In this section, we provide greater detail on the digital servo architecture, based on the interferometer setup shown in Fig. 2. With an acousto-optic modulator (AOM) driven at frequency , the heterodyne beat frequency on the in-loop detector is . The phase of this beat signal may be expressed as
where is the round-trip fiber noise, is the correction signal applied to the return light, is the correction signal applied to the out-going light after a double-pass through the fiber, is the laser phase noise from the interferometer’s reference arm, and is the laser phase noise after double-pass through the fiber link. Going forward, we make the usual assumption that the laser source has sufficiently low noise that . Satisfying this requirement necessitates a laser with phase noise that is more than 10 dB lower than the free-running phase noise of the fiber link. Most cavity-stabilized lasers meet this requirement for offset frequencies within the lock bandwidth of the fiber noise cancellation.
At the front end of the servo, the heterodyne signal is digitized and the phase is extracted. To achieve the proper weighting of the control signal, and are added to digitally. Note that is chosen by the user and must be known. After signal weighting, the phase is
In the Fourier domain this becomes
An implicit assumption here is that the temporal coherence of is greater than the fiber round-trip time, such that the correction applied at can be linearly extrapolated from the signal at time . This holds for noise frequencies where , in which case it is equivalent to the first-order approximation of .
This signal is then conditioned by the digital phase-lock loop (PLL) to provide a frequency-dependent gain, and applied to a voltage-controlled oscillator (VCO) to drive the AOM. The VCO’s output frequency is proportional to the control signal input. In terms of the phase, the VCO’s output is the derivative of , such that the Fourier domain equation of the PLL is written as
where is the open-loop gain of the PLL. Solving for yields
In the high gain limit where , this reduces to
which is the desired form. The standard feedback configuration with can be considered a special case of our more general description where can vary, typically between 0 and 1.
In our experiment, we do not directly measure and scale by . Rather, we use the digital control signal used to drive the VCO. The control signal was scaled by the volts-to-hertz conversion gain of the VCO, then integrated to retrieve the phase. For our implementation, was entered from the user-interface of the computer-controlled digital PLL. Memory of the correction signal for a time delay of is required to apply the term. In our system, the delay memory, control signal scaling and integration, and digital PLL are all implemented with an FPGA. Lastly, we note that the optimal correction applied along the outgoing path is sensitive to errors in the estimated round-trip time. This issue is less critical at low frequencies, because an error would introduce an error of only into the subtracted phase.
Generalized Analog Hardware Implementation
Analog hardware implementations of the optimized servo architecture are possible, though they are much more cumbersome than the digital servo approach. These implementations require a physical separation of the noise-corrected signal from the noise measurement-and-correction AOM used in the standard scheme. This increases the number of AOMs needed at the local end of the fiber. Also, the drive frequencies of these AOMs are determined by the desired value of , necessitating a change in hardware whenever is appreciably changed.
The general scheme for an arbitrary value of is shown in Extended Data Fig. 1. In addition to the double-pass AOM at the fiber input (denoted AOM1), a double-pass AOM is included in the interferometer reference arm (AOM2). Signal weighting is achieved by driving AOM1 and AOM2 with correction signals scaled by and , respectively. This can be achieved, for example, by using a single VCO whose output is split and routed to different frequency scalers. With this arrangement, the phase of signal on the detector from the reference arm becomes , and the signal from the fiber arm is . The resulting phase of the heterodyne beat on the in-loop detector then matches that of equation (2).

Figure 1 (PDF p. 14). Setup for analog hardware implementation of the general case. AOM, acousto-optic modulator; BPF, bandpass filter; VCO, voltage-controlled oscillator; PC, polarization controller.
Neither AOM1 nor AOM2 receives the properly scaled signal for noise cancellation. Therefore, a third AOM (AOM3) is required for delivery of the noise-corrected signal. This signal must be coupled into the fiber after AOM1, adding an uncompensated light path. A fourth AOM is included at the remote end simply to distinguish light that made a full round trip, thereby minimizing the influence of stray reflections. When heterodyning against another laser at the remote end, the different frequency shifts imparted by the separate AOMs provide for straightforward selection of the correct signal. In other use cases, multiple cores of a multicore fiber (MCF) may be used as a separate channel for light delivery. The use of MCF relies on the high degree of noise correlation between the core used to measure the noise and the core used for signal delivery. This has been shown to work for high-fidelity frequency transfer using multiple cores [18]. A small benefit of the hardware scheme is that prior knowledge of is not required.
With , the number of required AOMs can be reduced from three to two, since a double-pass through AOM2 provides the correct unity scaling of . This allows AOM2 to also serve as the delivery AOM. This experimental setup is shown in Extended Data Fig. 2. We have successfully realized this analog hardware implementation

Figure 2 (PDF p. 15). Setup for hardware implementation of case. AOM, acousto-optic modulator; BPF, bandpass filter; VCO, voltage-controlled oscillator; PC, polarization controller. The two feedback AOM signals are obtained from one VCO by driving the 55 MHz VCO at high power and filtering out the fundamental tone and third harmonic.
tation of to stabilize the BRAN deployed fiber path discussed in the main text. Phase noise and modified Allan deviation results from this experiment are shown in Extended Data Figure 3. For this demonstration, the reference-arm AOM operated at a frequency of 55 MHz, whereas the in-line AOM operated at 165 MHz, leading to the delivered light having a frequency shift of 110 MHz from double-passing the reference arm AOM. A third AOM was again used in this demonstration at the remote end to distinguish light that made a full round trip.

Extended Data Figure 3. Performance of analog implementation in deployed fiber. (a) Optical phase noise in the standard case () and the optimized case () for the BRAN deployed fiber. Noise at offset frequencies below 8 Hz is dominated by out of loop and non-common fiber paths. (b) Modified Allan deviation derived from the phase noise traces in (a).
Lock Bandwidth
As shown in Figure 3-Figure 4, the locking bandwidth decreases with increasing . In this section, we show how this behavior is expected from the dynamics of the feedback loop. Using standard phase-lock loop analysis [42], the in-loop phase may be represented in the frequency domain as
where is given by
The feedback loop will oscillate when . This can happen when the phase of is equal to . Assuming minimal phase delay in , finding the loop oscillation frequency reduces to finding the frequency where
Defining , the equation that gives may be expressed as
When (standard case), this reduces to . The frequency for which this phase is equal to is therefore , as expected [36]. However, when , the phase becomes , and the oscillation frequency is reduced to . The oscillation frequency for intermediate values of will fall between these extremes. Numerical solutions of equation (13) are shown in Extended Data Figure 4.

Extended Data Figure 4. Feedback bandwidth as a function of . The normalized bandwidth is defined as where the feedback phase is delayed by radians, and is derived by solving equation (13).
Optimized Noise Suppression: Uniform Distribution
In this section, we show is the optimal value for noise that is uniformly distributed along the fiber link. A similar conclusion is reached in Ref. [38] where it was shown that a temporal advance of the roundtrip noise can be used in post-processing to reduce the fiber noise.
At the remote end of the noise-canceled fiber link, the phase of the signal may be represented as
where is the one-way noise added by the fiber. To focus on the link noise, we drop the term in the analysis below. With this term removed, the Fourier transform of equation (14) is
Using the generalized control signal in the high gain limit of equation (9), the noise at the remote end becomes
This simplifies to
Further analysis requires expressions for and . We follow the typical representation of the one-way noise [36], where phase fluctuations imparted at position (measured from the local end) are represented as , where is the speed of light in the fiber. These fluctuations are integrated over , the length of the fiber, yielding
Taking the Fourier transform yields
The round-trip noise consists of phase perturbations on the return light as well as the out-going light after a double-pass, and is likewise integrated over the length of the fiber, yielding
with Fourier transform
Inserting equations (19) and (21) into equation (17) yields
where
The power spectral density (PSD) of the noise at the remote end is given by . This is expressed as
Note that * here denotes complex conjugation. We now specify the spatial correlation of the fiber noise as
where is a dimensionless, smooth function that goes to zero sufficiently fast as , and . We further define the characteristic length of the spatial correlations as
If we now assume , the function can be replaced by a Dirac delta function,
Equation (24) then simplifies to
With spatially uniform noise, this becomes
where we have used the fact that the PSD of the free-running one-way fiber noise is given by
Straightforward integration of equation (29) yields
Expanding to lowest order in , we obtain
which is presented as equation (3) in the main text. The noise is minimized for ,
while yields the standard result,
Equation (33) has previously been derived for two-way transfer and systems utilizing post-processing [37, 38], both of which require ultrastable lasers at each end of the fiber. Here we show that the limit imposed by equation (33) is achievable with one-way frequency transfer in real time.
Effective Temporal Shift
We now examine the effective temporal shift of the round-trip noise at the remote end of the fiber when feedback is applied. For , equation (17) can be expanded as
In the time domain this is
Note that since the correction signal originates at the local end it is delayed by as it propagates to the remote end of the fiber. The temporal advance of the round-trip noise at the local end is therefore (this can be derived directly from equation (9) as well). Thus, even for the standard architecture, the local-end correction signal exhibits an effective temporal advance.
Noise Center of Mass
The main text provides a pictorial illustration of the noise correlation function between the opposite ends of the fiber. In this section, we provide a more rigorous derivation. Using a normalization for a finite-time Fourier transform over a large time window , the cross-correlation between the phase noise of the out-going and the round-trip fields is the inverse Fourier transform of . Using equations (16) and (18), this becomes
After integration, the cross-spectrum becomes
Inverse Fourier transform of the above produces the convolution between
and the inverse Fourier transform of , which is the temporal correlation function of the phase perturbations. Here, is a rectangular function of unit amplitude and unit width. Assuming that the fiber perturbations are much faster than , their temporal correlation function can be approximated by a delta function. As a result, we obtain
The function , sketched in Fig. 1c of the main text, represents the effect of a delay on the temporal overlap between phase noise and . The two terms of the correlation function have the same area, with the first term centered on and the second on . The total temporal correlation center of mass is therefore centered on . The minus sign indicates a delay of relative to . The optimal correction signal therefore requires a temporal advance of . In a system where the noise is band limited, the inverse Fourier transform of is not a delta function. However, the noise center of mass does not change.
Optimized Noise Suppression: Localized Fluctuations
In this section, we derive the expected noise suppression for the situation where the phase noise fluctuations are dominated by a specific section of the fiber. Examples include fibers with aerial sections, near the shore of undersea cables, or near geologically active sites.
For noise located at measured from the local end, we first define as the fractional length to the noise location, or . We limit the spatial extent of the fiber noise over a length by defining as
If we then assume is constant over , the PSD of the noise at the remote end becomes
Note that the noise is fully suppressed for . (Recall that this applies in the high gain limit in steady state.) Expansion to second order gives
The lowest noise is achieved when , where the PSD is reduced to
The maximum residual noise is when is in the center of the fiber link, where the noise is . In practice, whenever , the localized suppression is so great that the noise from other sections of the fiber will usually dominate at low frequencies.
With digital signal processing, it is possible to further improve this limit for localized noise by adding another term to the heterodyne phase with delay of . In this case, it is theoretically possible to achieve perfect cancellation of noise for any value of , so long as the fiber noise is perfectly localized. The phase of the weighted signal of equation (2) then becomes
We now set and , yielding
In the high gain limit, the frequency domain representation of the correction signal is
The noise-corrected PSD at the remote end of the fiber is therefore
which is zero for all values of .
It is interesting to note that the average of the term proportional to in equation (43) with a uniform distribution of the position of the noise localized at is
producing once again that the optimum value of is one half. The impossibility of suppressing the term proportional to for uncorrelated fluctuations can be interpreted as a consequence of the fact that the value of that produces optimal suppression depends on the position of the perturbations along the fiber. For spatially distributed noise, it is impossible to pick one value of that perfectly suppresses noise at multiple locations simultaneously. Therefore, the best compromise comes from choosing a that corresponds to the noise distribution center of mass.
Spatially Correlated Noise
The phase noise may be correlated along the fiber path, as is expected in the use of spooled fiber. Here we show that, when the fluctuations are fully correlated in space, the use of the proposed technique with is much more effective than the standard architecture. In this case, there is no spatial dependence to the phase fluctuations , hence equation (22) can be explicitly integrated, yielding
where we have used . This holds because of the spatial coherence of the fluctuations. Taking the magnitude squared and expanding for small we obtain
For , the result is nearly the same as it is for uniformly distributed, uncorrelated noise, with only the prefactor replaced with . For , however, the term falls away, and we obtain the qualitatively different result of
This can lead to drastically lower noise. For example, for (10 km of fiber), results in an extra 36 dB of suppression of the fiber noise at 100 Hz offset compared to .
Optimality of the Scheme
In equation (2), the control signal is written as the round-trip noise scaled by and temporally shifted by . The most general form of the control signal is a linear transformation of the round-trip noise, such that the noise at the remote end of the fiber may be written as
where is the impulse response of the control system after propagating to the remote end of the fiber, and denotes convolution. The most general form of the control kernel is therefore
where is the Fourier transform of . To ensure full noise cancellation at , the condition must be satisfied. In this case, a useful form of is
where is an arbitrary, generally complex, function of its argument. In this section, we show our scheme produces the value of this function that results in the lowest noise. We restrict our discussion to the case of uncorrelated, uniformly distributed noise. For localized noise, the noise can in principle be fully suppressed. For spatially correlated noise, at small the dependence renders this noise negligible. Performing the integral in equation (22) with the use of equations (54) and (55) and expanding to second order gives
Here we have used the fact that only affects the second-order term. A nonzero real part of can only increase the noise, so that we may assume as purely imaginary: . With this choice, equation (56) becomes
Thus, the minimum noise is achieved for and is equal to . It is impossible to obtain, for spatially uncorrelated perturbations, a faster decay of the noise spectrum at low frequency for any choice of .
References
- [1]Marrа, G., Clivati, C., Luckett, R., Tampellini, A., Kronjäger, J., Wright, L., Mura, A., Levi, F., Robinson, S., Xuereb, A., et al.: Ultrastable laser interferometry for earthquake detection with terrestrial and submarine cables. Science 361(6401), 486–490 (2018)DOI
- [2]Marra, G., Fairweather, D., Kamalov, V., Gaynor, P., Cantono, M., Mulholland, S., Baptie, B., Castellanos, J., Vagenas, G., Gaudron, J.-O., et al.: Optical interferometry–based array of seafloor environmental sensors using a transoceanic submarine cable. Science 376 (6595), 874–879 (2022)
- [3]Lindsey, N.J., Martin, E.R.: Fiber-optic seismology. Annual Review of Earth and Planetary Sciences 49(1), 309–336 (2021)
- [4]Mehlstäubler, T.E., Grosche, G., Lisdat, C., Schmidt, P.O., Denker, H.: Atomic clocks for geodesy. Reports on Progress in Physics 81(6), 064401 (2018)
- [5]Clivati, C., Calonico, D., Frittelli, M., Mura, A., Levi, F.: A fiber link for the remote comparison of optical clocks and geodesy experiments. In: 2015 Joint Conference of the IEEE International Frequency Control Symposium & the European Frequency and Time Forum, pp. 579–582 (2015). IEEE
- [6]Grotti, J., Koller, S., Vogt, S., Häfner, S., Sterr, U., Lisdat, C., Denker, H., Voigt, C., Timmen, L., Rolland, A., et al.: Geodesy and metrology with a transportable optical clock. Nature Physics 14(5), 437–441 (2018)DOI
- [7]Manley, J., Stump, R., Petery, R., Singh, S.: Searching for scalar ultralight dark matter via refractive index changes in fibers. Physical Review D 108(7), 075008 (2023)DOI
- [8]Roberts, B.M., Delva, P., Al-Masoudi, A., Amy-Klein, A., Baerentsen, C., Baynham, C., Benkler, E., Bilicki, S., Bize, S., Bowden, W., et al.: Search for transient variations of the fine structure constant and dark matter using fiber-linked optical atomic clocks. New Journal of Physics 22(9), 093010 (2020)DOI
- [9]BACON Collaboration, et al.: Frequency ratio measurements at 18-digit accuracy using an optical clock network. Nature 591(7851), 564–569 (2021)
- [10]Rosenband, T., Hume, D., Schmidt, P., Chou, C.-W., Brusch, A., Lorini, L., Oskay, W., Drullinger, R.E., Fortier, T.M., Stalnaker, J.E., et al.: Frequency ratio of Al⁺ and Hg⁺ single-ion optical clocks; metrology at the 17th decimal place. Science 319(5871), 1808–1812 (2008)DOI
- [11]Nemitz, N., Ohkubo, T., Takamoto, M., Ushijima, I., Das, M., Ohmae, N., Katori, H.: Frequency ratio of Yb and Sr clocks with 5 × 10⁻¹⁷ uncertainty at 150 seconds averaging time. Nature Photonics 10(4), 258–261 (2016)DOI
- [12]Dörscher, S., Huntemann, N., Schwarz, R., Lange, R., Benkler, E., Lipphardt, B., Sterr, U., Peik, E., Lisdat, C.: Optical frequency ratio of a ¹⁷¹Yb⁺ single-ion clock and a ⁸⁷Sr lattice clock. Metrologia 58(1), 015005 (2021)DOI
- [13]Kim, M.E., McGrew, W.F., Nardelli, N.V., Clements, E.R., Hassan, Y.S., Zhang, X., Valencia, J.L., Leopardi, H., Hume, D.B., Fortier, T.M., et al.: Improved interspecies optical clock comparisons through differential spectroscopy. Nature Physics 19(1), 25–29 (2023)
- [14]Lindvall, T., Pizzocaro, M., Godun, R.M., Abrall, M., Akamatsu, D., Amy-Klein, A., Benkler, E., Bhatt, N.M., Calonico, D., Cantin, E., et al.: Coordinated international comparisons between optical clocks connected via fiber and satellite links. Optica 12(6), 843–852 (2025)DOI
- [15]Caldwell, E.D., Triano, T.M., Sinclair, L.C.: High-precision optical time and frequency transfer. Advances in Optics and Photonics 17(2), 375–440 (2025)
- [16]Clivati, C., Aiello, R., Bianco, G., Bortolotti, C., De Natale, P., Di Sarno, V., Maddaloni, P., Maccaferri, G., Mura, A., Negusini, M., et al.: Common-clock very long baseline interferometry using a coherent optical fiber link. Optica 7(8), 1031–1037 (2020)
- [17]Li, W., Chen, Y., Dai, H., Wang, F., Wang, B.: Phase noise control of the fiber-based frequency dissemination system. IEEE Transactions on Instrumentation and Measurement 73, 1–7 (2024)DOI
- [18]Hooghooghi, N., Mazur, M., Fontaine, N., Liu, Y., Lee, D., McLemore, C., Nakamura, T., Hayashi, T., Sciullo, G.D., Shaji, D., Mecozzi, A., Antonelli, C., Quinlan, F.: Ultrastable optical frequency transfer and attosecond timing in deployed multicore fiber. Optica 12(6), 894–899 (2025) https://doi.org/10.1364/OPTICA.558821
- [19]Schioppo, M., Kronjaeger, J., Silva, A., Ilieva, R., Paterson, J., Baynham, C., Bowden, W., Hill, I., Hobson, R., Vianello, A., et al.: Comparing ultrastable lasers at 7 × 10−17 fractional frequency instability through a 2220 km optical fibre network. Nature communications 13(1), 212 (2022)DOI
- [20]Cantin, E., Tønnes, M., Le Targat, R., Amy-Klein, A., Lopez, O., Pottie, P.-E.: An accurate and robust metrological network for coherent optical frequency dissemination. New Journal of Physics 23(5), 053027 (2021)DOI
- [21]Clivati, C., Pizzocaro, M., Bertacco, E., Condio, S., Costanzo, G., Donadello, S., Goti, I., Gozzelino, M., Levi, F., Mura, A., et al.: Coherent optical-fiber link across Italy and France. Physical Review Applied 18(5), 054009 (2022)DOI
- [22]Cizek, M., Pravdova, L., Minh Pham, T., Lesundak, A., Hrაბina, J., Lazar, J., Pronebner, T., Aeikens, E., Premper, J., Havlis, O., et al.: Coherent fibre link for synchronization of delocalized atomic clocks. Optics Express 30(4), 5450–5464 (2022)DOI
- [23]Bercy, A., Stefani, F., Lopez, O., Chardonnet, C., Pottie, P.-E., Amy-Klein, A.: Two-way optical frequency comparisons at 5 × 10−21 relative stability over 100-km telecommunication network fibers. Phys. Rev. A 90, 061802 (2014) https://doi.org/10.1103/PhysRevA.90.061802
- [24]Nichol, B.C., Srinivas, R., Nadlinger, D., Drmota, P., Main, D., Araneda, G., Ballance, C., Lucas, D.: An elementary quantum network of entangled optical atomic clocks. Nature 609(7928), 689–694 (2022)DOI
- [25]Nakamura, T., Hoghhooghi, N., Fontaine, N., Hayashi, T., Nagashima, T., Nardelli, N.V., Reddy, D.V., Stevens, M.J., Fortier, T., Shalm, L.K., et al.: Sub-femtosecond stabilization of multicore fiber for high-fidelity quantum networking at 100% duty cycle. arXiv preprint arXiv:2509.17989 (2025)DOI
- [26]Bersin, E., Grein, M., Sutula, M., Murphy, R., Huan, Y.Q., Stevens, M., Suleymanzade, A., Lee, C., Riedinger, R., Starling, D.J., et al.: Development of a Boston-area 50-km fiber quantum network testbed. Physical Review Applied 21(1), 014024 (2024)DOI
- [27]Kucera, S., Haen, C., Arenskötter, E., Bauer, T., Meiers, J., Schäfer, M., Boland, R., Yahyapour, M., Lessing, M., Holzwarth, R., et al.: Demonstration of quantum network protocols over a 14-km urban fiber link. npj Quantum Information 10(1), 88 (2024)arxiv.org/abs/2404.04958
- [28]Chung, J., Kanter, G., Lauk, N., Valivarthi, R., Wu, W., Ceballos, R.R., Pena, C., Sinclair, N., Thomas, J., Xie, S., et al.: Illinois Express Quantum Network (IEQNET): metropolitan-scale experimental quantum networking over deployed optical fiber. In: Quantum Information Science, Sensing, and Computation XIII, vol. 11726, p. 1172602 (2021). SPIEarxiv.org/abs/2104.04629
- [29]Duan, L.-M., Lukin, M.D., Cirac, J.I., Zoller, P.: Long-distance quantum communication with atomic ensembles and linear optics. Nature 414(6862), 413–418 (2001)
- [30]Nardelli, N.V., Reddy, D.V., Grayson, M., Sorensen, D., Stevens, M.J., Mazurek, M.D., Shalm, L.K., Fortier, T.M.: Phase-stable optical fiber links for quantum network protocols. arXiv preprint arXiv:2510.16230 (2025)arxiv.org/abs/2510.16230
- [31]Liu, J.-L., Luo, X.-Y., Yu, Y., Wang, C.-Y., Wang, B., Hu, Y., Li, J., Zheng, M.-Y., Yao, B., Yan, Z., et al.: Creation of memory–memory entanglement in a metropolitan quantum network. Nature 629(8012), 579–585 (2024)DOI
- [32]Stolk, A.J., Enden, K.L., Slater, M.-C., Raa-Derckx, I., Botma, P., Van Rantwijk, J., Biemond, J.B., Hagen, R.A., Herfst, R.W., Koek, W.D., et al.: Metropolitan-scale heralded entanglement of solid-state qubits. Science advances 10(44), 6442 (2024)DOI
- [33]Ma, L.-S., Jungner, P., Ye, J., Hall, J.L.: Delivering the same optical frequency at two places: accurate cancellation of phase noise introduced by an optical fiber or other time-varying path. Opt. Lett. 19(21), 1777–1779 (1994) https://doi.org/10.1364/OL.19.001777
- [34]Bergquist, J., Itano, W., Wineland, D.: Laser stabilization to a single ion. North Holland, Amsterdam (1994)
- [35]Newbury, N.R., Williams, P.A., Swann, W.C.: Coherent transfer of an optical carrier over 251 km. Optics letters 32(21), 3056–3058 (2007)
- [36]Williams, P.A., Swann, W.C., Newbury, N.R.: High-stability transfer of an optical frequency over long fiber-optic links. J. Opt. Soc. Am. B 25(8), 1284–1293 (2008) https://doi.org/10.1364/JOSAB.25.001284
- [37]Stefani, F., Lopez, O., Bercy, A., Lee, W.-K., Chardonnet, C., Santarelli, G., Pottie, P.-E., Amy-Klein, A.: Tackling the limits of optical fiber links. J. Opt. Soc. Am. B 32(5), 787–797 (2015) https://doi.org/10.1364/JOSAB.32.000787
- [38]Calosso, C., Bertacco, E., Calonico, D., Clivati, C., Costanzo, G.A., Frittelli, M., Levi, F., Micalizio, S., Mura, A., Godone, A.: Doppler-stabilized fiber link with 6 dB noise improvement below the classical limit. Optics Letters 40(2), 131–134 (2015)DOI
- [39]Jabir, M., Ahn, D., Annafianto, N.F.R., Burenkov, I., Battou, A., Polyakov, S.: Enabling phase stabilization of quantum networks via displacement-enhanced photon counting. Optica 12(5), 570–576 (2025)
- [40]Lindsey, N.J., Dawe, T.C., Ajo-Franklin, J.B.: Illuminating seafloor faults and ocean dynamics with dark fiber distributed acoustic sensing. Science 366(6469), 1103–1107 (2019)DOI
- [41]Liu, J., Li, H., Noh, H.Y., Santi, P., Biondi, B., Ratti, C.: Urban sensing using existing fiber-optic networks. Nature Communications 16(1), 3091 (2025)DOI
- [42]Oppenheim, A.V., Schafer, R.W., Buck, J.R.: Discrete-time Signal Processing. Prentice-Hall, Inc., Upper Saddle River, NJ, USA (1999)