PhaseLock: Recurring Phase Control Motif

• PhaseLock is a design motif that maintains phase coherence by using a reference phase, error extraction, and corrective feedback across diverse physical systems.
• It employs a two-stage strategy that separates coarse tracking from high-precision estimation, ensuring robust performance despite phase ambiguities.
• Applications include atomic clocks, digital phase-locked platforms, submillimeter astronomical oscillators, sequential beam combining, and diffusion-model motion preservation.

PhaseLock is a label used in several research areas for architectures that preserve, recover, or impose phase coherence in systems whose free evolution would otherwise drift, decohere, or become ambiguous. In the cited literature, the name is attached to an atomic-ensemble phase-tracking protocol for clocks and interferometers, a digital phase-locked-loop platform for optical metrology, a fielded Gunn-oscillator stabilization system in submillimeter astronomy, a sequential coherent beam-combination scheme, and a training-free image-to-video diffusion framework that preserves early motion priors during denoising (Kohlhaas et al., 2015, Tourigny-Plante et al., 2018, Hunter et al., 2011, Klenke et al., 2021, Han et al., 4 Jun 2026). This suggests that “PhaseLock” functions less as a single standardized method than as a recurring design motif for phase-sensitive control.

1. Scope and common structure

The named PhaseLock systems in the literature operate in physically different state spaces, but they share a common control logic. An internal or external phase reference is identified, an observable correlated with phase error is extracted, and a corrective action is applied before the phase leaves a regime in which inversion, constructive interference, or physically meaningful evolution can still be maintained. In atomic clocks, that observable is a weak nondestructive readout of $J_z$; in optical metrology hardware it is an FPGA-based digital phase estimate from I/Q demodulation; in filled-aperture beam combination it is a rejection-port lock-in error signal; and in image-to-video diffusion it is a latent inter-frame delta used as a proxy for inter-frame phase differences (Kohlhaas et al., 2015, Tourigny-Plante et al., 2018, Klenke et al., 2021, Han et al., 4 Jun 2026).

Context	PhaseLock form	Reported feature
Atomic ensemble	Repeated weak Ramsey interrogations with LO phase feedback	Extends the unambiguous interval beyond a single Ramsey fringe
Optical metrology	Open FPGA-based digital PLL on Red Pitaya	$565~\text{ns}$ total latency with VCO path
Submillimeter astronomy	Digital PLL for Gunn oscillators	Reacquisition target within $2~\mu\text{s}$ after Walsh switching
Beam combining	Sequential phase locking with reused dither frequencies	Comparable convergence with lower actuator-frequency requirement
Image-to-video diffusion	Training-free latent-delta guidance	Average Physics-IQ improvement of $6.2$ points

A recurrent feature is the separation of coarse tracking from fine estimation. Intermediate actions are often allowed to be approximate so long as they keep the system in a recoverable regime, while a later stage performs high-precision inference or high-fidelity synthesis. That pattern is explicit in the atomic protocol, where intermediate weak measurements only need sufficient precision to keep the phase inside $[-\pi/2,\pi/2]$, and in the diffusion framework, where a 2-step prior is preserved while later denoising remains available for appearance refinement (Kohlhaas et al., 2015, Han et al., 4 Jun 2026).

2. Atomic-ensemble PhaseLock

In the atomic-clock implementation, the basic limitation is the Ramsey readout itself. The accumulated relative phase $\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$ is not measured directly; instead, the experiment reads a population imbalance $J_z$ after a projection pulse, with signal proportional to a sinusoid such as $\sin\varphi$ or $\cos\varphi$. The inversion is therefore unique only in the interval $[-\pi/2,\pi/2]$; outside it, the same population difference can correspond to multiple phases, and interrogation time must normally be shortened to avoid ambiguity (Kohlhaas et al., 2015).

The protocol addresses this by converting the atomic coherent spin state from a one-shot phase sensor into a continuously phase-tracked reference. A balanced superposition of the two $565~\text{ns}$0Rb clock states $565~\text{ns}$1 and $565~\text{ns}$2, separated by $565~\text{ns}$3 GHz, is prepared in an optical dipole trap. The experiment then alternates a Ramsey-type projection pulse, a weak nondestructive dispersive measurement of $565~\text{ns}$4, and an opposite “reintroduction” $565~\text{ns}$5 pulse that returns the collective spin to the equatorial plane so that coherence is preserved for the next interval. The local-oscillator phase is updated after the $565~\text{ns}$6-th interrogation according to

$565~\text{ns}$7

typically with $565~\text{ns}$8, and after the final interrogation the total accumulated drift is reconstructed as

$565~\text{ns}$9

A conventional frequency correction is then applied over the full extended interrogation time $2~\mu\text{s}$0 (Kohlhaas et al., 2015).

The reported apparatus uses about $2~\mu\text{s}$1 laser-cooled $2~\mu\text{s}$2Rb atoms at $2~\mu\text{s}$3. Probe destructivity is limited to about $2~\mu\text{s}$4 of the ensemble coherence per readout, the $2~\mu\text{s}$5 microwave pulses are $2~\mu\text{s}$6 long, and the feedback delay is about $2~\mu\text{s}$7. In demonstrations with an intentional $2~\mu\text{s}$8 Hz LO detuning, the uncontrolled phase leaves the inversion region after about $2~\mu\text{s}$9 ms, whereas with phase lock active the trajectory becomes sawtooth-like and remains within the unambiguous interval for the full $6.2$0 ms shown. In an imposed phase-jump test with alternating $6.2$1 steps, the controller detects and cancels the jumps, keeping the corrected phase close to zero. In an atom-clock proof-of-concept, a standard Ramsey clock with $6.2$2 ms was compared with a clock using $6.2$3 correlated interrogations on the same ensemble, giving $6.2$4 ms; the phase-lock clock improved instability by a factor $6.2$5, with the shortfall from the ideal factor $6.2$6 attributed to probe-induced decoherence, trap dephasing, and finite phase-shifter resolution of $6.2$7 mrad (Kohlhaas et al., 2015).

The significance of this PhaseLock form is that it treats the atoms as a phase memory for the LO rather than a terminal readout element. A plausible implication is that the method relaxes local-oscillator coherence requirements and changes the stability budget from being dominated by fringe ambiguity toward being dominated by residual decoherence and atomic lifetime.

3. Digital and fielded PhaseLock instrumentation

A second usage of PhaseLock refers to hardware PLL platforms. The open digital PhaseLock platform for optical metrology is implemented on Red Pitaya STEMlab hardware with a Xilinx Zynq 7010 FPGA, two 14-bit ADCs and two 14-bit DACs clocked at $6.2$8 MHz, and $6.2$9 MHz anti-aliasing filters on the inputs. Incoming signals are demodulated in I/Q form with a digitally generated reference of rational tuning

$[-\pi/2,\pi/2]$0

low-pass filtered with selectable bandwidths of $[-\pi/2,\pi/2]$1 MHz, $[-\pi/2,\pi/2]$2 MHz, or $[-\pi/2,\pi/2]$3 MHz, converted to phase by arctangent, and differentiated numerically so that the FPGA representation remains bounded. The loop filter is a $[-\pi/2,\pi/2]$4 controller, parameterized by physically meaningful quantities such as proportional gain and crossover frequencies rather than opaque register coefficients. The platform also includes an optional internal VCO path, dither/lock-in functionality for plant-gain sign detection, an integrated VNA, spectra and PSD diagnostics, and TCP-connected Python control with standalone operation after disconnection of the host PC (Tourigny-Plante et al., 2018).

Its reported latency is central. The core loop has $[-\pi/2,\pi/2]$5, and $[-\pi/2,\pi/2]$6 when the VCO path is included, implying theoretical control-bandwidth limits of about $[-\pi/2,\pi/2]$7 kHz and $[-\pi/2,\pi/2]$8 kHz, respectively, via $[-\pi/2,\pi/2]$9. In a $\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$0 m fiber-noise-cancellation experiment, the actuator and propagation delays dominate: the AOM behaves mainly like a delay of about $\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$1, the fiber contributes about $\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$2, and the total effective delay is about $\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$3, giving a predicted bandwidth of roughly $\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$4 kHz and an observed cancellation bandwidth of about $\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$5 kHz (Tourigny-Plante et al., 2018). The important point is that PhaseLock here is not merely a control law; it is a reusable instrumentation stack for lock construction, plant identification, and servo diagnostics.

A more application-specific PhaseLock system appears in submillimeter astronomy as a compact digital PLL for Gunn oscillators used in SMA receiver chains. The oscillator frequency is linearized near a tuning point as

$\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$6

and the loop continuously adjusts the Gunn bias voltage to compensate cavity-length drift. The design target is reacquisition within $\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$7 after a Walsh phase switch, with standard second-order-loop relations

$\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$8

and near-optimal damping $\varphi=\varphi_{\rm LO}-\varphi_{\rm at}$9. The system was used on SMA front-end receivers for more than a decade, enabled the first phase closure observations at $J_z$0 GHz, and more than three dozen units were constructed and deployed across several observatories and experiments (Hunter et al., 2011). In this usage, PhaseLock denotes a mature field instrument whose significance lies in maintaining interferometric phase coherence under thermal drift, mechanical perturbation, and observing-mode transients.

Related optical phase-locking work sharpens the same theme. In long-term stabilization of a mode-locked fiber laser, replacing a conventional electronic phase detector and PZT cavity tuning with an optic-microwave phase detector and pump-power tuning reduced the phase-detector floor from $J_z$1 fs rms to $J_z$2 fs rms over $J_z$3 Hz to $J_z$4 kHz, and reduced long-term phase drift from $J_z$5 ps rms to $J_z$6 fs rms over $J_z$7 hours (Hou et al., 2014). Although not named PhaseLock, it exemplifies the same architectural priority: shift the phase comparison and actuation to less noisy degrees of freedom.

4. Sequential PhaseLock for coherent beam combination

In coherent beam combining, PhaseLock denotes a sequential phase-locking scheme for filled-aperture intensity coherent combination of beam arrays. The motivating limitation is standard LOCSET-style control, where each channel is assigned a distinct dither frequency. Because these frequencies must be separated by the actuator bandwidth, the number of channels is restricted when phase actuators are bandwidth-limited, particularly for piezo-based devices (Klenke et al., 2021).

The proposed scheme uses a filled-aperture, sequential combination geometry with a segmented mirror splitter/combiner. The reflectivity of the $J_z$8-th zone is chosen as

$J_z$9

so that beams can be split or recombined with equal power distribution. A crucial analytical result is that measurement at the final combined-beam output does not produce a usable sign for later channels: the lock-in error for channel $\sin\varphi$0 contains contributions from earlier channels, so correction can acquire the wrong sign. The remedy is rejection-port detection at each combination stage. For a sinusoidal dither $\sin\varphi$1, the rejection-port error signal becomes

$\sin\varphi$2

so the current channel enters with the opposite sign relative to previously combined channels, giving the correct correction direction (Klenke et al., 2021).

To keep the signal amplitude from degrading with channel number, the method adds an alternating $\sin\varphi$3 phase-offset scheme. Channels are partitioned into “$\sin\varphi$4” and “$\sin\varphi$5” dither classes using $\sin\varphi$6 and $\sin\varphi$7, respectively, with synchronized offsets in both modulation and demodulation. For a “$\sin\varphi$8” channel, the small-error expression becomes

$\sin\varphi$9

which preserves the correct sign while keeping the signal amplitude approximately constant as $\cos\varphi$0 increases. The same dither frequencies can then be reused sequentially. In a two-dimensional array, a second modulation frequency separates intra-column locking from inter-column locking, reducing the problem to two sequential one-dimensional stages (Klenke et al., 2021).

The reported simulation for a $\cos\varphi$1 array uses $\cos\varphi$2 kHz, $\cos\varphi$3 kHz, and dither amplitude $\cos\varphi$4. The normalized combined power converges comparably to a standard LOCSET simulation that requires distinct frequencies from $\cos\varphi$5 kHz to $\cos\varphi$6 kHz, but with a much lower maximum actuator-frequency requirement (Klenke et al., 2021). PhaseLock here is therefore a scaling strategy: it trades unique per-channel frequency labeling for spatially staged inference plus local detector placement.

5. PhaseLock in image-to-video diffusion

A recent usage of PhaseLock appears in image-to-video diffusion as a training-free inference-time framework for preserving physically correct motion during denoising. The key empirical finding is that, for the same model, seed, and conditioning, a 2-step generation can be more physically consistent than a 50-step generation. On Physics-IQ, CogVideoX reports approximately $\cos\varphi$7 at step 2 and $\cos\varphi$8 at step 50, while visual quality improves as LPIPS drops from $\cos\varphi$9 to $[-\pi/2,\pi/2]$0 (Han et al., 4 Jun 2026). The paper interprets this as evidence that diffusion models may establish a valid motion prior early and then partially erase it during later refinement.

The spectral diagnosis is explicit. Writing the latent Fourier representation as

$[-\pi/2,\pi/2]$1

the paper associates magnitude $[-\pi/2,\pi/2]$2 with low-level appearance statistics and phase $[-\pi/2,\pi/2]$3 with structural organization and motion trajectory. Across steps $[-\pi/2,\pi/2]$4, magnitude correlation drops only about $[-\pi/2,\pi/2]$5–$[-\pi/2,\pi/2]$6, whereas phase coherence drops by about $[-\pi/2,\pi/2]$7 from step 2 to step 50; for CogVideoX it falls roughly from $[-\pi/2,\pi/2]$8 to $[-\pi/2,\pi/2]$9, and for Wan 2.1 from $565~\text{ns}$00 to $565~\text{ns}$01. A complementary corruption experiment reports $565~\text{ns}$02 EPE under $565~\text{ns}$03 phase corruption and $565~\text{ns}$04 EPE under $565~\text{ns}$05 magnitude corruption, indicating that motion is much more sensitive to phase degradation than to magnitude degradation (Han et al., 4 Jun 2026).

PhaseLock operationalizes this by extracting a motion prior from a few denoising steps and reimposing it during full denoising through latent delta guidance. With latent video $565~\text{ns}$06, the inter-frame delta operator is

$565~\text{ns}$07

A few-step sample with $565~\text{ns}$08 yields $565~\text{ns}$09, from which the prior

$565~\text{ns}$10

is extracted. During the full $565~\text{ns}$11-step trajectory, the current motion is $565~\text{ns}$12, the residual guidance is

$565~\text{ns}$13

and only frames $565~\text{ns}$14 are updated: $565~\text{ns}$15 With the assumption of similar adjacent-frame spectral magnitudes, the paper shows that the delta magnitude approximately tracks inter-frame phase difference: $565~\text{ns}$16 for small phase shifts (Han et al., 4 Jun 2026).

The default schedule uses $565~\text{ns}$17, $565~\text{ns}$18, and $565~\text{ns}$19, so guidance decays linearly and does not constrain late refinement excessively. Reported gains average $565~\text{ns}$20 Physics-IQ points across diverse models, with $565~\text{ns}$21 for CogVideoX-5B, $565~\text{ns}$22 for LTX-Video, $565~\text{ns}$23 for Wan 2.1, and $565~\text{ns}$24 for Wan 2.1 distilled (4-step), at overheads of $565~\text{ns}$25 time and $565~\text{ns}$26 memory. Direct Fourier-phase intervention performs much worse: on CogVideoX, full phase substitution yields $565~\text{ns}$27, low-frequency phase injection $565~\text{ns}$28, magnitude-preserving phase blend $565~\text{ns}$29, and iterative refinement $565~\text{ns}$30, compared with $565~\text{ns}$31 for PhaseLock (Han et al., 4 Jun 2026). In this domain, PhaseLock is not a classical servo but a denoising-trajectory regularizer that treats early motion as a phase-like invariant to be protected.

6. Broader phase-locking context and recurring misconceptions

The broader phase-locking literature clarifies what PhaseLock is and is not. In oscillator synchronization, phase locking may be extremely strong: a spin-transfer vortex nano-oscillator driven by an external microwave field exhibited a locking range larger than $565~\text{ns}$32 of the oscillator frequency, linewidth collapse to $565~\text{ns}$33 Hz in the locked regime, and persistence under detuning as large as $565~\text{ns}$34 MHz, which the authors described as “perfect” and “robust” phase-locking (Hamadeh et al., 2013). Microresonator combs have likewise been interpreted through self-injection locking between overlapping comb bunches, with an Adler-type model and a measured locking range of about $565~\text{ns}$35 kHz (Del'Haye et al., 2013). These results show that phase locking can denote entrainment of nonlinear oscillators, not only classical PLL circuitry.

A common misconception is that phase locking always means a unique phase state. Parametric systems provide counterexamples. In voltage-controlled parametric magnetization oscillation, the pump frequency is twice the oscillation frequency, so the response admits two stable phase states differing by approximately $565~\text{ns}$36, and which state is realized depends on the initial condition. Frequency sweeping can transiently break that symmetry and achieve a locked rate of $565~\text{ns}$37 in noiseless simulation, whereas merely narrowing the initial-state distribution through stronger anisotropy is much less effective (Taniguchi, 2023). The Josephson parametric phase-locked oscillator similarly possesses two stable self-oscillation phases, $565~\text{ns}$38 and $565~\text{ns}$39, and uses a weak locking signal to bias the system into one or the other; this bistability enabled BPSK demodulation with $565~\text{ns}$40 transmitted bits and 4 errors, and dispersive qubit readout with Rabi-oscillation contrast of $565~\text{ns}$41 (Lin et al., 2014). Phase locking, in other words, may mean state selection within a discrete phase manifold rather than convergence to a unique global phase.

Another misconception is that phase-locking statistics are automatically spectrally local. In analytic-signal processing, the phase-locking value is obtained after amplitude normalization,

$565~\text{ns}$42

and that normalization is nonlinear. The resulting broad spectral leakage can create bias or spurious locking, including decentering bias associated with nonzero circular mean and effective leakage to $565~\text{ns}$43 Hz. The amplitude-weighted phase-locking value,

$565~\text{ns}$44

was proposed precisely to retain a phase-vector interpretation while avoiding the spectral distortion induced by phase-only normalization (Kovach, 2017). This matters because some modern “PhaseLock” constructions, especially in generative modeling, rely on spectral language but do not operate via literal Fourier-phase substitution.

Finally, the classical PLL vocabulary itself requires care. The lock-in range was historically regarded as a useful but vague notion, and later work defined it rigorously in nonlinear state space as the largest interval of frequency steps for which a loop starting from a locked state reacquires lock without cycle slipping (Kuznetsov et al., 2017). More recent full-state-space modeling argues that phase-reduced PLL descriptions can miss higher-frequency components generated in the phase detector and can obscure synchronization boundaries that become visible in non-reductionist simulations (Piqueira et al., 2024). Experimental system identification of a bandpass-filter PLL further found that the reconstructed nonlinear phase function is non-harmonic and highly asymmetric relative to the textbook model (Mishchenko et al., 2021). These results indicate that PhaseLock, across domains, should not be reduced to the informal statement that “the phase is kept constant.” What matters is the observable, the admissible phase manifold, the corrective pathway, and the precise criterion by which lock is declared.