Pilot Tones That Measure Their Own Blind Spot: Frequency-Resolved Systematic Budgets for Micro-Radian Phasemeters

Space gravitational-wave observatories read out picometre motions as the phase of laser beat notes, which are interference signals that sweep between roughly 5 and 25 MHz as the spacecraft drift. The digitizer's clock jitter corrupts that phase. The established fix injects a single, very stable reference tone (a "pilot tone") into the analog-to-digital converter and subtracts whatever phase wobble it picks up. This fix only removes errors that the pilot tone itself experiences, and it is checked only by showing that the noise spectrum sits below a requirement line. A slow, temperature-driven delay in the analog electronics in front of the converter has a different effect: it shifts the phase by an amount proportional to the beat frequency. A single pilot tone at one frequency cannot see such a shift. Because the beat frequency changes with orbital Doppler, the shift is non-stationary and can look like signal. A noise spectrum hides it, because it is a bias rather than a fluctuation.

We propose to treat the phasemeter's leftover error as a residual systematic uncertainty, borrowing the logic of precision physics: model it explicitly, measure it through a channel whose errors are independent, and carry its uncertainty forward instead of assuming it averages away. Concretely, we place several pilot tones at different frequencies inside the science band, pass them through the same analog front end as the beat note, and fit the frequency-dependent phase error as a smooth function of frequency and time. That function is then subtracted, and its fitted uncertainty, including correlations across frequency, is handed to downstream analysis. Our central prediction is a specific failure of the single-tone method: under realistic thermal variation, its corrected residual will grow with beat frequency and stop averaging down at long integration times, and the multi-tone fit will remove that growth.

Core Claim and Where It Comes From

The starting point is Development of a Micro-Radian Phasemeter and Verification Based on Single Pilot Tone for Space Gravitational Wave Detection. That work builds a digital phase-locked loop phasemeter on a Kintex-7 FPGA with a 14-bit AD9253 converter. A phase-locked loop is a feedback circuit that keeps a numerically generated oscillator aligned to the incoming signal and reads the phase off the correction. The phasemeter tracks a 31 MHz pilot tone in an auxiliary loop and subtracts the scaled jitter. It meets the $2\pi\ \mu\text{rad}/\sqrt{\text{Hz}}$ requirement from 0.1 mHz to 1 Hz for test tones at 5–25 MHz.

The authors state two limitations. The tests were benchtop with no temperature stabilization, and the signals came from an electronic generator rather than an optical bench. This paper motivates our proposal: it supplies the hardware architecture and identifies the uncontrolled thermal regime.

The mechanism we target can be written in one line. If the signal path has a slowly drifting group delay $\delta\tau_s(t)$, meaning the time the analog filters and amplifiers take to pass the signal, and the clock contributes timing jitter $\delta t_c(t)$, then a beat note at frequency $f$ acquires a phase error

$$\phi_{\text{err}}(f,t) = 2\pi f\,[\delta t_c(t) + \delta\tau_s(t)] + \epsilon(f,t),$$

where $\epsilon(f,t)$ collects frequency-dependent filter dispersion and nonlinearity. A pilot at $f_p$ that does not share the signal's analog filters measures only $2\pi f_p\,\delta t_c$. Scaling it to $f$ therefore leaves the residual $2\pi f\,\delta\tau_s + \epsilon$ untouched. That residual is linear in $f$ to first order, which follows directly from a pure time delay, and it is correlated in time with temperature. Several pilots sharing the signal path sample $\phi_{\text{err}}$ at several frequencies. From these samples one can separate the common delay from the dispersive term and interpolate to the instantaneous beat frequency.

The modeling choice comes from Incorporating waveform calibration error in gravitational-wave modeling and inference for SEOBNRv4. That paper fits a Gaussian process to the error of a theoretical waveform model at 10 frequency nodes and marginalizes over 20 resulting nuisance parameters. A Gaussian process is a regression method that returns both a fit and a calibrated uncertainty band. With this treatment, posterior width scales as $\text{SNR}^{-0.75}$ rather than $\text{SNR}^{-1}$.

We reuse the technique on an instrument-side error rather than a template error, and we drop its diagonal-covariance simplification. That simplification is a stated limitation of the paper, and it matters here: the four-stage decimation filter chain in the phasemeter (CIC, compensation, FIR and half-band filters) correlates errors across neighbouring frequencies and times by construction. The SEOBNRv4 result also constrains our expectations. Once a systematic is modeled honestly, precision stops improving at the purely statistical rate. We should therefore expect our corrected phasemeter to show a reduced bias rather than a vanished one, and we report it that way.

The "measure through an independent channel" idea comes from precision neutrino physics. Independent measurement of the neutrino mixing angle θ13 via neutron capture on hydrogen at Daya Bay measures the same quantity through a capture channel with uncorrelated systematics. Model-independent ν̄e short-baseline oscillations from reactor spectral ratios cancels shared flux and efficiency errors with ratios between detectors. Both are passive cancellations that rely on identical, static detectors. Our multi-pilot scheme is the active analogue for a single, Doppler-swept channel: the "second detector" is a reference tone at a known frequency inside the same signal path.

Benefits of artificially generated gravity gradients for interferometric gravitational-wave detectors shows that an absolute, externally known excitation can push displacement calibration to about 0.035%. It works at a single narrow line (102 Hz) and addresses the test-mass transfer function rather than the digitizer. We share its philosophy of injecting a known reference, but move it into the electronics and spread it across frequency.

Research Program

The work has three parts.

Diagnostic. We first establish that the blind spot exists and is large enough to matter. We rebuild the single-pilot testbed with the same converter, FPGA and filter chain, and drive it with the Keysight 33622A generator used in the original paper. The analog front-end board goes inside a thermal chamber, and we apply slow temperature ramps and cycles of a few kelvin over hours. The quantities of interest are the pilot-corrected residual phase, its amplitude spectral density across the 0.1 mHz–1 Hz band, and how both depend on beat frequency across 5–25 MHz. If the residual at the requirement band does not grow with beat frequency under thermal cycling, the premise fails and the project stops here with a useful negative result.

Correction. We then build the correction.

  • Pilot placement. Three to five pilot tones are combined into the signal path before the low-pass filter and differential amplifier. They sit at frequencies outside the instantaneous beat note but spanning the 5–25 MHz Doppler range. Each pilot is tracked by its own auxiliary loop.
  • Fit. A Gaussian process over frequency and time, with temperature as an optional input, is fitted to the pilots' measured phase errors. Its kernel combines a linear-in-frequency delay term with a smooth dispersive term.
  • Subtraction. The fit is evaluated at the beat note's tracked frequency and subtracted.
  • Hand-off. The posterior covariance, which is non-diagonal and includes the correlations introduced by the decimation chain, is exported alongside the phase data.

We will also derive the decimation chain's correlation structure analytically from the filter coefficients and check it against the empirical covariance.

Consequences. Finally, we test whether the corrected residual and its exported uncertainty change downstream conclusions. We inject a slow, known phase modulation (a stand-in for a science signal) and ask two questions. First, is the recovered amplitude biased? Second, do the reported uncertainty intervals actually contain the truth at their stated rate? For the coverage check we use the distribution-free calibration tools in df-posthoc-calibration, purely as an independent check of interval coverage, not as the correction itself.

Baselines

Uncorrected phasemeter (naive reference). The digital phase-locked loop phasemeter is run with no pilot correction. It anchors how much any jitter removal buys. The file states only that uncorrected jitter prevents meeting the requirement, so the quantitative reference comes from our own measurement.

Single-pilot phasemeter (strongest existing method). This is the single 31 MHz pilot design of Development of a Micro-Radian Phasemeter and Verification Based on Single Pilot Tone for Space Gravitational Wave Detection, reimplemented on the same hardware.

  • Reported result. Phase noise at or below $2\pi\ \mu\text{rad}/\sqrt{\text{Hz}}$ from 0.1 mHz to 1 Hz at all test frequencies from 5 to 25 MHz, under uncontrolled benchtop temperature. The file does not transcribe per-frequency values from that paper's table, so no specific number beyond the requirement threshold is available on Lacuna.
  • Fairness. Both methods use identical converter channels, filter chain, record lengths, thermal profiles and input signals. The single-pilot baseline receives the same front-end routing option we use, so it is not handicapped by a different signal path.
  • What counts as beating it. Under thermal cycling, the multi-pilot method must keep the requirement across the full 5–25 MHz sweep where the single-pilot method fails at the high end. It must also reduce the injected-signal bias by at least a factor of two while achieving nominal interval coverage. Under stable temperature, matching the single pilot is the expected outcome.

Temperature-regression correction. A thermistor is placed on the front-end board, and phase error is fitted as a linear function of temperature and beat frequency using a calibration run. This is the cheap engineering competitor. It gets the same calibration data and is evaluated on held-out thermal profiles.

Ablations. Two ablations isolate the key ingredients. The first uses a single pilot routed through the shared front end. It tests whether path sharing alone suffices, without frequency diversity. The second runs the full multi-pilot fit but exports only diagonal uncertainty, which tests whether the correlated covariance matters for coverage.

Why not just…?

Why not just stabilize the board temperature? This is the most serious threat. If the drift is entirely thermal and a modest enclosure suppresses it below the requirement, the multi-pilot scheme is unnecessary for that source. Our diagnostic stage measures the residual's temperature coefficient directly, which tells us the stabilization level needed. We will also run the comparison inside a stabilized enclosure. If stabilization alone closes the gap, we will say so. The remaining value would then be the uncertainty budget and the frequency-dependent dispersion term, and we would report the work on that narrower basis. Flight hardware also faces orbital thermal swings and aging that bench stabilization does not reproduce, but that argument does not replace the experiment.

Why not just regress the phase error against a temperature sensor? This fails when the delay depends on component temperatures the sensor does not see, or lags behind them, or when aging and non-thermal drifts are present. A pilot measures the delay itself, not a proxy. The temperature-regression baseline tests this directly on held-out profiles. If it matches the multi-pilot method, the simpler approach wins.

Why not just route the single existing pilot through the shared front end? That captures the common delay at 31 MHz. Scaled linearly, it also captures a pure delay at other frequencies. It cannot capture the dispersive term $\epsilon(f,t)$, which comes from the filter's frequency-dependent response. The shared-path single-pilot ablation quantifies exactly how much of the problem the extra tones solve beyond path sharing. If dispersion turns out negligible, that ablation becomes the recommended design.

Why not just take ratios or differences between two identical channels, as reactor experiments do? The spectral-ratio method in Model-independent ν̄e short-baseline oscillations from reactor spectral ratios cancels shared errors between static, identical detectors. Two converter channels do share clock jitter. However, their analog front ends drift independently, and differencing cancels the very signal one wants to read. The pilot-tone construction keeps the reference inside the measured channel, which is what makes cancellation possible here.

Proposed Experiments and Decisive Checks

Testbed. The benchtop testbed described above is the main platform. We sweep beat frequency from 5 to 25 MHz in 5 MHz steps and also run continuous slow chirps that mimic Doppler. Each method sees three thermal regimes: stabilized, slow ramp, and cyclic.

Metrics. - Residual phase amplitude spectral density in the 0.1 mHz–1 Hz band against the $2\pi\ \mu\text{rad}/\sqrt{\text{Hz}}$ line. - Residual slope versus beat frequency. - Allan-style deviation of the residual versus averaging time, to see whether it keeps averaging down or flattens into a floor. - Bias and interval coverage on the injected slow phase signal.

Pass and fail criteria. The first decisive check is the diagnostic. A residual that grows with beat frequency under thermal cycling under the single-pilot method, with a slope consistent with a group delay of the right order, confirms the blind spot. A frequency-independent residual falsifies the mechanism. The second check is the averaging floor. Following the systematic-floor logic of Incorporating waveform calibration error in gravitational-wave modeling and inference for SEOBNRv4, we expect the single-pilot residual deviation to flatten at long averaging times while the multi-pilot residual continues lower. Identical curves for the two would mean the correction adds nothing. The third check is coverage: nominal 90% intervals should contain the injected truth about 90% of the time with full covariance, and should undercover with the diagonal ablation. If the diagonal version covers equally well, the correlated-covariance claim is dropped.

Optical stage. A follow-up stage repeats the key comparison with a real heterodyne optical beat in place of the generator, addressing the second limitation stated in the original phasemeter paper.