This page develops the theoretical language you need before entering the lab. It is organised in three parts: frequentist tools for characterising clock stability, Bayesian tools for inference under uncertainty, and the causal-geometry framework that unifies both perspectives within a single comparison-network logic.
Classical stability analysis — the language of frequency metrology
Begin with a signal of nominal frequency ν0. Any real oscillator departs from this ideal, and the instantaneous output can be written as
where φ(t) encodes all phase deviations from the ideal. The fractional frequency deviation is then
Every stability measure below reduces to a statement about the statistics of y(t) or φ(t).
The one-sided power spectral density (PSD) of fractional frequency fluctuations, Sy(f), characterises noise in the Fourier domain. A power-law model covers most physical oscillators:
| α | Noise type | Physical origin (examples) |
|---|---|---|
| +2 | White phase | Additive broadband noise in detection electronics |
| +1 | Flicker phase | Active-device noise in amplifiers |
| 0 | White frequency | Shot noise; thermal noise in passive resonators |
| −1 | Flicker frequency | Environmental perturbations; resonator ageing |
| −2 | Random-walk frequency | Environmental drift; temperature fluctuations |
The classical variance of y(t) diverges for noise types with α ≤ −1. Allan variance resolves this by using second differences of phase:
where ¯yn is the average fractional frequency over the n-th interval of duration τ. The Allan deviation σy(τ) maps directly onto the noise-type exponent α: a log-log plot of σy vs τ reveals distinct slopes for each regime.
| Noise type | Slope of σy(τ) on log-log |
|---|---|
| White phase (α=+2) | τ−1 |
| Flicker phase (α=+1) | τ−1 (with log correction) |
| White frequency (α=0) | τ−1/2 |
| Flicker frequency (α=−1) | τ0 (flat) |
| Random-walk frequency (α=−2) | τ+1/2 |
The standard Allan variance cannot distinguish white phase from flicker phase noise (both give τ−1). The modified Allan variance (MVAR) applies additional averaging to separate these regimes. In practice, the overlapping Allan variance provides better statistical confidence from finite datasets by reusing samples.
The allantools Python library computes all variants. In the tutorial you will implement the standard estimator from scratch before using library functions, so that you understand what the library is doing.
When comparing clocks pairwise within a network of N clocks, the comparison results must satisfy closure relations. For three clocks A, B, C:
Triangular closure provides a powerful self-consistency check that requires no external reference. Residual closure violations indicate unmodelled systematic effects or transmission noise. In a network of N clocks, the number of independent closure constraints grows as (N−1)(N−2)/2, rapidly over-constraining the system.
Inference under uncertainty — updating knowledge as data arrive
Given data D and model parameters θ:
The posterior p(θ|D) combines the likelihood p(D|θ) with prior knowledge p(θ). The normalisation p(D) — the evidence — becomes decisive when comparing models.
The Bayesian perspective is not opposed to frequentist methods; it answers a different question. Frequentist tools ask “how probable is this data given the model?” Bayesian tools ask “how probable is the model given the data?” Clock metrology benefits from both.
In clock parameter estimation, priors encode physical constraints:
A prior is not a bias; it is a documented assumption. The tutorial exercises will test sensitivity to prior choice.
For models with more than a few parameters, direct computation of the posterior is intractable. MCMC samplers (Metropolis-Hastings, Hamiltonian Monte Carlo) generate samples from the posterior by constructing a Markov chain whose stationary distribution is p(θ|D).
Practical diagnostics for convergence: trace plots, autocorrelation lengths, the Gelman-Rubin &Rcirc; statistic. You will use emcee or PyMC in the tutorial.
The evidence (marginal likelihood) ranks competing models:
The Bayes factor B12 = p(D|M1) / p(D|M2) quantifies how strongly the data prefer one model over another. This becomes directly relevant when deciding whether a clock’s noise is better described by, say, white + flicker frequency vs. white + random-walk frequency.
Generate synthetic datasets from the fitted posterior and compare their statistical properties (Allan deviation slopes, closure residuals) to the observed data. If the posterior predictive distribution does not contain the observations, the model is falsified — regardless of how good the fit appeared.
Unifying clocks through comparison geometry, not oscillator physics
Conventional clock science focuses on the oscillator: what oscillates, at what frequency, with what stability. This framework shifts the focus to the comparison: what can be compared, over what distance, in what time?
The central claim:
The framework separates three physically distinct length scales:
| Length | Symbol | Meaning |
|---|---|---|
| Source length | Lsource | Spatial extent of the oscillator itself (atom, pendulum, pulsar) |
| Apparatus length | Lapparatus | Size of the instrument chain (cavity, electronics, detector) |
| Comparison length | Lcomparison | Distance between the two clocks being compared |
In most treatments, these are conflated. The framework insists they are independently variable and independently constraining. Each generates its own noise floor, its own systematic budget, and its own scaling.
Define the causal efficiency of a comparison:
where τ is the averaging time. This parameter ranges from 0 (local comparison, unlimited averaging) to 1 (comparison at the causal limit — one light-travel-time, no averaging possible).
η(τ) is an architectural control parameter, not a figure of merit. It tells you where in the design space a given comparison sits:
The framework is falsifiable in the location and behaviour of ηopt. If the predicted crossover does not appear in simulation, or appears at a different τ, the framework is either incomplete or wrong. This is the standard the lab will hold you to: not “does the framework feel right?” but “does ηopt appear where predicted, and do deviations teach us something?”
For a network of N clocks, each pairwise comparison has its own ηij(τ). The network-level question becomes: given the graph topology and noise budgets, what is the optimal allocation of comparison resources (averaging time, link bandwidth)?
Triangular closure (Part A.5) and Bayesian model comparison (Part B.4) become tools within this network perspective. Closure residuals map directly to consistency of the η values; Bayesian evidence selects the noise model most compatible with the network geometry.
Frequentist tools (σy, Sy, closure) characterise individual clocks and pairwise links. Bayesian tools (posterior inference, model comparison, predictive checks) select among competing descriptions of the noise. The causal-geometry framework provides the architectural context: it tells you which comparisons are possible, what the fundamental noise floor is for a given geometry, and where to look for the signature of the comparison limit.
In the Tutorial, you will compute all three. In the Lab, you will bring them together in a clock network simulation.