technical note 02 · signal processing, time metrology & geophysics
Two-Axis Canonical Coherence and the Source-Phase Stationarity Gap Across a 3,002 km Magnetometer Baseline
Deriving GL(2, ℂ)-invariant canonical coherence over horizontal subterranean magnetometers, mapping null-subtracted cross-spectra into a 76.6 µs Carter Cramér–Rao timing floor, and analyzing why diurnal thunderstorm migration defeated unmodeled time-domain lag stationarity (Gate D2b).
1. The Algebraic Blind Spot of Two-Way Packet Timing
Every two-way packet synchronization protocol between clocks at sites A and B — including NTP (RFC 5905) and IEEE 1588 PTP — exchanges directional timestamps to measure two scalar intervals:
⇒ θ_hat = (Δt_fwd − Δt_rev) / 2 = θ + (d_fwd − d_rev) / 2, RTT = d_fwd + d_rev
Because two linear equations cannot solve for three unknowns (θ, d_fwd, d_rev), packet timing protocols must assume path symmetry (d_fwd = d_rev). An on-path adversary who buffers forward packets by +δ and accelerates reverse packets by −δ (for example, rerouting from a congested queue to a shorter physical path while holding the forward queue) biases the receiver's clock estimate by δ / 2 while leaving RTT unchanged. Packet-layer cryptographic authentication (NTS, RFC 8915; PTPsec) authenticates the timestamp integers written by the sender, not the physical propagation symmetry of the medium.
Detecting an asymmetric delay attack without relying on GNSS (itself vulnerable to localized RF spoofing and inaccessible inside subterranean vaults) requires an exterior physical observable visible across continental baselines. Global lightning (~50–100 strokes/s, ~1011 W electromagnetic excitation) continuously excites the Earth–ionosphere spherical waveguide at the Schumann eigenmodes (f_1 ≈ 7.8 Hz, f_2 ≈ 14.1 Hz, f_3 ≈ 20.3 Hz, f_4 ≈ 26.4 Hz), attenuating at only ~0.2–0.5 dB/Mm and penetrating deep underground enclosures.
2. Why Scalar Coherence Fails on Elliptically Polarized Horizontal Modes
Consider two-axis horizontal induction/fluxgate magnetometers recording x_H(t) = [x_HX(t), x_HY(t)]^T at LIGO Hanford (Washington) and x_L(t) = [x_LX(t), x_LY(t)]^T at LIGO Livingston (Louisiana), separated by a 3,002 km baseline (10.0 ms vacuum light-time; ~13.0 ms ELF waveguide delay at v_ph ≈ 0.77c).
During a preregistered six-hour quiet epoch in Observing Run O3 (2020-01-29 01:53:02–07:53:02 UTC, GPS 1264298000–1264319600, decimated from 4,096 Hz to 128 Hz and analyzed via 16 s Hann Welch windows with 50% overlap, Δf = 0.0625 Hz, N_avg = 2,699), evaluating ordinary scalar magnitude-squared coherence across the fundamental band [7.3, 8.3] Hz produces an apparent paradox:
| Axis Pair | Mean γ²_ij(f₁) in [7.3, 8.3] Hz | Physical Interpretation |
|---|---|---|
H_X – L_X | 0.0215 | Near noise floor despite identical channel labels. |
H_Y – L_Y | 0.0974 | Weak parallel-axis projection. |
H_X – L_Y | 0.1477 | Cross-polarized coupling exceeds parallel axes. |
H_Y – L_X | 0.1584 | Strongest single scalar pair, still missing orthogonal energy. |
| Canonical λ₁(f₁) | 0.3291 | 2.08× best scalar pair; captures full 2D polarization ellipse. |
In a spherical cavity, horizontal magnetic waves arrive elliptically polarized at site-dependent azimuths, and auxiliary facility magnetometers have differing local vault orientations and transfer gains. Picking any single scalar pair (i, j) discards wavefield power projected onto the orthogonal sensor axis.
Following Koopmans (1974) and Brillinger (1981), we instead form optimal frequency-dependent virtual antennas u(f), v(f) ∈ ℂ² maximizing inter-site coherence:
C(f) = S_HH(f)^(-1/2) · S_HL(f) · S_LL(f)^(-1/2), λ₁(f) = σ_max²( C(f) )
Calibration and rotation invariance: Suppose the physical sensor vectors at each vault undergo arbitrary invertible linear distortions x_H ↦ A_H x_H and x_L ↦ A_L x_L for any A_H, A_L ∈ GL(2, ℂ) (encompassing compass rotation, unequal coil gain, axis sign flips, and frequency-dependent cross-talk). Then S_HH ↦ A_H S_HH A_H† and S_HL ↦ A_H S_HL A_L†, which merely bijectively re-parameterizes the optimization vectors u ↦ A_H^(-†) u and v ↦ A_L^(-†) v, leaving the maximum eigenvalue λ₁(f) identically unchanged.
3. Surrogate Nulls, Vault Shielding Gain, and the 76.6 µs Carter Bound
Because maximizing over ℂ² × ℂ² introduces positive finite-sample eigenvalue bias even under zero coupling, all gates were frozen in preregistration/o3_vault_run.py (SHA-256 f45dca61...) before inspecting O3 data:
- Off-band empirical null (
μ₀): Averagingλ₁(f)across inter-mode troughs yieldsμ₀ = 0.0399 ± 0.0118, placing observed fundamental coherenceγ²_can(f₁) = 0.3291atz = 24.54(GateP1supported). - Vector block-permutation (
N1) and phase-randomization (N2) surrogates: Permuting 64 s two-axis L1 blocks (4× longer than the 16 s Welch window, preserving intra-site X/Y covariance and auto-spectra) collapses fundamental coherence to0.0020(max0.0065across 20 permutations,p = 1/21 = 0.0476); Fourier phase randomization collapses it to0.0014. - Subterranean vault SNR leverage (
P2): In an earlier asymmetric GW170814 pilot where H1 used an unshielded corner-station fluxgate exposed to 60 Hz building electronics, null-subtracted coherence was0.0058. Moving H1 to the isolated subterranean vault induction coil (H1:PEM-VAULT_MAG_1030X195Y_COIL_X/Y_DQ) raised null-subtracted coherence to0.2892— a49.9×like-for-like gain.
Subtracting the empirical off-band floor g(f) = max(λ₁(f) − μ₀, 0) defines the conservative coherent SNR SNR_eff(f) = g(f) / (1 − g(f)). Substituting into Carter's (1987) Cramér–Rao lower bound across [6, 25] Hz bounds the minimum variance of any unbiased delay estimator over integration window T = 21,600 s:
Scaled to a single 30-minute sub-window (T_w = 1,800 s = T / 12), the theoretical white-noise timing floor is σ_τ,30m = 76.6 µs × √12 = 265 µs, which fixed our preregistered 3σ inter-window stability threshold at max(3 σ_τ,30m, 0.3 ms) = 0.796 ms.
4. Anatomy of a Failed Gate: Why D2b Wandered by 4.95 ms
Gate D2b tested whether an unmodeled, orientation-free time-domain Generalized Cross-Correlation with Smoothed Coherence Transform (GCC-SCOT) could achieve stationary sub-millisecond lag tracking across twelve consecutive 30-minute windows:
D2b required a three-part conjunction. Two criteria passed and the third failed decisively:
- Per-window statistical significance (
> 6/12atp ≤ 0.05): 11 of 12 windows exceeded all 19 in-window block-permutation surrogates. - Physical propagation bound (
|τ_med| ≤ 15 ms): Observed median lag was−3.85 ms, well inside the Earth–ionosphere waveguide bound. - Inter-window lag stationarity (
MADN(τ_hat) ≤ 0.796 ms): Failed — observed normalized median absolute deviation was4.953 ms(6.2× above threshold, 18.7× the0.265 msper-window CRLB), with individual 30-minute peaks spanning−9.89 msto+12.23 mswhile both LIGO observatories remained GPS-locked to sub-microsecond precision.
Why does the spectral information bound permit 76.6 µs (D1) while unmodeled time-domain lag wanders by 4.95 ms (D2b)? Carter's CRLB bounds phase estimation precision conditioned on a known or time-invariant inter-site transfer function H_HL(f). In the real Earth–ionosphere cavity, excitation is dominated by three localized tropical continental thunderstorm chimneys (Maritime Continent/Asia, Africa, Americas) whose relative intensity migrates westward across the 01:53–07:53 UTC window. Because the Schumann field is a partially standing interference pattern rather than a single plane wave, shifting source geometry rotates the complex cross-spectral phase arg S_HL(f) by milliseconds of apparent group delay.
Crucially, two follow-on attempts archived in k-luecke/Geo using contemporaneous World Wide Lightning Location Network (WWLLN) stroke catalogs also failed to close this gap without full spherical-waveguide forward modeling: 30-minute WWLLN spatial energy conditioning (wwlln_coherency_v12) did not improve complex coherency prediction, and 128 Hz time-only transient common-view association (transient_commonview_v1) yielded p_emp = 0.4216 against a 1,000-shift circular null. Reaching the sub-100 µs physical bound in operational clock auditing therefore requires either multi-station low-rank source-subspace projection or explicit vector waveguide transfer inversion.
Geo investigations, and acknowledges the Gravitational Wave Open Science Center (gwosc.org) of LIGO, Virgo, and KAGRA for public O3 vault magnetometry. Verify the two-byte post-verdict serialization diff directly via cmp -l preregistration/o3_vault_run.py analysis/o3_vault_run_v4_1.py.