3a Block-bootstrap surrogates (B=804 bins≈11 yr, n=5000): raw r(+15d)
p=0.022 — marginally significant on undetrended series, driven by
the shared solar-cycle trend not a causal signal.
3b Partial correlation (sunspot-regressed, no filter): r(+15d) drops
from 0.079 to 0.029 (63%) — confirms solar-cycle confounding without
preprocessing circularity.
3c Spectral coherence in solar-cycle band = 0.840 (>0.776 threshold);
kNN mutual information at τ=+15d = 0.000 nats (p=1.000) — no
nonlinear dependence at the claimed lag.
3d Missing-data impact: 0% NaN at station thresholds 2/3/5; r(+15d)
unchanged — missing data is not a confound.
3e Bin-size sensitivity: dominant peak at τ≈-520d for 1/5/27-day bins;
r at +15d scales with bin size (solar-cycle leakage, not physical).
3f GK declustering removes 28.4% of events as aftershocks; r(+15d)
changes by only Δ=0.014 — aftershock clustering not a confound.
3g Per-solar-cycle analysis (cycles 21–24): r(+15d) all positive
(0.018–0.073) but dominant peak lags scattered (-65/-125/+125/-125d)
— phase-drifting solar-cycle artefact, not physical precursor.
Script fixes: vectorised block bootstrap (401×5000 → NumPy matmul),
kNN MI with sorted searchsorted (O(N log N), no inflated values),
coherence nperseg 512→2048 (resolves solar-cycle band), axhspan→axvspan.
Paper: new Methods subsections (block bootstrap, partial correlation,
nonlinear dependence); new Results subsections for each check; updated
Conclusions with 7-item robustness summary; Kraskov2004 and
GardnerKnopoff1974 added to refs.bib; Homola2023 updated to arXiv
preprint 2204.12310; eq:energy duplicate label fixed. PDF: 35 pages.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
1d01fcf102
