01 Open-source research framework
Dynamic Resonance Rooting finds the rhythms in multivariate time series, traces which signal leads which, and tells you how sure it can be. Every number on this page is computed live, in your browser, by a port of the Python package that is tested against it.
02 Live lab
Three channels stream at 100 Hz. dim_0 mixes a 7 Hz rhythm with broadband noise, dim_1 hears it through a delay you choose, and dim_2 is an unrelated 11 Hz distractor. DRR only sees the numbers. Every 512 new samples it measures resonance depth and tests every directed edge against 99 circular-shift surrogates.
Collecting samples
Set coupling to 0 and every edge DRR reports is a false alarm. At α = 0.05 you should see about one in twenty.
dim_1 — · dim_2 —
03 Evidence
Each figure below is read from a committed artifact that anyone can regenerate with one command. That includes the result that went against us.
Share of simulated null systems where the rooting test reported any edge.
Family-wise false-alarm rate of each surrogate null on independent channels whose dependence breaks its assumption by degrees.
Share of trials that recovered the injected edge at its true lag.
Median and 5–95% range of resonance depth over 200 draws.
✕Not supported
On 572 months of NOAA stratospheric wind data, DRR raised too many alarms outside the published disruption windows. We record the claim as not supported, and it stays inside that one domain.
Read the protocol and the artifact04 The method
FFT, Welch, and Morlet-wavelet spectra locate recurring frequencies. Peaks are refined below the grid spacing, so a tone scores the same wherever it falls.
Lagged correlation or transfer entropy scores every directed pair. Circular-shift surrogates keep each series' own memory and give family-wise adjusted p-values.
Resonance depth combines spectral concentration, persistence, phase coherence, and amplitude stability. The score is unchanged if you relabel Hz as cycles per month.
A directed edge is a lead–lag relationship under a chosen estimator and null model, not a proof of causation. A result in one domain does not carry over to another.
Run it on your own data
$ pip install -e .
$ drr-reproduce --no-artifacts
$ python scripts/run_calibration_study.py