# CSI subcarrier redundancy and effective-dimensionality audit This package asks a narrow question about the public GL-MT6000 captures: > Do 256 reported CSI amplitude bins behave like 256 independent dimensions? It does not map the array index to a physical frequency, antenna position or 802.11 resource unit. It also does not provide respiration, fall, occupancy or identity ground truth. ## Inputs The script expects two Capture v1 container files: ```text ``` The public de-identified inputs are: | Capture | Rows used | Input SHA-256 | | :--- | ---: | :--- | | 10-minute primary | 7,608 | `527466ceb8b5f835e55ff348f02808736b947db39e7baac1f34e375de146a440` | | 60-second comparison | 347 | `41c35d9d23e5c03ddd34ae97617e772fecdb55e8545156181828b060605a87a6` | The 10-minute file is distributed through Zenodo: . The 60-second file is available from the project's dataset page: . ## Method 1. Walk every outer record and close a measurement group when `chain_info` bit 15 is set. 2. Select `tx0/rx0` only. A missing or duplicate fixed chain is counted and the group is not silently replaced with another chain. 3. Keep complete, non-zero 256-bin amplitude vectors. 4. Divide each vector by its own median so a shared gain change does not dominate the comparison. 5. Centre and scale each subcarrier over time, then compute the 256 x 256 Pearson correlation matrix. 6. Eigendecompose that correlation matrix. Report the first component count reaching 90%, 95% and 99% cumulative variance, plus the participation-ratio effective rank. This is an amplitude-only, correlation-based audit. High correlation may come from the radio front end, shared gain, multipath, the capture format or other joint causes. Correlation alone does not establish causality. ## Reproduce ```bash python3 analyze_csi_subcarrier_redundancy.py \ csi_udp_20261010_10m.anon.bin \ csi_demo_60s_anonymized.bin \ --out-dir . \ --img csi-subcarrier-redundancy.svg ``` Only Python 3.9+ and NumPy are required. ## Results On the 10-minute correlation matrix: - 6 components explain 90% of the variance; - 9 components explain 95%; - 20 components explain 99%; - the participation-ratio effective rank is 4.568732; - the first component explains 33.933% and the first 10 explain 96.858%; - the median absolute off-diagonal correlation is 0.204452; - 17.451% of pairs have `|r| >= 0.75`, while 10.064% have `|r| >= 0.9`; - the median adjacent-bin correlation is 0.900779. The distribution is not one uniform correlation block. A low median does not contradict the low PCA dimension: the eigenspectrum is dominated by a smaller number of strongly structured directions. The normalized median amplitude profile correlates at 0.954781 between the two captures, and the per-subcarrier temporal-CV profile correlates at 0.984305. That is evidence of stable structure across these two sessions, not proof that the same number of components will hold on another router, room or firmware. ## Outputs - `summary.json`: machine-readable capture, correlation, PCA and limitation data; - `subcarrier_metrics.csv`: one row for each of the 256 array indices; - `pca_variance.csv`: all 256 eigenvalues and cumulative variance; - `SHA256SUMS`: integrity checks for the script, README and generated analysis files. The public article with methods, figures and limits is: - Chinese: - English: ## Evidence boundaries - The two inputs come from one GL-MT6000, one client, one home, one bandwidth and similar firmware. - The subcarrier index is not treated as a physical frequency. - There is no noise-only baseline or human-event label. - PCA variance depends on row normalization and per-subcarrier standardization; the published script and data are required for any alternative comparison. - These results support feature-redundancy checks and dimension reporting. They do not prove breathing, fall, occupancy, identity or medical capability.