Biorthogonal Lattice-Shock Decoupling, 0.376 us O(1) C99/SIMD Reads, and Constant-Memory Dual Telemetry with Zero-PyTorch ScarCodec v6.3 Across Ten Public Scientific Archives
Authors/Creators
Description
Scientific observatories, laser gravitational-wave interferometers, broadband seismic networks, climate oceanography grids, spacecraft buses, and clinical biopotential monitors emit IEEE-754 floating-point streams spanning 29 orders of magnitude (10^-21 strain to 1.5e8 photon counts). Lossless codecs (zlib, Blosc2) stall at 1.1x-1.3x on float mantissas, fixed-rate 4^d block transforms (ZFP) smear bright point sources, and modern variable-length predictors or biorthogonal-wavelet coders (SZ3, SPERR CDF 9/7, CFITSIO fpack) compress smooth fields tightly at the cost of 0.4-21.8 ms random-slice inflation and zero-mask wavelet bleed.
Here we present and forensically audit ScarCodec v6.3—a Zero-PyTorch (NumPy + multithreaded C99/SIMD) library, CRC32 mmap container (.scar), and Zarr plugin—across 16 SHA-256-verified files from 10 public archives across 8 physical sciences (26.23 MB raw binary, zero synthetic noise, zero oracle steering):
1. Biorthogonal Smooth-Shock-Lattice & Oblique Nulling (Delta = +0 B .scar): Coupled biorthogonal projections and Behrens-Scharf oblique projectors P_biorth improve every archive without adding a single header byte—yielding +8.56 dB SNR gain on the Mw 9.1 IRIS Tohoku seismogram (-38.69% bytes under epsilon = 5,000), -27.06% byte footprint on MIT-BIH 119 Holter ECG under clinical IEC 60601-2-47 (+/-50 uV, 0 violations across 1.30M samples), 0.8821 deg C ocean RMSE at 23.60x on 3D NOAA OISST v2 with exact 0.0000 deg C Antarctica land bleed (vs. 0.159 deg C SPERR and 0.079 deg C SZ3), and 100.0% oblique target retention under -60 dB broadband jamming (vs. 28.0% orthogonal P_perp and cos = -0.010 Nita-Gary spectral kurtosis).
2. Constant-Memory O(1) Dual SCache (34,816 B, Delta = 0 B): Active-slot dual Gram readout lifts blind 32->32 PhysioNet cortical EEG recall under oblique superposition from cos = 0.7433 to 0.9822 (beating Frequent Directions SVD sketching at 0.9570), MIT-BIH fusion-beat recall to 0.9858 (5.9x faster query than SVD solve), and corrupted NASA Space Shuttle fault recall from 0.7258 to 0.99994 (159x lower L2 error; 200x higher rare-fault capture than reservoir sampling).
3. Conway-Sloane O(8) C99 Speed, 2026 Exascale SOTA & 0.376 us mmap: Exact C99 D8 parity-coset decoding plus System 1 Gate G0 active compaction encode real archives at 44.4-72.5 MB/s on a single CPU, beat byte-matched SZ3/SPERR bright-source peak error by 7.88x-10.81x on CSO Bolocam (0.0031 Jy, exact 0.0 vacuum) and 2.19x-3.16x on 17.1x ROSAT Gosset E8 (7.80 vs. 17.06 SZ3 and 24.68 SPERR counts), and read random 64-float mmap slices in 0.3765 us (11,418x faster than zlib, 6,132x faster than SZ3, and 25,991x faster than SPERR), while candidly documenting smooth-field uniform-L_inf rate-access trade-offs and passing 37/37 adversarial IEEE-754/CRC32/concurrency checks.
Files
ScarCodec_v6_3_Biorthogonal_10Archives_Preprint_v3.0.pdf
Files
(3.2 MB)
| Name | Size | |
|---|---|---|
|
md5:12f7391a0641d13ff76f51bb61be14a7
|
1.1 MB | Preview Download |
|
md5:1fce51afc53f547af48f018061fd2952
|
924.6 kB | Preview Download |
|
md5:99f5f56ef1733a0b94a47a907da48d36
|
77.2 kB | Preview Download |
|
md5:d4796bb30163d22e89c2d42bdd485cbd
|
1.1 MB | Preview Download |
|
md5:cea149d60479a3bc3d4b9b3e6d458e8c
|
549 Bytes | Preview Download |
|
md5:67eae12606552ffbc133dbf99864aa7d
|
1.2 kB | Preview Download |
Additional details
Related works
- Is supplement to
- Preprint: https://github.com/byte-od/mixture-of-scars (URL)
- Is supplemented by
- Preprint: 10.5281/zenodo.24172784 (DOI)
Software
- Repository URL
- https://github.com/byte-od/mixture-of-scars/scar_codec
- Programming language
- Python