Waveform adaptation, iterative receivers and reusable computation
Experiment record · Receiver methods · Reproduce · Release & data · 中文
WaveForge-X is a CPU research platform for OFDM, OTFS and AFDM. It brings together nonstationary waveform selection, finite receiver budgets, decision-aware PCG stopping and reusable sparse time-domain preparation. NumPy/SciPy implementations, frozen settings and saved experimental outcomes make the comparisons inspectable.
| Topic | What the evidence establishes | Record |
|---|---|---|
| Waveform adaptation | Less switching can help; the tested controller does not beat the best fixed waveform. Corrected six-order probing matches Explore-Then-Commit. | Adaptation |
| Compute budgets | Receiver iterations create BER-cost tradeoffs. The tested joint selector is worse than a validation-selected fixed configuration at 2.49× work. | Receiver budgets |
| Anytime reception | Valid error sets constrain disagreement with a specified LMMSE hard-decision reference. Preparation can outweigh iteration savings. | Receiver study |
| Reuse and changing CSI | Legal static reuse saves preparation. Against strong cached baselines, further phase/refinement work does not show a general end-to-end advantage. | State reuse |
These are distinct experiments, including negative results. Comparisons retain their channel assumptions, statistical units and cost conventions rather than pooling everything into one claimed speedup.
- Gray BPSK/QPSK/16/64-QAM; unitary OFDM, reduced-prefix rectangular OTFS and chirp-prefix AFDM with complete useful-time CP/CPP channel operators.
- AWGN/fading, synthetic nonstationary multipath and finite-FIR TR 38.901 profile-based TDL. The latter is not a full standard-conformance claim.
- Dense LMMSE references, sparse time-domain PCG, symbol-domain sparsification, posterior decision checks, structured refinement and checked reuse/fallback.
- Paired Monte Carlo, noisy selected-action feedback, independent reevaluation, uncertainty/switching controls, cost decomposition and episode-cluster analysis.
Certificates concern the supplied model's reference hard decisions, not zero BER,
ML/MAP optimality or reliability on an unknown physical channel.
floating_point_certified remains False. Modeled work and measured CPU time
are reported separately. No hardware, CUDA or deep-learning framework is needed.
| Location | Contents |
|---|---|
src/waveforge6g/ |
Physical layer, policy experiments, iterative receivers and reuse |
configs/ |
Channel values, scenario grids and frozen method/analysis parameters |
scripts/ |
Fresh study runners, diagnostics, artifact checks and figure regeneration |
data/tables/ |
Seed/trajectory/episode summaries, paired results and ablations |
data/*.zip |
Checksummed synthetic evidence, including recorded failures |
docs/ |
Consolidated English methods, experiments, limitations and reproduction |
tests/ |
Channel/transform, policy, solver, decision-boundary and reuse checks |
Internal phase identifiers appear where required to trace original evidence. The public project remains one integrated edition. Experiment record explains which raw arrays are complete and which are representative prefixes.
Python 3.11 or 3.12. In Bash:
source scripts/env.sh
python -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/python -m pip install --no-deps -e .
.venv/bin/python -m pytest -q
.venv/bin/python scripts/verify_evidence.py --extract
.venv/bin/python scripts/make_figures.pyOn Windows use . ./scripts/env.ps1 and .venv/Scripts/python.exe.
The import and CLI name waveforge6g is retained for numerical compatibility;
the public package is waveforge-x.
.venv/bin/python -m waveforge6g run configs/smoke.yaml
.venv/bin/python -m waveforge6g research configs/research/smoke.yamlEnvironments, dependencies, caches and generated outputs stay in the checkout. Reading evidence and regenerating figures does not repeat the PHY experiments. See reproduction for expensive fresh studies.
Copyright © 2026 cabal312512. Code, original artwork and synthetic experiment artifacts are under the MIT License. Third-party papers/standards are cited rather than redistributed. Cite the software version using CITATION.cff. AI assistance contributed to implementation, diagnostics and technical documentation; the published numerical evidence records actual executed experiments.