PureOne's picture
Upload verified Ilyaxune phase-retrieval research release
9775ad6 verified
|
Raw History Blame Contribute Delete
11.4 kB
metadata
pretty_name: Ilyaxūne ∿ Φ — Full-Support Blindness in STFT Phase Retrieval
language:
  - en
size_categories:
  - n<1K
tags:
  - mathematics
  - phase-retrieval
  - short-time-fourier-transform
  - inverse-problems
  - identifiability
  - signal-processing
  - reproducibility
  - synthetic-data
  - ai-assisted-research
configs:
  - config_name: research_records
    default: true
    data_files:
      - split: test
        path: data/research_records.jsonl
  - config_name: verification_checks
    data_files:
      - split: test
        path: data/verification_checks.jsonl
  - config_name: signal_samples
    data_files:
      - split: test
        path: data/signal_samples.jsonl
  - config_name: perturbations
    data_files:
      - split: test
        path: data/perturbations.jsonl
  - config_name: bridge_measurements
    data_files:
      - split: test
        path: data/bridge_measurements.jsonl

Ilyaxūne ∿ Φ

Full-Support Blindness in STFT Phase Retrieval

Positive counterexamples, a sharp decay boundary, exact minimax loss, and phase-bridge repair

Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Release: 1.0.0 · 15 November 2026
Contents: a 16-page mathematical manuscript, editable LaTeX, executable finite verification, recorded outputs, figures, and structured research data.

Read the complete paper · LaTeX source · Recorded results · Reproduce · Data schema

This release presents a constructive affirmative answer to the existence question stated at the end of Section 3.2 of David Bartusel's 2023 paper: can a window satisfy

Dg=supp⁡(g)−supp⁡(g)=ZD_g=\operatorname{supp}(g)-\operatorname{supp}(g)=\mathbb Z

and nevertheless fail global STFT phase retrieval? The construction has the stronger property that the window and both indistinguishable signals are strictly positive at every integer. The manuscript supplies the analytic proof and extensions. The numerical files reproduce a separate finite cyclic analogue.

The repository is a mathematical research dataset and reproducibility archive. It contains synthetic mathematical examples and recorded computations. It includes no trained model or observations of AI self-improvement.

Central construction

For a signal and window in $\ell^2(\mathbb Z)$, use

Vgf(k,ω)=∑n∈Zfngn−k‾e−2πinω,k∈Z,ω∈R/Z.V_gf(k,\omega)=\sum_{n\in\mathbb Z}f_n\overline{g_{n-k}}e^{-2\pi i n\omega},\qquad k\in\mathbb Z,\quad \omega\in\mathbb R/\mathbb Z.

Choose an irrational $0<a<1/8$ and define

gn=(sin⁡(πan)πan)2,g0=1,g_n=\left(\frac{\sin(\pi an)}{\pi an}\right)^2,\qquad g_0=1,

f±(n)=gn(1±12(−1)n).f_\pm(n)=g_n\left(1\pm\frac12(-1)^n\right).

The window and signals belong to $\ell^1\cap\ell^2$ and are positive everywhere. They satisfy

∣Vgf+(k,ω)∣=∣Vgf−(k,ω)∣for every k,ω,|V_gf_+(k,\omega)|=|V_gf_-(k,\omega)|\quad\text{for every }k,\omega,

although the signals are different modulo global phase. Their pointwise ratio is $3$ on even indices and $1/3$ on odd indices.

The mechanism is spectral separation. The Fourier transform of $g$ is supported in $[-a,a]$. The STFT responses of $g$ and $(-1)^ng_n$ lie in disjoint frequency arcs, so no observed intensity contains their interference term. Complete time support therefore does not guarantee sufficient information for recovery.

Results and their scope

Result Exact scope
Positive full-support counterexample Equality holds for every integer shift and every frequency on the circle.
Rapid-decay extensions Positive Schwartz examples and examples bounded by $Ce^{-c
Decay boundary Exponential decay with $D_g=\mathbb Z$ gives uniqueness; this positive direction is established prior work, reproved under a pointwise exponential bound. Sharpness is within the power-exponential scale.
Arbitrary ambiguity dimension Positive-real signal families with any prescribed finite number of hidden phase parameters.
Exact minimax risk For independent invisible orthogonal complex sectors: $E-E_{\max}$, also for randomized estimators under worst-case expected squared loss. With fixed output norm $\sqrt E$: $2E-2\sqrt{EE_{\max}}$.
Positive-pair risk The specified normalized positive pair has exact deterministic minimax squared error $1/5$ with unrestricted output norm.
Near-blindness Unit signals separated by $2/\sqrt5$ have amplitude-data distance at most $2\varepsilon$ when the window changes by at most $\varepsilon$ in $\ell^2$.
Conditional phase repair $2(K-1)$ additional real intensities and $O(K)$ arithmetic recover relative phases of known components with known nonzero magnitudes using a spanning tree.

The independent-complex-sector minimax formulas are not asserted for the restricted positive-real torus. The measurement count is optimal within the specified pairwise tree architecture; it is not claimed to be a universal lower bound for arbitrary measurement designs. The paper explains these distinctions and gives complete proofs.

Executed finite verification

The supplied NumPy program evaluates all $256^2=65,536$ shift-frequency pairs of a finite cyclic construction with a unitary DFT. The finite window is constructed by a geometric Fourier sum; the sequence is not an arbitrary truncation of the infinite sinc example.

Recorded quantity Value
Minimum window entry 0.0023499012731199452
Signal distance modulo global phase 0.8944271909999159
Predicted signal distance $2/\sqrt5$
Maximum amplitude-data discrepancy $1.734723475976807\times10^{-17}$
Full-grid amplitude-data $\ell^2$ discrepancy $2.716684976246151\times10^{-16}$
Tree phase-reconstruction error $2.8521042017553903\times10^{-16}$
Numerical checks passed 11 / 11

These are floating-point results for the shipped finite example. The infinite-dimensional equality is proved analytically, independently of those residuals. Small roundoff differences can occur across environments.

Finite cyclic verification: distinct positive signals, magnitude data, and residual

Dataset configurations

Each configuration contains one test split. This is a reproducibility collection; test does not imply a held-out machine-learning benchmark or a claim of generalization.

Configuration Rows Contents
research_records 12 Structured statements, assumptions, proof locations, evidence categories, and limitations. These summaries do not replace the paper's proofs.
verification_checks 11 Recorded Boolean results from the finite verification program.
signal_samples 256 Every cyclic sample of the finite window and its two positive signals.
perturbations 8 Four perturbation sizes, each with raw and normalized windows, measured discrepancies, and theoretical bounds.
bridge_measurements 2 The two tree edges in the known three-sector phase-recovery example, with two interference intensities per edge.

Load a configuration after replacing the repository identifier:

from datasets import load_dataset

repo_id = "YOUR_USERNAME/Ilyaxune-Phase-Retrieval"
records = load_dataset(repo_id, "research_records", split="test")
samples = load_dataset(repo_id, "signal_samples", split="test")

Load directly from a downloaded copy without custom dataset code:

from datasets import load_dataset

records = load_dataset(
    "json", data_files={"test": "data/research_records.jsonl"}, split="test"
)

The plain JSONL files are also readable using Python's standard json module. DATA_SCHEMA.md describes every field. tools/build_dataset.py regenerates the numeric exports from the recorded result file and the original finite construction.

Reproduction and repository contents

python -m pip install -r requirements.txt
python tools/reproduce.py --output-dir ../Ilyaxune_reproduced

The wrapper runs the original verification in an isolated temporary directory and writes fresh results to the requested output directory. The shipped manuscript and recorded results remain available for comparison. See REPRODUCE.md for rebuilding the LaTeX paper and numeric dataset exports.

The top-level PDF, TeX, verify.py, verification.json, and figure files are the research artifacts. data/ contains derived exports; tools/ contains reproduction helpers. PROVENANCE.json identifies the source artifacts. RELEASE_MANIFEST.json records SHA-256 hashes and sizes for every uploadable file except the manifest itself.

Relevance to AI-assisted research and RSI

The potential indirect use is identifying when an optimization or self-supervision loop is trying to infer a distinction erased by its observations. This construction provides a precise test case: different ground truths have identical losses against the complete measurements. The minimax theorem quantifies the information deficit in its specified model, and additional interference measurements repair the relative-phase deficit under explicit assumptions.

Suitable uses include inverse-problem teaching, testing uncertainty reporting, checking phase-retrieval implementations, and developing measurement-selection experiments. The shipped dataset contains no AI training, autonomous self-modification, or measured RSI acceleration. A practical acceleration claim would require a separate experiment comparing reconstruction and measurement-selection methods under matched resources.

Evidence, provenance, and limitations

The manuscript states a complete proof answering the question documented in the 2023 paper and repeated in the 2024 dissertation. Its targeted literature search found no later answer to that exact question; an exhaustive publication-priority claim is not established. Fourier separation, Moyal's identity, polarization, and tree propagation are established tools. The exponential uniqueness result is explicitly attributed to prior work.

The source visual-topological attachment is audited in the manuscript. Its undefined constructions are not premises of the theorems. This release preserves the research byline and discloses its AI-assisted origin. No external peer review, formal proof-assistant verification, or general AI capability gain is asserted by this dataset card.

No reuse license was specified in the source package. This release preserves that status; citation information is supplied in CITATION.cff and citation.bib. It does not add a new license grant.

References

  1. David Bartusel (2023), Injectivity Conditions for STFT Phase Retrieval on $\mathbb Z$, $\mathbb Z_d$ and $\mathbb R^d$. Journal of Fourier Analysis and Applications 29, article 53. DOI: 10.1007/s00041-023-10026-2. The existence question appears at the end of Section 3.2.
  2. David Wilhelm Bartusel (2024), The role of connectedness in phase retrieval. RWTH Aachen dissertation. DOI: 10.18154/RWTH-2024-03402. The question is repeated in Section 6.2, printed page 80.
  3. David Bartusel (2026), Uncertainty Principles as a Tool for STFT Phase Retrieval. arXiv:2605.31507v1.

The complete bibliography and proof references are in the manuscript.