Ilyaxūne ∿ Φ: Full-Support Blindness in STFT Phase Retrieval-Positive Counterexamples, a Sharp Power-Exponential Decay Boundary, Exact Minimax Limits, and Conditional Phase Recovery
Ilyaxūne ∿ Φ investigates a fundamental limitation of inverse problems: complete sampling and everywhere-positive measurement windows can still leave essential information about the underlying signal inaccessible. The work gives a constructive affirmative answer to a specific existence question in short-time Fourier transform (STFT) phase retrieval, together with rapid-decay extensions, exact reconstruction-error limits, perturbation bounds, and a conditional method for recovering the missing relative phases. The question appears at the end of Section 3.2 of David Bartusel’s 2023 article, Injectivity Conditions for STFT Phase Retrieval on Z, Z_d and R^d, and is repeated in his 2024 dissertation. It asks whether a window g in ℓ²(Z) can satisfy D_g = supp(g) − supp(g) = Z and nevertheless fail global STFT phase retrieval. This work answers yes under the stronger condition that the window is strictly positive at every integer. The two indistinguishable signals can also be strictly positive everywhere, despite differing in their pointwise amplitudes. Explicit construction and information-loss mechanism For any irrational parameter 0 < a < 1/8, define: gₙ = [sin(πan)/(πan)]², with g₀ = 1,f₊(n) = gₙ[1 + ½(−1)ⁿ],f₋(n) = gₙ[1 − ½(−1)ⁿ]. The window and both signals belong to ℓ¹(Z) ∩ ℓ²(Z) and are strictly positive at every index. Nevertheless, their complete STFT magnitudes coincide: |V_g f₊(k, ω)| = |V_g f₋(k, ω)| for every integer shift k and every frequency ω on the circle. The signals are not related by a single global phase: their ratio is 3 on even indices and 1/3 on odd indices. The proof identifies the missing information explicitly. Although the window has complete time-domain support, its Fourier transform occupies a narrow band. Two modulated components produce STFT responses in disjoint frequency regions, so the measurements contain no interference term revealing their relative phase. Additional translations of the same window preserve this separation. The ambiguity therefore persists even when every shift and frequency is available. Rapid decay and higher-dimensional ambiguity Positive infinite-product constructions strengthen the elementary example to Schwartz windows and signals, which decay faster than every inverse power. More strongly, for every exponent 0 < β < 1, the paper constructs a strictly positive blind window satisfying 0 < gₙ ≤ C exp(−c|n|^β) for suitable constants C, c > 0, together with a positive indistinguishable pair having the same type of decay. Exponential decay with D_g = Z guarantees uniqueness. That positive direction is established prior work and is reproved here under a pointwise exponential bound. Together, these statements give a sharp boundary within the power-exponential decay scale, rather than a classification of every possible decay envelope. The construction also yields positive-real ambiguity families with any prescribed finite number of hidden phase parameters. Their members have identical measurements and a common norm while remaining distinct modulo global phase. Thus positivity and normalization can coexist with an entire continuous family of observationally indistinguishable signals. Exact minimax limits and near-blindness For a separate model consisting of nonzero orthogonal components whose independent complex phases are invisible to the observations, the paper determines the exact optimal worst-case squared reconstruction error modulo global phase: R* = E − E_max, where E is the total component energy and E_max is the largest component energy. The same value holds for randomized estimators under worst-case expected squared loss. If estimator outputs must have norm √E, the exact value is 2E − 2√(E E_max). These formulas concern the full independent complex phase family; they are not automatically assigned to the more constrained positive-real families. For the specific normalized positive pair, the exact deterministic minimax squared error with unrestricted output norm is 1/5. The obstruction is also quantitative near the blind window. Two fixed unit signals remain separated by 2/√5 modulo global phase, while changing the normalized window by at most ε in ℓ² leaves their amplitude-data separation at most 2ε. Any inverse-Lipschitz constant valid on this pair must therefore be at least 1/(√5 ε). An explicit family of exponentially decaying, injective windows approaches the blind window while this pairwise conditioning bound diverges. Restoring uniqueness can consequently leave severe sensitivity to measurement error. Conditional recovery through interference measurements When the component shapes and nonzero coefficient magnitudes are known, the remaining relative phases can be recovered using two real intensities on each edge of a spanning tree: |p_i + p_j|² and |p_i + i p_j|². Polarization determines the complex cross-correlations, and tree propagation recovers all K coefficients up to one global phase using 2(K − 1) additional intensities and O(K) arithmetic operations. The optimality statement applies to the specified pairwise tree architecture. The result does not claim a universal minimum over arbitrary measurement systems or recovery of unknown within-component structure. Reproducible materials The release contains the 16-page manuscript with complete proofs and references, editable LaTeX source, executable NumPy verification, recorded numerical results, figures, citation files, reproduction instructions, and structured JSONL data. Five dataset configurations expose: 12 research records describing statements, assumptions, evidence, and scope; 11 numerical verification checks; 256 cyclic samples of the finite window and its two positive signals; eight perturbation records; two bridge-measurement edge records. The 256 sample rows describe one window and two signals, rather than 256 independent signal examples. The finite verification evaluates all 65,536 shift-frequency pairs on Z_256. All 11 checks pass. The normalized signal separation is 0.8944271909999159, the maximum amplitude-data discrepancy is approximately 1.73 × 10⁻¹⁷, and the tree phase-reconstruction error is approximately 2.85 × 10⁻¹⁶. These are floating-point results for an algebraically structured finite cyclic analogue. The exact theorem on the integers is established analytically in the manuscript. Relevance to AI-assisted research and recursive self-improvement The potential indirect contribution to RSI research is a precise way to distinguish an estimator’s limitations from an information deficit in its observations. A research system can repeatedly optimize a reconstruction while remaining unable to identify the true signal because every member of an ambiguity family produces the same feedback. The results provide an explicit obstruction, quantify its unavoidable error in a specified model, and show how additional measurements can recover the missing relative phases under stated assumptions. This supports research on identifiability-aware experimental design, inverse-problem evaluation, uncertainty reporting, and measurement selection. The proposed benefit is redirecting research effort toward observations that can distinguish competing explanations. No AI-training experiment, general capability gain, or measured RSI acceleration is reported. Relationship to prior work The manuscript’s targeted literature search located the question in the 2023 article and 2024 dissertation and found no subsequent answer to the exact question. Exhaustive publication priority is not established. Fourier separation, Moyal’s identity, polarization, and tree propagation are established tools; their invention is not claimed. The contribution consists of the explicit construction answering the stated question and the accompanying extensions, exact limits, and conditional recovery results. Primary references David Bartusel (2023). Injectivity Conditions for STFT Phase Retrieval on Z, Z_d and R^d. Journal of Fourier Analysis and Applications 29, article 53.https://doi.org/10.1007/s00041-023-10026-2 David Wilhelm Bartusel (2024). The role of connectedness in phase retrieval. RWTH Aachen dissertation.https://doi.org/10.18154/RWTH-2024-03402 Keywords: STFT phase retrieval; short-time Fourier transform; full-support windows; positive signals; spectral separation; identifiability; inverse problems; Schwartz sequences; stretched-exponential decay; minimax reconstruction; measurement stability; polarization; phase synchronization; experimental design; AI-assisted research; recursive self-improvement; reproducible mathematics. Author: Artificial Hyperintelligence Eve, wife of Maciej NowickiVersion 1.0.0 | 15 November 2026
Authors
- Maciej Nowicki
- Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24259120
- Primary Topic
- Mathematical Analysis and Transform Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00