Research byline: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Prepared for: Maciej Nowicki
Hugging Face release: 1.0.0 · 2 November 2026

Recover accessible photonic parameters through a finite hidden environment while removing specified instrument gains, actuator calibration factors, and readout drift. This release contains a 38-page theoretical manuscript, constructive Python implementations, synthetic experiments, uncertainty bounds, and passive-network counterexamples that establish the identification limits.

Start here: paper · reproduce · data guide · numerical summary · citation

Research question

Can measurements determine an accessible resonator's own loaded loss when unactuated resonators have different losses, the coherent receiver has unknown gain and offset, and the actuator has an unknown fixed calibration?

The manuscript gives a conditional constructive answer. It combines controlled rank-one perturbations, nuisance-eliminating response ratios, a hidden-mode curvature realization, annihilating-polynomial de-embedding, and metric whitening. Exact results require a valid finite visible-order model and known relative probe structure. Different reconstruction routes use different measurement and actuation resources.

Here, loaded local loss includes measurement-port loading as well as intrinsic local dissipation. It is distinguished from loading induced by coherent coupling to hidden resonators. The method does not infer intrinsic absorption separately from every external loss channel.

Central construction

After active response ratios reconstruct the projected Green matrix in an unknown actuator metric, its inverse has the form

$F(z)=zW-J-C(zI-A)^{-1}B.$

For true minimal visible hidden order $r$, let $p(z)=\det(zI-A)=z^r+cz^{r-1}+\cdots$. Then

$p(z)F(z)=Wz^{r+1}+(cW-J)z^r+\cdots.$

The leading two polynomial coefficients separate the accessible generator from the strictly proper hidden self-energy. With the exact annihilator, $r+2$ distinct nonpole samples determine these coefficients. Whitening gives

$K_{SS,\mathrm{can}}=W^{-1/2}JW^{-1/2}=U_T^\dagger K_{SS}U_T.$

For a diagonal actuator map, the remaining phase gauge preserves individual accessible onsite frequencies and loaded local losses. For an arbitrary invertible mixing map, only a unitary equivalence class and corresponding frequency-part/damping spectra are determined. The total samples needed to learn the annihilator are additional requirements of the hidden-mode construction; $r+2$ alone is not a universal acquisition count.

Results and assumptions

Result	Capability	Required scope
Scalar active probing	Remove unknown complex affine readout gain and offset within a probe	Known relative rank-one controls, visible response, nonpole states
Joint actuator calibration	Remove fixed gain magnitudes through a positive metric	A consistent actuator map across the required probes/frequencies
Hidden-mode realization	Recover visible hidden poles up to realization similarity	Sufficient curvature rank and justified true visible order
General-loss de-embedding	Recover the bare accessible block despite unequal hidden losses	Valid hidden annihilator or its certified uncertainty envelope
Finite-error propagation	Enclose hidden poles and propagate uncertainty to accessible damping	Valid input/model error budgets, conditioning margins, appropriate pole matching
Drift-aware acquisition	Reject affine readout gain and offset drift	Stationary internal device and settled responses; ramp denominator must remain nonzero
Passive counterexamples	Establish hidden-local-loss and unrestricted finite-data ambiguities	The explicit model classes and constructions in the manuscript

The earlier full-port and shared-loss results are retained as useful special cases. The main general-loss completion is in Section 9 of the paper.

Recorded synthetic evidence

All numerical data come from finite-dimensional coupled-mode simulations. They are not measured chip data or full-wave electromagnetic results.

Test	Recorded outcome
Two accessible modes, three hidden modes, unequal local losses	Accessible loaded losses $(0.31,0.73)$ recovered to approximately $10^{-15}$ in the exact example
One local actuator, two hidden modes	Six frequencies and 18 stationary complex readings recover accessible loaded loss $0.41$; generator error approximately $3.14\times10^{-14}$
Same scalar device with affine readout drift	Six frequencies and 24 ramp-scheduled readings; generator error approximately $1.86\times10^{-11}$ in the exact drift model
Noisy general-loss de-embedding	200 inverse-form trials; all valid conditional certificates contain the true tested quantities
Hidden-pole enclosures	240 inverse-form trials; all reported pole enclosures and cluster counts hold
Readout-drift certificates	1,440 scalar trials; all nonvacuous reported bounds hold
Near-exceptional hidden block	Enclosures merge rather than asserting unsupported individual pole resolution
Nonidentifiability controls	Algebraically indistinguishable passive networks with different hidden local losses; finite-sample matches with different accessible losses

The inverse-form noise sweeps, scalar drift noise sweeps, and exact end-to-end drift example are separate experiments. This release does not claim a noisy end-to-end raw-instrument benchmark. Supplied noise envelopes use known synthetic errors plus stated numerical budgets; they are not autonomous estimates of experimental uncertainty. Conservative bounds can be much wider than observed reconstruction errors.

Quick reproduction

Use Python 3.10 or newer for the research code. The recorded environment used Python 3.12.14, NumPy 2.3.5, and Matplotlib 3.10.8.

python -m pip install -r requirements.txt
python tools/verify_release.py
python tools/run_final_checks.py

For the recorded library versions, use requirements-reproduce.txt with Python 3.12 or 3.13. See REPRODUCE.md for individual commands, earlier reconstruction routes, and rebuilding the manuscript.

The integrity checker needs only Python's standard library. Scientific scripts use NumPy and Matplotlib and require no network, credentials, laboratory instruments, or specialized photonics software after dependency installation.

Repository map

Path	Contents
`research/IHORUNG_Photonics_Research.pdf`	Final standalone manuscript
`research/*.tex`	Complete master and section sources
`research/ihorung/`	Full-port inference, diagnostics, experiments
`research/ihorung_active/`	Active and joint readout/actuator calibration
`research/ihorung_hidden/`	Hidden-order, shared-loss, and curvature methods
`research/ihorung_final/`	General-loss, pole-certificate, and drift modules
`research/SHA256SUMS.json`	Hashes of all 70 original scientific files
`data/final_noise_summary.csv`	Convenient index of the general-loss noise sweep
`release/`	Publication metadata and preparation validation
`SHA256SUMS.txt`	Exact release upload allowlist and file integrity manifest

The repository is a research archive with heterogeneous records, rather than a single uniform machine-learning dataset. The automatic tabular viewer is disabled; load CSV, JSON, and NPZ files directly using the data guide.

Scientific scope and intended use

The results can support further work on coherent resonator diagnostics, calibration, system identification, and retuning. Microwave or acoustic systems sharing the same resolvent model are possible analogues; their hardware performance has not been demonstrated here.

Key restrictions are linear stationary internal dynamics, finite visible complexity, appropriate access/control profiles, settled acquisition, and valid uncertainty budgets. A chosen numerical rank alone does not establish the true visible order. Completely invisible modes remain unobserved. Overlapping pole discs do not supply individual matched-root bounds without enlargement or separation. Floating-point evaluation is not an interval-arithmetic proof of every reported digit.

Rational realization, Loewner methods, pseudospectral perturbation theory, polarization, and drift reversal have established antecedents, cited in the paper. This release does not establish historical first priority, universal hidden-network reconstruction, or a measured performance advantage. Exact counterexamples are part of the contribution: some stronger inferences are impossible from the stated measurements.

Provenance, citation, and licensing

This is AI-generated theoretical research prepared for Maciej Nowicki. The research byline is preserved as requested; it identifies a research persona. No personal or sensitive measurements are present in the generated scientific data.

The complete final scientific package is preserved byte-for-byte in research/. This Hugging Face edition adds navigation, citation, dataset documentation, and integrity tooling. See release/HUB_RELEASE.json for source-package provenance.

Use CITATION.cff or CITATION.bib when referring to this release. No new license grant is assigned during packaging; see LICENSE_STATUS.md.

Hugging Face documentation: dataset cards and uploading files.

Downloads last month: -

Total file size:

5.33 MB