Metadata-Version: 2.4
Name: entroptics
Version: 0.2.4
Summary: The universe, in focus.
Author-email: Ikailo John Sessford <john@ikailo.com>
License: 
                                         Apache License
                                   Version 2.0, January 2004
                                http://www.apache.org/licenses/
        
           TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
        
           1. Definitions.
        
              "License" shall mean the terms and conditions for use, reproduction,
              and distribution as defined by Sections 1 through 9 of this document.
        
              "Licensor" shall mean the copyright owner or entity authorized by
              the copyright owner that is granting the License.
        
              "Legal Entity" shall mean the union of the acting entity and all
              other entities that control, are controlled by, or are under common
              control with that entity. For the purposes of this definition,
              "control" means (i) the power, direct or indirect, to cause the
              direction or management of such entity, whether by contract or
              otherwise, or (ii) ownership of fifty percent (50%) or more of the
              outstanding shares, or (iii) beneficial ownership of such entity.
        
              "You" (or "Your") shall mean an individual or Legal Entity
              exercising permissions granted by this License.
        
              "Source" form shall mean the preferred form for making modifications,
              including but not limited to software source code, documentation
              source, and configuration files.
        
              "Object" form shall mean any form resulting from mechanical
              transformation or translation of a Source form, including but
              not limited to compiled object code, generated documentation,
              and conversions to other media types.
        
              "Work" shall mean the work of authorship, whether in Source or
              Object form, made available under the License, as indicated by a
              copyright notice that is included in or attached to the work
              (an example is provided in the Appendix below).
        
              "Derivative Works" shall mean any work, whether in Source or Object
              form, that is based on (or derived from) the Work and for which the
              editorial revisions, annotations, elaborations, or other modifications
              represent, as a whole, an original work of authorship. For the purposes
              of this License, Derivative Works shall not include works that remain
              separable from, or merely link (or bind by name) to the interfaces of,
              the Work and Derivative Works thereof.
        
              "Contribution" shall mean any work of authorship, including
              the original version of the Work and any modifications or additions
              to that Work or Derivative Works thereof, that is intentionally
              submitted to Licensor for inclusion in the Work by the copyright owner
              or by an individual or Legal Entity authorized to submit on behalf of
              the copyright owner. For the purposes of this definition, "submitted"
              means any form of electronic, verbal, or written communication sent
              to the Licensor or its representatives, including but not limited to
              communication on electronic mailing lists, source code control systems,
              and issue tracking systems that are managed by, or on behalf of, the
              Licensor for the purpose of discussing and improving the Work, but
              excluding communication that is conspicuously marked or otherwise
              designated in writing by the copyright owner as "Not a Contribution."
        
              "Contributor" shall mean Licensor and any individual or Legal Entity
              on behalf of whom a Contribution has been received by Licensor and
              subsequently incorporated within the Work.
        
           2. Grant of Copyright License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              copyright license to reproduce, prepare Derivative Works of,
              publicly display, publicly perform, sublicense, and distribute the
              Work and such Derivative Works in Source or Object form.
        
           3. Grant of Patent License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              (except as stated in this section) patent license to make, have made,
              use, offer to sell, sell, import, and otherwise transfer the Work,
              where such license applies only to those patent claims licensable
              by such Contributor that are necessarily infringed by their
              Contribution(s) alone or by combination of their Contribution(s)
              with the Work to which such Contribution(s) was submitted. If You
              institute patent litigation against any entity (including a
              cross-claim or counterclaim in a lawsuit) alleging that the Work
              or a Contribution incorporated within the Work constitutes direct
              or contributory patent infringement, then any patent licenses
              granted to You under this License for that Work shall terminate
              as of the date such litigation is filed.
        
           4. Redistribution. You may reproduce and distribute copies of the
              Work or Derivative Works thereof in any medium, with or without
              modifications, and in Source or Object form, provided that You
              meet the following conditions:
        
              (a) You must give any other recipients of the Work or
                  Derivative Works a copy of this License; and
        
              (b) You must cause any modified files to carry prominent notices
                  stating that You changed the files; and
        
              (c) You must retain, in the Source form of any Derivative Works
                  that You distribute, all copyright, patent, trademark, and
                  attribution notices from the Source form of the Work,
                  excluding those notices that do not pertain to any part of
                  the Derivative Works; and
        
              (d) If the Work includes a "NOTICE" text file as part of its
                  distribution, then any Derivative Works that You distribute must
                  include a readable copy of the attribution notices contained
                  within such NOTICE file, excluding those notices that do not
                  pertain to any part of the Derivative Works, in at least one
                  of the following places: within a NOTICE text file distributed
                  as part of the Derivative Works; within the Source form or
                  documentation, if provided along with the Derivative Works; or,
                  within a display generated by the Derivative Works, if and
                  wherever such third-party notices normally appear. The contents
                  of the NOTICE file are for informational purposes only and
                  do not modify the License. You may add Your own attribution
                  notices within Derivative Works that You distribute, alongside
                  or as an addendum to the NOTICE text from the Work, provided
                  that such additional attribution notices cannot be construed
                  as modifying the License.
        
              You may add Your own copyright statement to Your modifications and
              may provide additional or different license terms and conditions
              for use, reproduction, or distribution of Your modifications, or
              for any such Derivative Works as a whole, provided Your use,
              reproduction, and distribution of the Work otherwise complies with
              the conditions stated in this License.
        
           5. Submission of Contributions. Unless You explicitly state otherwise,
              any Contribution intentionally submitted for inclusion in the Work
              by You to the Licensor shall be under the terms and conditions of
              this License, without any additional terms or conditions.
              Notwithstanding the above, nothing herein shall supersede or modify
              the terms of any separate license agreement you may have executed
              with Licensor regarding such Contributions.
        
           6. Trademarks. This License does not grant permission to use the trade
              names, trademarks, service marks, or product names of the Licensor,
              except as required for reasonable and customary use in describing the
              origin of the Work and reproducing the content of the NOTICE file.
        
           7. Disclaimer of Warranty. Unless required by applicable law or
              agreed to in writing, Licensor provides the Work (and each
              Contributor provides its Contributions) on an "AS IS" BASIS,
              WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
              implied, including, without limitation, any warranties or conditions
              of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
              PARTICULAR PURPOSE. You are solely responsible for determining the
              appropriateness of using or redistributing the Work and assume any
              risks associated with Your exercise of permissions under this License.
        
           8. Limitation of Liability. In no event and under no legal theory,
              whether in tort (including negligence), contract, or otherwise,
              unless required by applicable law (such as deliberate and grossly
              negligent acts) or agreed to in writing, shall any Contributor be
              liable to You for damages, including any direct, indirect, special,
              incidental, or consequential damages of any character arising as a
              result of this License or out of the use or inability to use the
              Work (including but not limited to damages for loss of goodwill,
              work stoppage, computer failure or malfunction, or any and all
              other commercial damages or losses), even if such Contributor
              has been advised of the possibility of such damages.
        
           9. Accepting Warranty or Additional Liability. While redistributing
              the Work or Derivative Works thereof, You may choose to offer,
              and charge a fee for, acceptance of support, warranty, indemnity,
              or other liability obligations and/or rights consistent with this
              License. However, in accepting such obligations, You may act only
              on Your own behalf and on Your sole responsibility, not on behalf
              of any other Contributor, and only if You agree to indemnify,
              defend, and hold each Contributor harmless for any liability
              incurred by, or claims asserted against, such Contributor by reason
              of your accepting any such warranty or additional liability.
        
           END OF TERMS AND CONDITIONS
        
           APPENDIX: How to apply the Apache License to your work.
        
              To apply the Apache License to your work, attach the following
              boilerplate notice, with the fields enclosed by brackets "[]"
              replaced with your own identifying information. (Don't include
              the brackets!)  The text should be enclosed in the appropriate
              comment syntax for the file format. We also recommend that a
              file or class name and description of purpose be included on the
              same "printed page" as the copyright notice for easier
              identification within third-party archives.
        
           Copyright 2026 Ikailo Inc.
        
           Licensed under the Apache License, Version 2.0 (the "License");
           you may not use this file except in compliance with the License.
           You may obtain a copy of the License at
        
               http://www.apache.org/licenses/LICENSE-2.0
        
           Unless required by applicable law or agreed to in writing, software
           distributed under the License is distributed on an "AS IS" BASIS,
           WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
           See the License for the specific language governing permissions and
           limitations under the License.
Project-URL: Repository, https://github.com/Agience/entroptics
Project-URL: Issues, https://github.com/Agience/entroptics/issues
Keywords: entroptics,entropy,optics,aperture,signal processing
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Operating System :: OS Independent
Classifier: Intended Audience :: Science/Research
Classifier: Topic :: Scientific/Engineering
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE.md
License-File: NOTICE
Requires-Dist: numpy>=2.0
Provides-Extra: torch
Requires-Dist: torch>=2.0; extra == "torch"
Provides-Extra: scipy
Requires-Dist: scipy>=1.10; extra == "scipy"
Provides-Extra: test
Requires-Dist: pytest>=8; extra == "test"
Requires-Dist: scipy>=1.10; extra == "test"
Requires-Dist: threadpoolctl>=3; extra == "test"
Provides-Extra: figures
Requires-Dist: matplotlib>=3.7; extra == "figures"
Requires-Dist: h5py>=3.8; extra == "figures"
Provides-Extra: dev
Requires-Dist: pytest>=8; extra == "dev"
Requires-Dist: scipy>=1.10; extra == "dev"
Requires-Dist: threadpoolctl>=3; extra == "dev"
Requires-Dist: torch>=2.0; extra == "dev"
Dynamic: license-file

# Entroptics

[![PyPI](https://img.shields.io/pypi/v/entroptics)](https://pypi.org/project/entroptics/)
[![Python](https://img.shields.io/pypi/pyversions/entroptics)](https://pypi.org/project/entroptics/)
[![License](https://img.shields.io/badge/license-Apache%202.0-blue)](https://github.com/Agience/entroptics/blob/main/LICENSE.md)
[![CI](https://github.com/Agience/entroptics/actions/workflows/ci.yml/badge.svg)](https://github.com/Agience/entroptics/actions/workflows/ci.yml)
[![Proofs](https://img.shields.io/badge/proofs-Lean%204%20%2F%20Mathlib-4B0082)](https://github.com/Agience/entroptics/tree/main/research/lean)
[![Paper](https://img.shields.io/badge/paper-PDF-b31b1b)](https://github.com/Agience/entroptics/blob/main/research/PAPER.pdf)
[![DOI](https://img.shields.io/badge/DOI-10.5281%2Fzenodo.21273400-blue)](https://doi.org/10.5281/zenodo.21273400)
[![Sponsor](https://img.shields.io/badge/Sponsor-Agience-EA4AAA?logo=githubsponsors&logoColor=white)](https://github.com/sponsors/Agience)

## *The universe, in focus.*

**Read any 2-D signal as a finite optical aperture whose resolution is fixed by the signal's own entropy.**

Entroptics (entropy + optics) treats a 2-D array `W` of shape `(T, F)`, one **ordered** axis (time / evolution) and one **feature** axis (channels / frequency), as a finite optical aperture. The signal sets its own focus from the entropy of its power marginals. Every quantity Entroptics reports is then a standard optical / wave measurement (étendue, Strehl, OTF, diffraction limit, propagation constant), and the same reads apply to any structured 2-D field: a spectrogram, a waterfall, an embedding stack, a market panel, an image.

It is a small, standalone library, **numpy only** at the core (scipy and torch optional), built entirely from geometry and standard theorems. Parameter-free and domain-agnostic.

## Statement of need

Selecting a rank is a prerequisite for a large class of analyses, and it is usually done with a threshold the analyst picks or a criterion that needs a known noise level. In practice a pipeline then carries a constant that was tuned on one instrument and is silently wrong on another.

Existing tools reflect this. `scikit-learn`'s PCA offers a variance-explained fraction or Minka's MLE; [`optht`](https://github.com/erichson/optht) implements the Gavish–Donoho optimal hard threshold, whose unknown-noise form estimates its scale from the median singular value; and the Wax–Kailath AIC and MDL criteria, standard in array processing, are derived for *n* snapshots of *p* variables with *n* > *p* and are undefined otherwise.

**Entroptics is for people who need a rank, a resolution and a reconstruction from records whose instrument they do not control, or across many substrates at once** — radio-astronomy waterfalls, spectrograms, sensor panels, embedding stacks — where a per-substrate constant is exactly what cannot be supplied. Its distinguishing property is that **no constant in it is fitted to data or calibrated to a substrate**: every fixed number is a derived mathematical quantity (a χ² median, an influence-function variance, a universal Tracy–Widom quantile) or a criterion stated in the documentation. What you supply is an operating point — a false-alarm level α, and the null it is taken against.

Two further needs it addresses are more specific. It handles the wide, short regime (*F* ≫ *T*) that a dedispersed burst cutout or a short multichannel record presents, where the finite-size Tracy–Widom edge holds the top eigenvalue but the asymptotic Marchenko–Pastur edge does not. And it treats a masked or never-measured cell as **absent** throughout, so every read divides by the extent actually observed rather than the array's nominal shape.

The intended audience is researchers and practitioners in signal processing, radio astronomy and applied statistics. For the DMD component specifically, [PyDMD](https://github.com/PyDMD/PyDMD) and [PyKoopman](https://github.com/dynamicslab/pykoopman) are more complete DMD libraries and are the better choice if DMD is your problem; Entroptics differs in maintaining the operator as a fixed sufficient statistic whose read cost is independent of stream length, and in truncating at its own derived floor rather than a supplied tolerance.

## Install

```bash
pip install entroptics              # core (numpy only)
pip install "entroptics[torch]"     # + GPU / torch tensors
pip install "entroptics[scipy]"     # + exact MAD constant
```

Requires Python ≥ 3.10 and numpy ≥ 2.0. From a checkout, `pip install -e ".[dev]"` adds pytest,
scipy and torch; `[figures]` adds matplotlib and h5py, which `research/figures/*.py` needs and the
library core does not.

## The apparatus

The optical chain: **a beam passes through an aperture, a lens converts it, and a screen receives it as a projection.**

A screen is viewed differently from each side, and `Projection` is that view made concrete — one signal on its own entropy-matched grid. 

| | |
|---|---|
| **`Aperture`** | What **bounds** a beam, and the read of everything *about* it (read-only). The single front door: batch `Aperture(W)` or streaming `Aperture(window=…).update(frame)`. Per-axis and screen-area reads, the mode spectrum, the diffraction limit, and a streaming dynamical operator for exact decay rates. Its size *is* the beam's étendue. |
| **`Projection`** | The screen **as one side sees it** — one signal on its own entropy-matched grid (read-only). Folds the signal onto its entropy-matched grid and reads its SVD factorization, coherence, and the modes above the noise floor. A denoised view is the `extract` filter — a projection onto those resolved modes. |
| **`Lens`** | A system's **conversion**: `entry` (surface → the screen's coordinates) and `inverse` (back out), plus that system's own laws — `energy`, `zero`, `null`. All domain code lives here. |
| **`Screen`** | Where beams **land**: the surface two or more systems share, and the crossing measurements between them. |
| **`Beam`** | What a side **carries**: energy, étendue, and the directions it occupies. A bundle of beams — each mode is itself a `Beam`, down to a leaf spanning one direction. |

`ap.projection()` and `Projection(W).aperture()` cross between the single-signal views.

### Which to reach for

| you have | you want | use |
|---|---|---|
| one signal | its structure, resolution, decay rates, étendue | **`Aperture`** |
| one signal | its factorization, embedding, or a denoised view | **`Projection`** (or `Aperture.extract`) |
| two or more systems | them to meet, convert, couple, or trade energy | **`Screen`** |
| a surface to convert | it mapped onto a screen's coordinates and back | **`Lens`** |

A `Screen` earns its place only when the question is *between* signals. For one signal it adds a
lens registration and a placement for a reading `Aperture` or `Projection` gives directly.

### The screen

`N` lenses carry `N` conversions and no pairwise table exists. Every screen read runs on the native, un-folded frame.

```python
from entroptics import Screen

s = Screen()                                                  # far= sets the reader's level
s.register("A", entry=to_concept_a, inverse=from_concept_a)   # a lens IS its conversion
s.register("B", entry=to_concept_b, inverse=from_concept_b,
           energy=my_energy_law, zero=my_zero, null=my_floor)  # ...and its own laws

concept   = s.place("A", surface_a)      # A.entry   : signal -> concept
surface_b = s.render("B", concept)       # B.inverse : concept -> B's signal
```

**Reads between sides**

```python
s.couple("A", "B")        # the MEASURED signed coupling (exact permutation null); 0 unresolved
s.coupling("A", "B")      # the evidence behind it: z, sign, strength, phase, tightness
s.transfer("A", "B")      # absorbed vs transmitted, the signed flux, and which concepts
                          # the energy condensed into
s.uncondensed("A", "B")   # what did NOT condense, as a frame ready for another screen
```

**Reads about one side**

```python
s.beam("A")               # what it carries: energy, flow, etendue, modes, basis, profile
s.aperture("A")           # the full optics of that side
s.directions("A")         # the directions it resolves, against ITS OWN floor
s.energy("A")             # its energy flow, by ITS OWN law
```

**Certificates on a lens**

```python
s.certify("A", surface)   # imaging: does inverse . entry return the surface?
s.lossless("A", surface)  # the same as a bare relative residual
s.realise("A", "B")       # what the crossing actually delivers vs the étendue bound
s.linear("A")             # does the lens pass the beam's modes independently?
```

**The screen as a whole**

```python
s.read()                  # the aperture measurement of the joint frame
s.basis()                 # the shared surface's coordinates
s.balance()               # each side at its own zero, and whether that zero closes
s.resolution()            # the settled state; None when nothing clears the floor
```

The sides need not share an ordered axis — bank transactions run on real time, language on information flow, and they still meet on the shared basis. Only the row-paired reads (`couple`, `joint`, `read`, `resolution`) need a common order, and they say so.

**Who decides what.** A lens declares its own noise (`null`) — a detector and a market count signal differently. The false-alarm level `far` is the reader's, by Neyman–Pearson. The caller owns the loop: these reads are pull, and the handoff in either direction is a plain `(T, D)` frame.

## Quickstart

### Batch

```python
import numpy as np
from entroptics import Aperture

W = np.random.default_rng(0).standard_normal((256, 64))

ap = Aperture(W)

o = ap.optics()               # the full intrinsic read, as a dict (33 fields)
ap.phi, ap.magnification      # fill fraction and its reciprocal (scale duality)
ap.etendue, ap.space_bandwidth
ap.strehl                     # dominant-mode coherence
ap.T, ap.F                    # per-axis AxisRead(H, n, delta, phi, sigma)
ap.spectral                   # contrast, attenuation α, phase β, dispersion, resolved_power, dominance
ap.a_delta, ap.correlation_length, ap.mercer   # diffraction limit + certificate
ap.scale_profile()            # structure vs observation window (resolution vs aperture size)
```

### The projection

```python
sc = ap.projection()
sc.K_signal        # SVD modes standing above the noise floor
sc.coherence       # ordered-axis structure z-score (deterministic)
sc.footprints      # per-mode (phi_T, phi_F) localization of each resolved mode
```

### The filter

Pull the resolved signal out of the field — a projection onto its own screen modes.

```python
clean, info = ap.extract()          # the whole frame, since `Aperture(W)` reads all of it
info["K_signal"]                    # resolved modes above the floor
info["contrast"]                    # leading singular value over the floor (σ₁ / Φ)
info["n_kept"], info["n_dropped"]   # transient modes kept, persistent (RFI) modes dropped
```

`clean = U · diag(S̃) · Vᴴ` uses the data's own screen modes `U, Vᴴ` and the Gavish–Donoho optimal singular-value shrinkage `S̃` against the derived floor: idempotent, with the persistent narrowband (`φ_F ≤ φ_T`) modes dropped, and exact recovery **where the signal stands well clear of the derived floor**.

That last condition is worth stating plainly, because "noise-free" is not what governs it. The floor is estimated from the data, so a signal that fills the ordered axis contributes to its own floor estimate and is shrunk against it. Measured on an exactly rank-1 signal with **no noise added**, varying only its concentration in time:

| `σ_top / floor` | recovery error |
|---|---|
| 1.2e+07 | 5.4e-15 |
| 1972 | 1.8e-07 |
| 5.8 | 1.7e-02 |
| 2.6 | 4.4e-02 |

Recovery is exact once the top mode stands ~10³ above the floor, and the error peaks near the Gavish–Donoho threshold (`σ_top/floor ≈ 2.5`), where up to ~12% of amplitude is attenuated. A sparse burst reaches `σ_top/floor ≈ 5.6e+04` and recovers to 2e-10; a sine filling the whole record reaches only 2.45 and loses 12%. Check `Projection(W).sigma_top / Projection(W).noise_floor` if exactness matters.

**`clean` comes back in `W`'s own units.** The modes are read on the screen, and the screen is `project(normalize(W))` — each channel's median removed and its robust scale divided out — so the projection lands on the whitened grid. That whitening is inverted before `extract` returns, and `info["centre"]` / `info["scale"]` report the map that did it. So `clean` plots against `W`, and:

```python
removed = W - clean          # exactly what the filter discarded: noise and the dropped modes
```

Two things follow, and both are the definition rather than caveats:

- **The entropy fold is not inverted.** It is what makes the screen the screen, its shape is in `info["screen_shape"]`, and undoing it would synthesise cells that were never resolved.
- **`clean` is the per-channel baseline plus the resolved modes.** A channel's median is not a mode — `normalize` removes it before the SVD runs, so no cut was ever offered it. A persistent narrowband tone has its *modulation* dropped by the `φ_F ≤ φ_T` cut while its DC level stays in the baseline. Wanting that gone too is baseline estimation, a different read.

### Streaming, resume, splice

Feed frames from the first one and propagate; the exact-rate operator updates online.

```python
ap = Aperture(window=512)          # window bounds the optics snapshot
for frame in signal:               # each frame an F-vector (numpy or torch)
    ap.update(frame)

ap.rates()                         # long_range (slowest) & short_range (fastest) decay
ap.predict(frame)                  # one-step forecast A·x (A = ap.propagator_full())

s = ap.state()                     # export the full operator state …
ap2 = Aperture.from_state(s)       # … and resume exactly (bit-for-bit)

whole = a.splice(b)                # a, b: two Aperture streams → the concatenated-stream operator (exact at forgetting=1)
```

## What it reads

All reads are intrinsic — derived from `W` alone — and each is tied to a standard theorem.

### One screen, in focus

| | reads | what it says |
|---|---|---|
| **Scale** | `ap.phi`, `ap.magnification` = `1/phi` | the fill fraction and its reciprocal reach |
| | `ap.H_T`/`H_F`, `n_T`/`n_F`, `delta_T`/`delta_F` | per-axis entropy, matched grid width, cell scale |
| **Aperture area** | `ap.etendue` = `phi_F · phi_T` | the bounded 2-D area the screen carries |
| | `ap.space_bandwidth` = `n_F · n_T` | degrees of freedom it *can* carry — a capacity |
| **Coherence** | `ap.strehl`, `ap.phi_T`/`phi_F`, `ap.sigma_T`/`sigma_F` | dominant-mode power fraction, per-axis fills, leading singular values |
| | `Projection.coherence` | closed-form z against the exact row-permutation null |
| **Concentration** | `ap.concentration` | `intensity` (σ₁²), `focus` (axial), `resultant` (directional) |

`space_bandwidth` is a capacity, not a content: an unfolded screen reports `T·F` whatever sits on
it. What the screen actually fills is `ap.etendue * ap.space_bandwidth` — 1 for a single mode on
an **unfolded** screen, rising with the modes present. The 1 is the large-`T` limit: measured on a
rank-1 signal it reads 1.020 at `T = 256`, 1.012 at 512 and 1.006 at 1024.

**When the feature axis folds, the product reads `F_eff / F` of that, and the shortfall is
informative rather than an error.** `space_bandwidth` is the *unfolded* capacity while `etendue`
is measured on the screen, so a fold dilutes the product by exactly its own ratio. Measured over
200 rank-1 draws with random channel weights: 187 did not fold and read 1.0197; the 13 that did
read 0.381, 0.509 and 0.637 at `F_eff` of 3, 4 and 5 — each exactly `F_eff/F` of 1.0197, recovering
1.017–1.020 when multiplied back. So a product well below 1 on a single-mode signal says the
screen folded, and `ap.projection().F_eff` says by how much.

### The mode spectrum

`ap.spectral` reads the correlation spectrum against a derived noise floor:

- `contrast`, `top_share`, `resolved_modes`, `noise_floor`, `resolved_power` (summed eigenvalue
  excess above the floor), `dominance` = `(λ₁−1)/(F−1)`.
- the propagation constant `γ = α + iβ` — attenuation `α`, phase `β` — and `dispersion`.
- `ap.attenuation_interval(band)` returns a Weyl-certified interval for `α`.

The floor comes from a `null=` provider (`entroptics.null_providers`): the derived
Marchenko–Pastur / Johnstone default `mp` when `null` is unset, or the deterministic
data-derived Tukey fence via `null=null_providers.robust`.

### The diffraction limit

- `ap.decay` — the OTF, an FFT-free autocorrelation.
- `ap.a_delta` (entropy width `1/2^{H(C²)}`) and `ap.correlation_length` (the decay length ξ).
- `ap.mercer` — a model-free temporal-vs-spectral cross-check.
- `ap.rayleigh_shape_factor` (`g = xi * a_delta`), `ap.fresnel_number(window)`, `ap.shape_factor`
  (the Abbe factor `a_delta / phi_F`).
- `ap.decay_scatter` — the decay is a sum over per-channel autocovariances, so the channels are
  replicates of it and their disagreement measures the read's own uncertainty. No null, nothing
  subtracted. `noise_share` far below `tail_share` means the correlation is structure the channels
  agree on; the two converging means the width is scatter, and `a_delta` overstates the correlation
  length. The cure is more channels.

### How it moves, and at what scale

- **Dynamics**, `Aperture.rates()`: exact per-mode decay rates `α_k = −log|μ_k|` and frequencies
  `β_k = arg(μ_k)` from a streaming online-DMD / Koopman operator; splice-able and resumable.
  `Aperture.dominant_decay_rate` reads the slowest mode's rate `α_1 = −log|μ_1|` straight off that
  spectrum, isolating that one mode. A lag-window read returns a blend of all of them. `Aperture.propagator_full()` and
  `predict(x)` expose and apply the full one-step operator `A = P_yx · P_xx⁺`.
- **Multi-scale**, `ap.scale_profile()`: structure as a function of observation window —
  `K_signal`, `coherence`, `a_delta`, `phi_T` per window, plus `resolved_window` and
  `dominant_window` (in ordered-axis cells).
- **Sweep**, `Aperture.sweep()`: fix the aperture to a bounded capacity and sweep it where the
  coherence gate finds structure; noise-only patches are skipped. Returns per-band `span`, entropy
  `width` and tail decay `tau`, in dimensionless window samples.

### Recovering the signal

- **Projection**, `Projection` (`footprints`, `significance`, `read`), and `Aperture.tensor()` — a
  delay-embedded Tucker/HOSVD exposing within-window fine structure the averaged screen loses.
- **Filter**, `Aperture.extract()`: project the field onto its own resolved screen modes with
  Gavish–Donoho optimal shrinkage against the derived floor, dropping persistent narrowband
  (`phi_F <= phi_T`) interference. Returns `(clean, info)`; `clean` is a linear projection of the
  measured data — it synthesises nothing — and `info` carries `K_signal`, `contrast`, `coherence`
  and the kept/dropped modes. `clean` comes back in the **input's own units** — the per-channel
  whitening the modes are read through is inverted before returning.

### Other shapes of input

- **N-D fields**, `entroptics.fields`: `slabs(field, plane_axes)` and
  `over_planes(field, plane_axes, read=, reduce=)` reduce a higher-D field to the 2-D screen
  **while keeping each plane intact** (what feature reads need, since they use within-plane
  correlation); `pool(field, ordered_axis)` flattens sites as samples, which is what ordered reads
  need.
- **A stack at once**, `resolved_batch(X)` for `X: (B, T, F)`: the same resolved read over many
  frames, backend-optimal — numpy on the CPU (bit-identical to a per-frame `Projection`), a torch
  tensor on its device. Two cost tiers: the cheap survey gate (`K_signal`, `sigma_top`,
  `noise_floor`) always, the resolved `basis` and per-row `energy` on demand and only for frames
  that cleared the gate. `ResolvedScreen` / `ResolvedScreenBatch` are the stateful siblings for a
  screen you revisit (an LLM KV head across turns): append rows, refresh the basis lazily, resume
  from `state()`. `ResourceLimits` bounds threads, memory and GPU use; chunking is
  output-transparent.
- **Koopman lift**, `entroptics.lift`: `delay_embed(W, d)` builds Takens/Hankel coordinates
  `(T,F) → (T-d+1, d·F)`, and `koopman_lift(W, d)` fits the operator there — the path from a
  nonlinear or near-orthogonal trajectory to a linear one. An oscillator that resolves no modes
  raw resolves them after the lift.

### Records that are not a 2-D field

- **Spectral proximity**, `entroptics.proximity`: a magnitude-carrying, width-free digest of a
  frame's spectrum (`mp_spectrum`, `spectral_distance`, `bulk_edge`, `effective_width`) and a
  probe over a set of them (`SpectrumProbe`).
- **Symbol sequences**, `entroptics.sequence`: the ordered-axis reads on a symbol stream —
  `entropy_rate`, `block_entropies`, `lempel_ziv_rate`, `redundancy_rate`, `effective_length`,
  `surrogate_test`.

## Real data: fast radio bursts

`research/figures/frb_panel.py` applies the plain library, with zero tuning, to the public CHIME/FRB Catalog 1 waterfalls at native 16384-channel resolution. It supplies only observer facts — dead channels are **dropped, never zero-filled** — and passes the surviving channels to the aperture front door. The catalog waterfalls arrive already dedispersed; Entroptics adds no dedispersion, no derotation, and no domain physics of its own.

```python
from entroptics import Aperture

live = np.isfinite(W).all(axis=0) & (np.nanstd(W, axis=0) > 0)   # observer fact: the RFI mask
clean, info = Aperture(W[:, live], window=None).extract()        # everything downstream is the library
info["K_signal"], info["contrast"], info["coherence"]
```

`extract` is the Gavish-Donoho projection onto the resolved modes with the `φ_F > φ_T` persistent-structure cut: the noise sea is attenuated and persistent narrowband interference removed, with the burst morphology intact. The read comes back on the waterfall's own amplitude scale, so the "Removed" panel is a plain `wf - clean` with nothing rescaled by hand. The per-burst reads behind the figure are in [`research/figures/frb_panel.csv`](research/figures/frb_panel.csv), and the method is §12.1 of [the paper](https://github.com/Agience/entroptics/blob/main/research/PAPER.pdf).

## Why it's principled

Every read is a classical result specialised to finite, discrete data. The backbone: a signal's autocorrelation *is* its optical transfer function (**Wiener–Khinchin** → **Fourier optics**), so the diffraction limit is one over that OTF's bandwidth (**Abbe/Rayleigh**). The full derivation, and the Lean-checked lemmas, are in the [paper](https://github.com/Agience/entroptics/blob/main/research/PAPER.pdf).

## Backends & determinism

- **One code path, numpy or torch.** Feed a numpy array → runs on CPU; feed a torch tensor → runs on its device (GPU), staying on-device; torch is imported lazily only when a tensor appears.
- **Deterministic.** The coherence is a closed-form permutation-null z-score, so results are reproducible and bit-identical across numpy and torch (to floating-point round-off).
- **Complex-safe** end to end. Reads take the record as given and infer nothing from it — whether you hold a field or an intensity is a fact about your instrument, not about the sign of your samples. For the incoherent read of an amplitude record, pass the intensity (`decay(W ** 2)`). `Aperture.sweep()` finds and reads its bands on a field like any other record — `span`, `coherence`, `contrast`, `K`, `noise_floor`, `phi_T/F` — and the span takes you back to the original vectors, complex intact. Its `peak`, `width` and `tau_decay` come off a per-sample brightness, which a field does not carry, so those three return `-1`/`NaN`; pass `abs(W)` or `abs(W)**2` when you want them.

## Axis convention

Every input `W` has shape `(T, F)`:

- axis-0 (rows, `_T`) is the **ordered** / evolution axis, "time" is a *role*, not literal physics;
- axis-1 (cols, `_F`) is the **feature** / channel axis, "frequency" is likewise a role.

Any 2-D array with one ordered axis works.

## Absent data

Mark what was never observed — either as `NaN` in `W`, or with a boolean `mask` where `True` means
*not observed*. The two say the same thing, and a mask works over a value that is finite and wrong
(a saturated sample, an RFI-flagged channel that still holds a reading):

```python
ap = Aperture(W, mask=flags)        # flags[t, f] True  ->  that cell was not observed
```

Every read then divides by the **measured extent** — the rows and columns that carry at least one
observation — not by the array's shape. A read taken through a mask equals the same frame with
those channels deleted, to floating point.

Absence is not an observation of zero. Substituting `0` for an unobserved channel is a different
frame: zero is a real reading of no power, it belongs in the extent, and it widens the axis the
signal is scored against. Mark it absent and the reads are unmoved; fill it with zeros and they move, correctly.

## Per-channel structure

Multiplying the whole record by a constant changes nothing. Every **dimensionless** read — a
z-score, a share, a mode count, a correlation, a focus — is identical from `1e+30` down to
`1e-140`, and every **dimensioned** one scales exactly as its own dimension (a noise floor as
`c`, a variance as `c²`). A constant baseline changes nothing either; the reads are taken on the
centred block.

That is a property test rather than a claim: [`src/tests/test_scale_invariance.py`](src/tests/test_scale_invariance.py)
sweeps both halves, and carries a control that each read still separates structure from noise —
a read that returned a constant would be trivially scale-free and useless, so invariance alone
is not enough to pass.

*Per-channel* structure is different. A channel with ten times another channel's gain is, to the
instrument, a channel carrying ten times the power — there is nothing in the data that says whether
that is the sky or the amplifier. So the reads describe it, because it is part of the frame they
were handed. `phi_F` is the exception: the feature-axis correlation has unit diagonal, so it
rescales every channel by construction.

If per-channel gain or baseline is instrumental, remove it before you read:

```python
from entroptics.entropy import normalize
ap = Aperture(normalize(W))          # per-channel median removed, robust scale equalised
```

## Tests

```bash
pip install -e ".[dev]"
pytest
```

The suite (`src/tests/`) pins the full optics read as a golden contract, checks numpy↔torch parity, the mathematical invariants (étendue = φ_F·φ_T, exact decay-rate recovery, PSD autocovariance, axial-vs-directional concentration, the read-side filter's exact projection and idempotence), round-trips (tensor reconstruct, factor pack/unpack), determinism, and degenerate-input robustness.

Four files pin the claims this README makes about *how* the reads behave, each with a control that
makes it able to fail:

| file | what it holds |
|---|---|
| `test_scale_invariance.py` | dimensionless reads identical across `1e+30`–`1e-140`; dimensioned ones exact to `c^dim`; and every read must still separate structure from noise, so invariance alone cannot pass |
| `test_fold_band_calibration.py` | `fold_band`'s stated false-fold rate on pure noise, and the power that must survive it — the significance and sufficiency terms are separated, because calibration alone cannot test the second |
| `test_coupling_reduces_to_pearson.py` | at one shared coordinate `coupling.strength` **is** Pearson's r and `z` is `r√(T−1)`; what the read adds there is the decision layer, and that is asserted too |
| `test_resolved_modes_is_not_a_rank.py` | `resolved_modes` counts modes above a **noise** floor — it is not a matrix rank and not a model-order selector; pinned so nothing is built on it expecting one |

## Formal certification

The governing lemmas of the theory ([`research/PAPER.pdf`](https://github.com/Agience/entroptics/blob/main/research/PAPER.pdf)) are **machine-checked in Lean 4 / Mathlib** ([`research/lean/`](https://github.com/Agience/entroptics/tree/main/research/lean), 44 theorems): the fill-fraction and Strehl bounds, positive-semidefiniteness of the biased autocovariance (peak-at-zero-lag OTF), the exact permutation-null mean of the coherence, the Weyl-certified attenuation interval, axial≠directional concentration, and exact decay-rate recovery + additive splicing. `lake build` compiles with **no `sorry`**, resting only on Mathlib's standard axioms.

## Documentation

- **[API reference](docs/API.md)** — every public name, with its signature and what it does. Generated from the library itself by `python docs/generate_api.py`, so it cannot drift from the code.
- **[`research/PAPER.pdf`](https://github.com/Agience/entroptics/blob/main/research/PAPER.pdf)** — the paper: definitions, proofs, and the provenance of every constant (§11.1). Typeset, with the mathematics rendered. Same text in Markdown: [`research/PAPER.md`](research/PAPER.md).
- **[`research/validation/RESULTS.md`](research/validation/RESULTS.md)** — what each read recovers against a planted ground truth.
- **[`research/supplemental/`](research/supplemental/)** — the applications and statistics papers, each with a `reproduce.py` and a `verify.py`.

## Getting help

- **Questions and bug reports:** open an issue at [github.com/Agience/entroptics/issues](https://github.com/Agience/entroptics/issues). A report is most useful with the array shape, the backend (numpy or torch), and the read you called.
- **Security issues:** email **connect@agience.ai** rather than opening a public issue.
- **Contributions:** see [CONTRIBUTING.md](CONTRIBUTING.md).
- **Conduct:** participation is governed by the [Code of Conduct](CODE_OF_CONDUCT.md).

## What runs without a download

The library, its test suite and the full validation suite need **no external data** — every input is generated from a fixed seed:

```bash
pytest                                    # 675 tests
python research/validation/run_all.py     # 19 experiments -> RESULTS.md
```

Two figure routines are the exception. `research/figures/frb_panel.py` and `frb_spotcheck.py` read the **CHIME/FRB Catalog 1 waterfalls**, a separate public download from the CANFAR archive (CISTI.CANFAR/21.0007) that is not part of this repository and is far too large to ship. They refuse, with instructions, if it is not configured — see [`research/supplemental/frb/README.md`](research/supplemental/frb/README.md) for the fetch and the one line of configuration. Everything else in the repository runs from a clean checkout.

## Contributing

Bug reports, validation experiments, backend-parity fixes, and new reads (with their theorems) are welcome — see [CONTRIBUTING.md](CONTRIBUTING.md). The short version: the library is parameter-free, deterministic, and numpy-only at the core, and contributions must keep it that way — every constant needs a provenance, every claim needs a theorem, and the golden-contract tests must stay bit-identical across backends.

## Star history

<a href="https://www.star-history.com/?repos=Agience%2Fentroptics&type=date&legend=top-left">
 <picture>
   <source media="(prefers-color-scheme: dark)" srcset="https://api.star-history.com/chart?repos=Agience/entroptics&type=date&theme=dark&legend=top-left&sealed_token=DRxmKEUu-jgYDGrGi5K7vVFwrww1YJiMFU2_nv85yGjwPbsvhmTkOSlVv2aQQGkVDHXd2jlGQnjZDHbYYOXwfObR6iE9wTeV5jyplb30xZ3GFdD1ebDZIAonKgIvYBZ5vH8Z7T-2lSgsWrktUeeoPUdPPRELXBa4LY0ILQatLXzOLeWpq4dU5eVFXTcH" />
   <source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/chart?repos=Agience/entroptics&type=date&legend=top-left&sealed_token=DRxmKEUu-jgYDGrGi5K7vVFwrww1YJiMFU2_nv85yGjwPbsvhmTkOSlVv2aQQGkVDHXd2jlGQnjZDHbYYOXwfObR6iE9wTeV5jyplb30xZ3GFdD1ebDZIAonKgIvYBZ5vH8Z7T-2lSgsWrktUeeoPUdPPRELXBa4LY0ILQatLXzOLeWpq4dU5eVFXTcH" />
   <img alt="Star History Chart" src="https://api.star-history.com/chart?repos=Agience/entroptics&type=date&legend=top-left&sealed_token=DRxmKEUu-jgYDGrGi5K7vVFwrww1YJiMFU2_nv85yGjwPbsvhmTkOSlVv2aQQGkVDHXd2jlGQnjZDHbYYOXwfObR6iE9wTeV5jyplb30xZ3GFdD1ebDZIAonKgIvYBZ5vH8Z7T-2lSgsWrktUeeoPUdPPRELXBa4LY0ILQatLXzOLeWpq4dU5eVFXTcH" />
 </picture>
</a>

Security issues: email **connect@agience.ai** rather than opening a public issue.

Licensed under Apache-2.0 — see [`LICENSE.md`](https://github.com/Agience/entroptics/blob/main/LICENSE.md),
[`NOTICE`](https://github.com/Agience/entroptics/blob/main/NOTICE), [`PATENTS.md`](https://github.com/Agience/entroptics/blob/main/PATENTS.md)
and [`PLEDGE.md`](https://github.com/Agience/entroptics/blob/main/PLEDGE.md).

## Declaration of generative AI use

The author used Anthropic's Claude Opus (versions 4.8 and 5) in the preparation of this work. Its
contribution was to write code, and to generate and validate content. The ideas, the construction
and the claims are the author's. No other generative AI tool was used. The author reviewed and
edited all output and takes full responsibility for the content of this publication.
