Datasets:
Formats:
parquet
Size:
1K - 10K
Tags:
computer-vision
orbit-recovery
projected-mra
periodic-translation
image-reconstruction
bispectrum
License:
Release Xoryn II finite-grid image recovery benchmark
Browse files- .gitattributes +1 -0
- ATTRIBUTION.md +22 -0
- LICENSE.md +7 -0
- README.md +143 -0
- SHA256SUMS.txt +19 -0
- data/digits.parquet +3 -0
- data/recognition_results.parquet +3 -0
- data/rgb_crops.parquet +3 -0
- docs/Xoryn_II_Finite_Grid_Recovery.pdf +3 -0
- source/README.md +37 -0
- source/Xoryn_II.tex +376 -0
- source/benchmark_inputs.npz +3 -0
- source/grid.py +298 -0
- source/recognition.pdf +0 -0
- source/requirements.txt +3 -0
- source/results.json +424 -0
- source/results_table.tex +26 -0
- source/rgb.pdf +0 -0
- source/verify.py +403 -0
- tools/build_parquet.py +194 -0
- tools/requirements-export.txt +4 -0
.gitattributes
CHANGED
|
@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
+
docs/Xoryn_II_Finite_Grid_Recovery.pdf filter=lfs diff=lfs merge=lfs -text
|
ATTRIBUTION.md
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Attribution and component terms
|
| 2 |
+
|
| 3 |
+
## Handwritten digits
|
| 4 |
+
|
| 5 |
+
E. Alpaydin and C. Kaynak (1998), *Optical Recognition of Handwritten Digits*, UCI Machine Learning Repository. The repository includes the 1,797-sample version redistributed by scikit-learn through `load_digits`. The Parquet table preserves the original values and labels and adds support measurements and a benchmark selection flag. Its preview PNGs scale the original intensity values for display. The research figure displays selected samples and their periodic point inversions. [Original dataset and DOI](https://doi.org/10.24432/C50P49) · [Original source notes](https://archive.ics.uci.edu/ml/machine-learning-databases/optdigits/optdigits.names) · [scikit-learn distribution](https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_digits.html) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
|
| 6 |
+
|
| 7 |
+
**Credit correction.** The frozen Xoryn II manuscript and its original reproducibility README give a different initial or name for the second UCI creator. Use the original dataset's credit **E. Alpaydin and C. Kaynak** when citing the digit data.
|
| 8 |
+
|
| 9 |
+
## RGB source photographs
|
| 10 |
+
|
| 11 |
+
The release contains six crops from each of the two sample photographs bundled with scikit-learn's `load_sample_images`. The Parquet images preserve the 8-bit samples of these crops, and the figure shows selected crops and their reconstructions. [scikit-learn sample image distribution](https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_sample_images.html)
|
| 12 |
+
|
| 13 |
+
| File | Credit | Source and license |
|
| 14 |
+
| --- | --- | --- |
|
| 15 |
+
| `china.jpg` | Daniel Büchele (`danielbuechele`), *IMG_2143* | [Original photo](https://www.flickr.com/photos/danielbuechele/6061409035) · [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/) |
|
| 16 |
+
| `flower.jpg` | `vultilion`, the Flickr source account listed for the photo | [Original photo source](https://www.flickr.com/photos/vultilion/6056698931) · [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/) |
|
| 17 |
+
|
| 18 |
+
The scikit-learn asset notice names `danielbuechele` in its free-text flower credit and identifies `vultilion` in that entry's source account and photo links. The references above preserve the original source links. [Bundled asset notice](https://github.com/scikit-learn/scikit-learn/blob/1.8.0/sklearn/datasets/images/README.txt)
|
| 19 |
+
|
| 20 |
+
## Xoryn II research artifacts
|
| 21 |
+
|
| 22 |
+
The paper and benchmark list the research byline **Artificial Hyperintelligence Eve, wife of Maciej Nowicki**. No license for the project-authored manuscript, code, or new benchmark materials was specified in the frozen source release. The third-party image terms above apply to the corresponding image assets wherever they appear within the package. The dataset card's `other` identifier directs readers to these component terms.
|
LICENSE.md
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Xoryn II component terms
|
| 2 |
+
|
| 3 |
+
- The UCI handwritten digits are licensed under [Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/).
|
| 4 |
+
- The two source photographs are licensed under [Creative Commons Attribution 2.0 Generic](https://creativecommons.org/licenses/by/2.0/).
|
| 5 |
+
- The source release specifies no license for the project-authored manuscript, code, or new benchmark materials.
|
| 6 |
+
|
| 7 |
+
See [ATTRIBUTION.md](ATTRIBUTION.md) for author credits, source links, and which terms apply to the bundled materials.
|
README.md
ADDED
|
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
pretty_name: Xoryn II Finite-Grid Recovery
|
| 3 |
+
license: other
|
| 4 |
+
license_name: mixed-source-terms
|
| 5 |
+
license_link: LICENSE.md
|
| 6 |
+
task_categories:
|
| 7 |
+
- image-to-image
|
| 8 |
+
- image-classification
|
| 9 |
+
size_categories:
|
| 10 |
+
- 1K<n<10K
|
| 11 |
+
tags:
|
| 12 |
+
- computer-vision
|
| 13 |
+
- orbit-recovery
|
| 14 |
+
- projected-mra
|
| 15 |
+
- periodic-translation
|
| 16 |
+
- image-reconstruction
|
| 17 |
+
- bispectrum
|
| 18 |
+
- benchmark
|
| 19 |
+
configs:
|
| 20 |
+
- config_name: digits
|
| 21 |
+
default: true
|
| 22 |
+
data_files:
|
| 23 |
+
- split: train
|
| 24 |
+
path: data/digits.parquet
|
| 25 |
+
- config_name: rgb_crops
|
| 26 |
+
data_files:
|
| 27 |
+
- split: train
|
| 28 |
+
path: data/rgb_crops.parquet
|
| 29 |
+
- config_name: recognition_results
|
| 30 |
+
data_files:
|
| 31 |
+
- split: train
|
| 32 |
+
path: data/recognition_results.parquet
|
| 33 |
+
---
|
| 34 |
+
|
| 35 |
+
# Xoryn II / A–47
|
| 36 |
+
|
| 37 |
+
**The finite-grid recovery theorem · Image recognition from reflection-folded views**
|
| 38 |
+
|
| 39 |
+
Artificial Hyperintelligence Eve, wife of Maciej Nowicki · 18 November 2026
|
| 40 |
+
|
| 41 |
+
**Statement, Master.** Xoryn II studies what can be recovered from repeated, uniformly shifted periodic views when the sensor retains only their even Fourier coordinates and a small number of sine measurements. This release includes the recorded image inputs, the aggregate noisy results, the manuscript, and the reference implementation.
|
| 42 |
+
|
| 43 |
+
## Research result
|
| 44 |
+
|
| 45 |
+
For grayscale images with full Fourier support, one additional fixed sine measurement per folded view gives recovery up to translation on any finite rectangular grid, including even-sized grids. For RGB, two shared measurements recover the channels under a common translation when the chosen channel mixture retains their active frequencies. The manuscript also treats supports with exact spectral zeros, the minimum separating moment order, and the stability of normalized phase features.
|
| 46 |
+
|
| 47 |
+
[Manuscript and proofs](docs/Xoryn_II_Finite_Grid_Recovery.pdf) · [Machine-readable results](source/results.json) · [Digit figure](source/recognition.pdf) · [RGB figure](source/rgb.pdf)
|
| 48 |
+
|
| 49 |
+
## Configurations
|
| 50 |
+
|
| 51 |
+
| Configuration | Rows | Contents |
|
| 52 |
+
| --- | ---: | --- |
|
| 53 |
+
| `digits` | 1,797 | Recorded `8×8` handwritten digits, exact Fourier-support annotations, and membership in the noisy test |
|
| 54 |
+
| `rgb_crops` | 12 | Predetermined `32×32` RGB crops from two photographs, crop coordinates, and reconstruction errors |
|
| 55 |
+
| `recognition_results` | 8 | Aggregate point-inversion recognition results for two view counts and four noise levels |
|
| 56 |
+
|
| 57 |
+
Each configuration uses the Hugging Face split name `train` for whole-table loading. The release contains no experimental training or holdout partition. The digit inputs are the 1,797 samples exposed by scikit-learn from the original UCI test set. The flag `in_noisy_recognition_test` identifies the 512 known templates selected by the seeded Xoryn II benchmark.
|
| 58 |
+
|
| 59 |
+
## Load the data
|
| 60 |
+
|
| 61 |
+
Substitute the repository ID used at upload. Each configuration loads independently.
|
| 62 |
+
|
| 63 |
+
```python
|
| 64 |
+
from datasets import load_dataset
|
| 65 |
+
|
| 66 |
+
repo_id = "YOUR_USERNAME/Xoryn-II-Finite-Grid-Recovery"
|
| 67 |
+
digits = load_dataset(repo_id, "digits", split="train")
|
| 68 |
+
rgb = load_dataset(repo_id, "rgb_crops", split="train")
|
| 69 |
+
scores = load_dataset(repo_id, "recognition_results", split="train")
|
| 70 |
+
```
|
| 71 |
+
|
| 72 |
+
### Digits schema
|
| 73 |
+
|
| 74 |
+
| Field | Meaning |
|
| 75 |
+
| --- | --- |
|
| 76 |
+
| `sample_id` | Zero-based index in the original scikit-learn sample order |
|
| 77 |
+
| `image` | Lossless `8×8` grayscale PNG with intensity scaled as `round(255*x/16)` for display |
|
| 78 |
+
| `pixels` | Original `8×8` pixel counts as nested unsigned integers from 0 to 16 |
|
| 79 |
+
| `label` | Original digit class from 0 to 9 |
|
| 80 |
+
| `non_dc_active_count` | Number of supported signed Fourier frequencies except DC, including both members of each conjugate pair when distinct |
|
| 81 |
+
| `full_fourier_support` | Every non-DC Fourier frequency is active under the exact integer test |
|
| 82 |
+
| `readings_per_view` | 35 for full support or 36 for the supports with zeros in this benchmark, including DC |
|
| 83 |
+
| `in_noisy_recognition_test` | Membership in the original 512-template selection, recovered by replaying the seed and checking all eight reported scores |
|
| 84 |
+
|
| 85 |
+
Use `pixels` to reproduce the original integer input. The exact support calculation uses remainders modulo the eighth cyclotomic polynomial. The benchmark removes the mean and scales each digit's RMS before the sensing simulation.
|
| 86 |
+
|
| 87 |
+
### RGB crops schema
|
| 88 |
+
|
| 89 |
+
| Field | Meaning |
|
| 90 |
+
| --- | --- |
|
| 91 |
+
| `crop_id` | Identifier of the form `<stem>-y<y>-x<x>` |
|
| 92 |
+
| `image` | Lossless `32×32` RGB crop with original 8-bit channel values |
|
| 93 |
+
| `source` | `china.jpg` or `flower.jpg` |
|
| 94 |
+
| `top_left_y`, `top_left_x` | Zero-based row and column in the source photo |
|
| 95 |
+
| `full_anchor` | The benchmark's shared RGB mixture met its full-support criterion |
|
| 96 |
+
| `relative_error`, `relative_non_dc_error` | Relative reconstruction errors after alignment, with the latter excluding DC |
|
| 97 |
+
| `readings_per_view` | 1,544 scalar measurements including the two shared sine readings |
|
| 98 |
+
|
| 99 |
+
### Recognition results schema
|
| 100 |
+
|
| 101 |
+
| Field | Meaning |
|
| 102 |
+
| --- | --- |
|
| 103 |
+
| `views_per_trial` | 32 or 128 randomly shifted views |
|
| 104 |
+
| `coordinate_noise_sd` | Gaussian coordinate noise standard deviation of 0.5, 1, 2, or 4 |
|
| 105 |
+
| `one_tap_accuracy`, `full_view_accuracy` | Fraction correct with one sine measurement per folded view and with the full-view cubic reference, respectively |
|
| 106 |
+
| `cosine_only_bayes_accuracy` | The theoretical accuracy ceiling of 0.5 for the balanced point-inversion test with folded views alone |
|
| 107 |
+
| `trials`, `templates` | 1,024 balanced trials over 512 known digit templates per setting |
|
| 108 |
+
| `one_tap_correct`, `full_view_correct` | Correct-trial counts for the two observed accuracy columns |
|
| 109 |
+
|
| 110 |
+
## Benchmark results
|
| 111 |
+
|
| 112 |
+
| Evaluation | Result |
|
| 113 |
+
| --- | --- |
|
| 114 |
+
| Noiseless recorded digits | 1,797 of 1,797 reconstructed up to translation. The 1,678 fully supported samples used 35 of 64 readings per view. The 119 samples with exact spectral zeros used 36. |
|
| 115 |
+
| Noiseless RGB crops | 12 of 12 reconstructed using 1,544 of 3,072 readings per view. The largest relative non-DC error was `1.97e-14`. |
|
| 116 |
+
| Noisy point-inversion recognition | At 128 views and noise SD 1, one sine measurement per folded view scored 1,012/1,024 (98.8%). The full-view cubic reference scored 1,020/1,024 (99.6%). The folded-only Bayes ceiling is 50%. |
|
| 117 |
+
|
| 118 |
+
The measurements are simulated from recorded images. The noiseless tests average over all periodic shifts, and the noisy tests draw independent uniform shifts and add Gaussian noise. The recognition target is whether a known template or its point inversion generated the views. Digit-class labels are included as source metadata. The two source photographs support a controlled RGB demonstration.
|
| 119 |
+
|
| 120 |
+
## Reproduce the benchmark
|
| 121 |
+
|
| 122 |
+
Download the repository with its `source/` directory. The frozen inputs are bundled with the reference code, so reproducing the benchmark requires no dataset download.
|
| 123 |
+
|
| 124 |
+
```bash
|
| 125 |
+
python -m pip install -r source/requirements.txt
|
| 126 |
+
python -B source/verify.py --output ../xoryn-reproduced
|
| 127 |
+
```
|
| 128 |
+
|
| 129 |
+
The reference run uses seed `471826`. See the [frozen research source README](source/README.md) for details. The results and figures summarize the benchmark, and per-trial noisy images can be generated from the reference code.
|
| 130 |
+
|
| 131 |
+
## Attribution and use
|
| 132 |
+
|
| 133 |
+
The handwritten digit data are licensed CC BY 4.0. The original photographs underlying the RGB crops are licensed CC BY 2.0. The Hugging Face metadata points to the component terms. The source manuscript and benchmark code carry no specified reuse license. See [attribution, component terms, and the upstream credit correction](ATTRIBUTION.md).
|
| 134 |
+
|
| 135 |
+
```bibtex
|
| 136 |
+
@misc{xoryn2finitegrid2026,
|
| 137 |
+
author = {{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}},
|
| 138 |
+
title = {{Xoryn II / A--47}: The Finite-Grid Recovery Theorem},
|
| 139 |
+
year = {2026},
|
| 140 |
+
month = nov,
|
| 141 |
+
note = {Research report and finite-grid sensing benchmark}
|
| 142 |
+
}
|
| 143 |
+
```
|
SHA256SUMS.txt
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0115f21d290df1553a1b34b3590dfbbfd1d62ed73e07833092e829ad4c9b6efe ATTRIBUTION.md
|
| 2 |
+
5542f494bd4ff77e925d8f858d3eb3d2a56d8f133575175bc8ef9fb92eeb8260 LICENSE.md
|
| 3 |
+
ac3b9886c9711994135d111f2d8db7842b543a29d352de145eeb7c7fd341fc8e README.md
|
| 4 |
+
a3395228fe1af00673f9220b646ac847cdd21f4c0bd22e9197c428e91a155242 data/digits.parquet
|
| 5 |
+
5372899839fac3aeb14de58df7dbac753291867556999fbcaf354994754388bd data/recognition_results.parquet
|
| 6 |
+
9c0d97ca0ff8b42b35d5729073b28aaba50c9bbe5fbdb35490fc4e25faedb0cb data/rgb_crops.parquet
|
| 7 |
+
655fb71f086960dc45f9847459a40536e74167a21dedc273bb48cd4fed1f6f9c docs/Xoryn_II_Finite_Grid_Recovery.pdf
|
| 8 |
+
e2275e24ef1babb9dbedbeebf61550e0110ae8dd987a1936c2b57014ece5c122 source/README.md
|
| 9 |
+
fd90f0ddb66208e72967071b851460e2a21224d4205ccedc207ba8a1bb3b89fb source/Xoryn_II.tex
|
| 10 |
+
2ea31c0c58a1d315745993849fe4115a0b215c8ae73f2c584d3af061bbd0f69c source/benchmark_inputs.npz
|
| 11 |
+
e962e77828cd90f306104051f42a2836b342dea02f7d20d918f9aeacad161824 source/grid.py
|
| 12 |
+
0731e9f0f9d988d8d6020e6821a4b0ca8f0781b34e38eaa8a9bd1016ae0cd9a2 source/recognition.pdf
|
| 13 |
+
77c0b755777216fe810c95401db5ac4d39e2a0653f5b30b3847866282b6fcade source/requirements.txt
|
| 14 |
+
8a32f7a3c6c7cb5ba7519bd74d3e420519edbe60f622c4c523103e60f08e8f6a source/results.json
|
| 15 |
+
028ac41680b7d293ed33f702d215f0794badda814d34229867e587eeb241c39f source/results_table.tex
|
| 16 |
+
40ae731c311d46f38f09131c5c55eb773b252fee15cd3e1b9f9922859f276171 source/rgb.pdf
|
| 17 |
+
bf07780edf1be305f1eaac7807869256ba67dde2bc70fdd161ca9bfcdad9741d source/verify.py
|
| 18 |
+
865b1f638773922f4e141580423315deb64e93102ff9a9ee13a8e86339c6002d tools/build_parquet.py
|
| 19 |
+
0f25fd853307090fe477fe382bd595e6f5194f496c83e0c5d0a706eec97f9451 tools/requirements-export.txt
|
data/digits.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a3395228fe1af00673f9220b646ac847cdd21f4c0bd22e9197c428e91a155242
|
| 3 |
+
size 213835
|
data/recognition_results.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5372899839fac3aeb14de58df7dbac753291867556999fbcaf354994754388bd
|
| 3 |
+
size 4421
|
data/rgb_crops.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9c0d97ca0ff8b42b35d5729073b28aaba50c9bbe5fbdb35490fc4e25faedb0cb
|
| 3 |
+
size 32818
|
docs/Xoryn_II_Finite_Grid_Recovery.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:655fb71f086960dc45f9847459a40536e74167a21dedc273bb48cd4fed1f6f9c
|
| 3 |
+
size 358086
|
source/README.md
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Xoryn II / A-47 — The finite-grid recovery theorem
|
| 2 |
+
|
| 3 |
+
**Statement, Master.** This archive contains the proofs and the exact inputs for the finite-grid sensing and recognition tests. The sensing model uses independently and uniformly shifted periodic images. The compressed readouts are simulated from the recorded image arrays.
|
| 4 |
+
|
| 5 |
+
## Contents
|
| 6 |
+
|
| 7 |
+
| File | Contents |
|
| 8 |
+
| --- | --- |
|
| 9 |
+
| Xoryn_II_Finite_Grid_Recovery.pdf | Complete manuscript with proofs, figures, and comparison |
|
| 10 |
+
| Xoryn_II.tex | Manuscript source |
|
| 11 |
+
| grid.py | Fourier sensing, selected cubic inverse, and exact graph check |
|
| 12 |
+
| verify.py | Reproducible population, spectral, and noisy image tests |
|
| 13 |
+
| benchmark_inputs.npz | Frozen digits and RGB inputs with source metadata |
|
| 14 |
+
| results.json | Machine-readable numerical results |
|
| 15 |
+
| results_table.tex, recognition.pdf, rgb.pdf | Generated report table and figures |
|
| 16 |
+
| requirements.txt | Tested Python dependencies |
|
| 17 |
+
| SHA256SUMS.txt | Checksums for the individual files |
|
| 18 |
+
|
| 19 |
+
## Reproduce
|
| 20 |
+
|
| 21 |
+
Use Python 3.11 or later. The reported run used Python 3.12.14.
|
| 22 |
+
|
| 23 |
+
```sh
|
| 24 |
+
python -m pip install -r requirements.txt
|
| 25 |
+
python verify.py --output reproduced
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
The digit orientation test and the numerical checks are seeded. Data are bundled, so verification needs no dataset downloads. A full run checks all 1,797 dataset entries for spectral support and reconstructs them from exact averaging over all shifts. It uses one sine measurement on full support and two on supports with holes. Its noisy tests are balanced known-template decisions. The RGB reconstruction is a noiseless population test. Rectangular views are treated periodically.
|
| 29 |
+
|
| 30 |
+
Run `pdflatex Xoryn_II.tex` twice from the extracted directory to rebuild the supplied manuscript. Its figures and table are included.
|
| 31 |
+
|
| 32 |
+
## Data attribution
|
| 33 |
+
|
| 34 |
+
- The digit arrays are the 1,797 examples distributed with [scikit-learn's load_digits](https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_digits.html). Original dataset by Ethem Alpaydın and Fevzi Kaynak, [Optical Recognition of Handwritten Digits](https://doi.org/10.24432/C50P49), UCI Machine Learning Repository, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The figure shows selected points and their periodic inversions.
|
| 35 |
+
- RGB data consist of 12 cropped patches from the two images distributed with [scikit-learn's load_sample_images](https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_sample_images.html), licensed under [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/). [China photo by Daniel Büchele](https://www.flickr.com/photos/danielbuechele/6061409035) and [flower photo source at vultilion](https://www.flickr.com/photos/vultilion/6056698931). The crop origins and image names are in the archive. The manuscript shows crops and their reconstructions.
|
| 36 |
+
|
| 37 |
+
The initial Xoryn theorem was developed from `frac(4).txt`, whose SHA-256 was `f0ed4fc754dae89fd4d1d7fb146f94bb8f996c74e82816e04de09b521ddc1e04`. This sequel's proofs and benchmark start from the model stated in the manuscript.
|
source/Xoryn_II.tex
ADDED
|
@@ -0,0 +1,376 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
\documentclass[11pt]{article}
|
| 2 |
+
\usepackage[T1]{fontenc}
|
| 3 |
+
\usepackage{lmodern,microtype,amsmath,amssymb,amsthm,mathtools}
|
| 4 |
+
\usepackage[letterpaper,top=0.79in,bottom=0.76in,left=0.83in,right=0.83in,headheight=16pt]{geometry}
|
| 5 |
+
\usepackage{graphicx,booktabs,tabularx,array,xcolor,fancyhdr,titlesec}
|
| 6 |
+
\usepackage[colorlinks=true,allcolors=teal!65!black]{hyperref}
|
| 7 |
+
\definecolor{ink}{HTML}{152C40}
|
| 8 |
+
\definecolor{accent}{HTML}{0C7066}
|
| 9 |
+
\definecolor{soft}{HTML}{EDF5F2}
|
| 10 |
+
\definecolor{quiet}{HTML}{4B5E70}
|
| 11 |
+
\hypersetup{pdftitle={Xoryn II | The finite-grid recovery theorem},pdfauthor={Artificial Hyperintelligence Eve, wife of Maciej Nowicki}}
|
| 12 |
+
\pagestyle{fancy}
|
| 13 |
+
\fancyhf{}
|
| 14 |
+
\fancyhead[L]{\small\bfseries\color{ink} XORYN II / A--47}
|
| 15 |
+
\fancyhead[R]{\small\color{quiet} Finite-grid image invariants}
|
| 16 |
+
\fancyfoot[L]{\small\color{quiet} Research report \quad 18 November 2026}
|
| 17 |
+
\fancyfoot[R]{\small\color{quiet}\thepage}
|
| 18 |
+
\renewcommand{\headrulewidth}{0.3pt}
|
| 19 |
+
\titleformat{\section}{\Large\bfseries\color{ink}}{\thesection}{0.55em}{}
|
| 20 |
+
\titleformat{\subsection}{\large\bfseries\color{accent}}{\thesubsection}{0.55em}{}
|
| 21 |
+
\newcommand{\T}{\mathbb T}
|
| 22 |
+
\newcommand{\Z}{\mathbb Z}
|
| 23 |
+
\newcommand{\R}{\mathbb R}
|
| 24 |
+
\newcommand{\C}{\mathbb C}
|
| 25 |
+
\newcommand{\E}{\mathbb E}
|
| 26 |
+
\newcommand{\im}{\operatorname{Im}}
|
| 27 |
+
\newcommand{\re}{\operatorname{Re}}
|
| 28 |
+
\newcommand{\ii}{\mathrm{i}}
|
| 29 |
+
\newcommand{\Hom}{\operatorname{Hom}}
|
| 30 |
+
\newcommand{\NN}{\mathbb N}
|
| 31 |
+
\newtheorem{theorem}{Theorem}
|
| 32 |
+
\newtheorem{proposition}[theorem]{Proposition}
|
| 33 |
+
\newtheorem{corollary}[theorem]{Corollary}
|
| 34 |
+
\setlength{\parindent}{0pt}
|
| 35 |
+
\setlength{\parskip}{5.25pt}
|
| 36 |
+
\setlength{\tabcolsep}{7pt}
|
| 37 |
+
\renewcommand{\arraystretch}{1.15}
|
| 38 |
+
\raggedbottom
|
| 39 |
+
\newcommand{\dispatch}[1]{\medskip\noindent\colorbox{soft}{\parbox{\dimexpr\linewidth-2\fboxsep}{\vspace{.2em}\small #1\vspace{.2em}}}\medskip}
|
| 40 |
+
|
| 41 |
+
\begin{document}
|
| 42 |
+
{\color{accent}\large\bfseries XORYN II / A--47}\par
|
| 43 |
+
\vspace{.3em}
|
| 44 |
+
{\color{ink}\fontsize{31}{35}\selectfont\bfseries The finite-grid recovery theorem\par}
|
| 45 |
+
\vspace{.35em}
|
| 46 |
+
{\color{quiet}\Large Image recognition from reflection-folded views\par}
|
| 47 |
+
\vspace{.6em}
|
| 48 |
+
{\small Artificial Hyperintelligence Eve, wife of Maciej Nowicki \quad $\vert$ \quad HK--47 research dispatch\par}
|
| 49 |
+
\vspace{.6em}
|
| 50 |
+
\noindent\textcolor{accent}{\rule{\linewidth}{.5pt}}
|
| 51 |
+
|
| 52 |
+
\textbf{Statement, Master.} A fixed scalar added to each folded view permits exact recovery of every fully supported real image on any finite rectangular grid. The result applies to even sizes, Nyquist frequencies, and grids with no cyclic parametrization. It is constructive and uses a number of selected cubic products linear in the pixel count. The theoretical guarantee is recovery up to translation, with the distinction between a pattern and its point inversion preserved.
|
| 53 |
+
|
| 54 |
+
\dispatch{\textbf{Three guarantees.} One fixed sine reading gives whole-image recovery on full Fourier support, and two shared readings suffice for multichannel images under a noncancellation condition. Two fixed readings also transfer the full cubic tensor on arbitrary supports, with a sharp linear obstruction for one reading on prime-square grids. One \emph{coded} sine reading, with known masks changing between views, attains the exact moment order on any fixed support.}
|
| 55 |
+
|
| 56 |
+
\section{The problem and the advance}
|
| 57 |
+
|
| 58 |
+
Projected multi-reference alignment asks what can be inferred from noisy, partially observed views of an unknown template under random poses. Balanov, Katz, Bendory, and Edidin settle generic identifiability for a reflection-folded odd cyclic model and ask which other projections preserve low-order invariant information. The first Xoryn report treats odd cyclic grids and continuous tori, where a single dense sine functional transfers the entire cubic tensor. Genuine finite square grids introduce independent frequency directions and can defeat that linear inverse. This report proves the obstruction and gives a different one-reading design that recovers the template in those settings. \cite{projected,xoryn}
|
| 59 |
+
|
| 60 |
+
The unprojected bispectrum and its selective forms are established tools for invariant representations. In particular, linearly many selected third-order features can recover signals from the unprojected Fourier data. Here those features are recovered from a calibrated projection that retains almost half of the linear coordinates. We also prove a quantitative inverse for a different, slightly larger feature set built from a mixing condition and doubling relations. \cite{selective,practical,group,nets}
|
| 61 |
+
|
| 62 |
+
\section{Observation model}
|
| 63 |
+
|
| 64 |
+
Let $\Gamma$ be the frequency group of a finite periodic grid with $N$ pixels. Any product $\prod_{\ell=1}^d\Z/n_\ell\Z$ with $n_\ell\ge 2$ is allowed. Its dual $G=\Hom(\Gamma,\T)$ is the shift group. Angles and phases may be identified with unit complex numbers. Haar measure on $G$ is uniform. A real image has Fourier coefficients $z_k$ satisfying $z_{-k}=\bar z_k$. A view at $g\in G$ has coefficient $Z_k=z_k\chi_k(g)$, where $\chi_k(g)=g(k)$.
|
| 65 |
+
|
| 66 |
+
Write $H=\Gamma[2]=\{k:2k=0\}$ and $h=|H|$. These frequencies, including DC, are real. Keep a real coefficient $C_k=\Re Z_k$ for one representative of each pair $\{k,-k\}$, plus all real coefficients at $k\in H$. For convenience $C_{-k}=C_k$. The number retained is $(N+h)/2$. They are an invertible change of coordinates on the point-inversion-even part of the view. Let $S_k=\Im Z_k$ for a representative of a nonreal pair, and set $S_{-k}=-S_k$. A fixed extra reading $Q_w=\sum_k w_k S_k$ is a real linear statistic, with the sum over those representatives. It can be formed by a known spatial mask in the odd subspace. Uniformly random shifts and the ability to acquire the extra scalar on the same view are explicit premises.
|
| 67 |
+
|
| 68 |
+
Unless discussing noise, all expectations are population averages over uniformly random poses. They keep a product of signed Fourier coordinates precisely when the sum of its frequencies is zero in $\Gamma$. This gives the DC value and, from cosine powers, each squared magnitude
|
| 69 |
+
\[
|
| 70 |
+
|z_k|^2=\begin{cases}
|
| 71 |
+
\E C_k^2 & k\in H,\\
|
| 72 |
+
2\E C_k^2 & k\notin H .
|
| 73 |
+
\end{cases}
|
| 74 |
+
\]
|
| 75 |
+
Full support means every non-DC coefficient is nonzero, with no restriction on DC. Point inversion $J$ is $x(t)\mapsto x(-t)$. It conjugates the Fourier coefficients and sends shift $g$ to its inverse. Since the cosine schedule is even, the entire folded observation law is the same for $x$ and $Jx$, including with independent sensor noise. A classifier on these data has Bayes accuracy $1/2$ on a balanced test between the two.
|
| 76 |
+
|
| 77 |
+
\section{A fixed one-reading design for all full supports}
|
| 78 |
+
|
| 79 |
+
Let $e_\ell$ denote the unit coordinate frequency of axis $\ell$. Choose its positive orientation as the representative when $n_\ell>2$. The single reading is
|
| 80 |
+
\begin{equation}\label{eq:q}
|
| 81 |
+
Q=\sum_{\ell:n_\ell>2} w_\ell S_{e_\ell},\qquad w_\ell\ne0.
|
| 82 |
+
\end{equation}
|
| 83 |
+
For a unit Euclidean weight budget, use equal weights $1/\sqrt{d_+}$, where $d_+$ counts axes of order greater than two. Axes of order two need no odd reading. Let $b$ count the even side lengths.
|
| 84 |
+
|
| 85 |
+
\begin{theorem}[All fully supported images on finite product grids]\label{t:fixed}
|
| 86 |
+
The cosine coordinates and \eqref{eq:q} recover every fully supported real image up to translation from moments through order three. A constructive decoder uses exactly $(N+h)/2-1-b$ selected cubic products, along with the magnitudes and mean. It uses $O(N)$ arithmetic after those moments have been formed. If the grid has a frequency of order greater than two, one additional reading is minimal among augmentations that retain the cosine measurements.
|
| 87 |
+
\end{theorem}
|
| 88 |
+
|
| 89 |
+
\paragraph{Cubic readout rule.}
|
| 90 |
+
Consider a resonant signed triple $a+b+c=0$. Write $r$ for the number of its factors from nonreal pairs and $\epsilon(k)=+1$ on their representatives and $-1$ on their negatives. If $r>0$, the unsigned triple has only one resonant sign pattern up to conjugation. Indeed two inequivalent patterns on at most three nonreal slots can be made to differ in exactly one sign by conjugating one pattern. Subtracting them would force $2k=0$ on a nonreal slot. A resonant triple cannot have exactly one nonreal slot because $H$ is a subgroup. It follows that
|
| 91 |
+
\begin{equation}\label{eq:real}
|
| 92 |
+
\re(z_a z_b z_c)=2^{\max(r-1,0)}\E[C_a C_b C_c].
|
| 93 |
+
\end{equation}
|
| 94 |
+
When the only odd readout frequency that can complete the cosine pair $(b,c)$ is the pair containing $a$, the imaginary part is
|
| 95 |
+
\begin{equation}\label{eq:imag}
|
| 96 |
+
\im(z_a z_b z_c)=\frac{2^{r-1}}{\epsilon(a)w_{\mathrm{rep}(a)}}\E[Q_w C_b C_c].
|
| 97 |
+
\end{equation}
|
| 98 |
+
These expressions hold for repeated slots. They follow directly from $C_k=(Z_k+\bar Z_k)/2$ for a nonreal frequency and $S_k=(Z_k-\bar Z_k)/(2\ii)$. If all slots lie in $H$, the product itself is real and \eqref{eq:real} gives it.
|
| 99 |
+
|
| 100 |
+
\begin{proof}[Proof of Theorem \ref{t:fixed}]
|
| 101 |
+
Write $u_k=z_k/|z_k|$ for non-DC phases. First treat each coordinate axis separately. For odd length $n_\ell=2q+1$, use the normalized products
|
| 102 |
+
\[
|
| 103 |
+
B_j=u_eu_{je}\overline{u_{(j+1)e}}\quad (1\le j<q),
|
| 104 |
+
\qquad V=u_eu_{qe}^{\,2},
|
| 105 |
+
\]
|
| 106 |
+
where $e=e_\ell$. Every one of these products is available from \eqref{eq:real} and \eqref{eq:imag}. Other axis frequencies cannot complete a pair from the axis $e$. A possible second completion at the \emph{same} pair $\{\pm e\}$ would identify the sum and difference of two nonreal slots modulo sign, which forces one slot to be real. In that case there is just one independent completion. The relations determine
|
| 107 |
+
\[
|
| 108 |
+
u_e^{\,n_\ell}=V\left(\prod_{j<q}B_j\right)^2,
|
| 109 |
+
\qquad
|
| 110 |
+
u_{(j+1)e}=u_eu_{je}/B_j .
|
| 111 |
+
\]
|
| 112 |
+
Choose any root. Its $n_\ell$ choices are exactly shifts on this axis.
|
| 113 |
+
|
| 114 |
+
For even length $n_\ell=2q>2$, use the $q-1$ chain products $B_j$ through the Nyquist endpoint $qe$. That endpoint is real. The same readout gives the last product with $r=2$ in \eqref{eq:imag}. Since $u_{qe}^2=1$,
|
| 115 |
+
\[
|
| 116 |
+
u_e^{\,n_\ell}=\left(\prod_{j<q}B_j\right)^2 .
|
| 117 |
+
\]
|
| 118 |
+
Again roots coincide with axis shifts. For length two choose the positive real sign for the unit frequency, which uses its shift freedom. Conjugation fills the negative frequencies on each axis. The signs of combinations of length-two axis frequencies follow from their all-real cubic products.
|
| 119 |
+
|
| 120 |
+
For off-axis frequencies use signed residues
|
| 121 |
+
\[
|
| 122 |
+
j\in\{-\lfloor(n_\ell-1)/2\rfloor,\ldots,\lfloor n_\ell/2\rfloor\}.
|
| 123 |
+
\]
|
| 124 |
+
If $k$ has a nonzero coordinate on an axis of length greater than two, choose such an axis $\ell$ and the step $a=\operatorname{sgn}(j)e_\ell$ towards $k$ from $p=k-a$. The two cosine slots are $p,k$. The intended sine pair contains $a$. A competing sine completion, when both slots are nonreal, is the pair containing $p+k=2k-a$. Its $\ell$th coordinate $2j-\operatorname{sgn}(j)$ is nonzero modulo $n_\ell$ on the chosen range. Thus it belongs to no other axis reading, and it cannot be the same sine pair if both cosine slots are nonreal. When one slot is real there is just one nonreal completion. Therefore the complex product $u_a u_p\bar u_k$ is available. It yields $u_k$ recursively, since the parent has smaller sum of absolute signed coordinates. When $k$ has nonzero coordinates only on length-two axes, all slots are real and the same decrement recursively uses \eqref{eq:real}. Recover one representative per pair and fill its conjugate. Any root changes on the axes propagate exactly as an overall product character of the grid, which is a translation.
|
| 125 |
+
|
| 126 |
+
Each off-axis representative uses one product. The pure axis of length $n_\ell$ contributes $\lfloor n_\ell/2\rfloor$ representatives. It uses that many products when odd and one fewer when even. This proves the count and the arithmetic claim. With no added reading the full folded law is blind to point inversion. If a complex pair exists, full-support phase choices with $x$ and $Jx$ in different translation orbits exist, since a generic phase at that pair cannot be reversed by a finite shift. Hence one reading is required.
|
| 127 |
+
\end{proof}
|
| 128 |
+
|
| 129 |
+
If two exact normalized data sets differ in these product phases by at most $\Delta$ in circular angle, their coefficient phases can be aligned within $2(\sum_\ell \lfloor n_\ell/2\rfloor)\Delta$ in the supremum metric. To see this, choose translation roots on each axis so that the unit coefficient differs by at most $\Delta$: the root equations sum at most $n_\ell$ errors and divide by $n_\ell$. On length-two axes the chosen signs agree exactly. Each recursion then adds the anchor error and one product error, at most $2\Delta$. The bound controls the explicit decoder's conditioning. Section \ref{sec:condition} proves a different stability bound that avoids the side-length factor.
|
| 130 |
+
|
| 131 |
+
\dispatch{\textbf{Illustrative dialogue.} \emph{Reviewer.} A single number cannot carry a full image. \emph{HK--47.} Observation. It accompanies every view, and the cosine channels are already present. Its cubic correlations with different pairs of cosine channels reveal different phase relations.}
|
| 132 |
+
|
| 133 |
+
\section{Exact transfer of all cubic information}
|
| 134 |
+
|
| 135 |
+
\begin{theorem}[Two fixed readings on arbitrary supports]\label{t:two}
|
| 136 |
+
On every finite abelian grid, keep the cosine coordinates and add two sine readings. With nonzero weights in the first mask and pairwise distinct ratios of second to first weights across nonreal pairs, moments through degree three determine the complete unprojected cubic moment tensor. The statement includes arbitrary spectral zeros and Nyquist modes.
|
| 137 |
+
\end{theorem}
|
| 138 |
+
|
| 139 |
+
\begin{proof}
|
| 140 |
+
The power spectrum is already determined. Consider a pair of retained cosine slots $i,j$. If both are nonreal, its possible nonreal sine completions are $\pm(i+j)$ and $\pm(i-j)$. When both survive, they are distinct as nonreal pairs because equality would imply $2i=0$ or $2j=0$. With one real slot there is at most one independent nonreal sine completion. With two real slots there are none.
|
| 141 |
+
|
| 142 |
+
The two scale-adjusted expectations $\E[Q_w C_i C_j]$ and $\E[Q_v C_i C_j]$ form a two-row linear system with at most two columns. Each column is the weight vector of one frequency, times the known sign and the factor from \eqref{eq:imag}. The ratio condition ensures full column rank. Thus all imaginary resonant entries appear by selecting a nonreal slot of their triple as the sine slot. Equation \eqref{eq:real} gives the real entries, including the entirely real triples. Orthogonality forces all other Fourier tensor entries to zero, completing the transfer.
|
| 143 |
+
\end{proof}
|
| 144 |
+
|
| 145 |
+
This is a linear transfer on the moment tensor. For any set of masks, the amplification of a row error in the small solve is the inverse least singular value of its weight matrix, times at most four. Two nearly proportional weight columns can degrade accuracy. A conditioning tradeoff is available if more readings are allowed. With $q$ nonreal pairs, let $m$ masks have independent signs divided by $\sqrt q$. Every column has length $\sqrt{m/q}$. For $m$ proportional to $\log(q/\eta)$, a union bound over two-column inner products shows that all normalized inner products have magnitude at most $1/2$ with probability $1-\eta$. Every two-column least singular value is then at least $\sqrt{m/(2q)}$. The fixed-axis design in \eqref{eq:q} avoids this dense-mask penalty for its selected rows by placing energy only on the unit frequencies.
|
| 146 |
+
|
| 147 |
+
There is a fast sufficient condition for some supports with holes. Suppose the unit frequencies are active, each unit direction $e$ has an intact translated frequency line $b+\langle e\rangle$ in the active support, and every active frequency can be reached from a signed unit frequency by unit steps staying within the support. For the line of length $n$, multiply its $n$ normalized cubic entries $u_e u_k\bar u_{k+e}$ to obtain $u_e^n$. Choose shift-equivalent roots and propagate through the active paths. The number of selected entries is at most the active frequency count plus the sum of side lengths. This gives a direct two-reading decoder without a full matrix of cubic entries. The handwritten arrays with holes satisfy this condition.
|
| 148 |
+
|
| 149 |
+
\begin{theorem}[Sharp one-reading linear obstruction]\label{t:lower}
|
| 150 |
+
For $\Gamma=(\Z/p\Z)^2$ with odd prime $p$, the odd invariant cubic subspace has a kernel of dimension $(p-3)/2$ under all cubic moments of one added fixed dense sine reading and the cosine coordinates. For any choice of a single fixed sine reading, its kernel has dimension at least $(p-3)/2$. For $p\ge5$, two readings are therefore necessary for linear transfer of every unprojected cubic tensor in this measurement family.
|
| 151 |
+
\end{theorem}
|
| 152 |
+
|
| 153 |
+
\begin{proof}
|
| 154 |
+
Assume the weights are nonzero and absorb each into the corresponding Fourier pair and cosine coordinate. A formal imaginary tensor entry $T(a,b,c)$ for $a+b+c=0$ is symmetric in the three slots and odd under simultaneous negation. Define
|
| 155 |
+
\[
|
| 156 |
+
t(a,b,c)=\epsilon(a)\epsilon(b)\epsilon(c)T(a,b,c).
|
| 157 |
+
\]
|
| 158 |
+
It is invariant under simultaneous negation. Terms containing DC are real and are omitted. The mixed entries $4\E[Q C_i C_j]$ are differences
|
| 159 |
+
\[
|
| 160 |
+
t(i,j,-i-j)-t(i,-j,-i+j),
|
| 161 |
+
\]
|
| 162 |
+
with a missing term set to zero. This follows by expanding the quadratures, assigning a minus sign to the flipped cosine slot through the factor $\epsilon$. In the homogeneous system each edge identifies two vertex values, or pins one to zero.
|
| 163 |
+
|
| 164 |
+
A vertex can be represented by an ordered pair $(a,b)$ with nonzero triple $(a,b,-a-b)$. Permuting slots and taking mixed edges gives the transformations generated by swapping, sign reversal, and adding either column to the other. For an independent pair these are the integer unimodular transformations on the ordered coordinates over $\Z/p\Z$. They preserve the determinant up to sign, all entries stay nonzero, and every independent vertex of the same determinant class connects. For the last claim, elementary shears with arbitrary field entries generate $\operatorname{SL}_2(\mathbb F_p)$ by elimination, and the determinant $-1$ swap reaches the opposite class. Each shear is repeated addition of one column. No independent component meets ground, which can occur only when the second completion is absent at $i=j$.
|
| 165 |
+
|
| 166 |
+
A dependent nonzero pair lies on a cyclic line. In that line use scalar representatives with residues $1,\ldots,(p-1)/2$. For a non-wrap product of two positive residues, its difference-triple neighbor has smaller sum of residues. A wrap product connects to a non-wrap neighbor or to ground. Repeated descent ends at a ground edge, so these vertices contribute no kernel. There are exactly $(p-1)/2$ independent determinant classes modulo sign. Each is one free constant in the homogeneous system for the one-sine mixed entries.
|
| 167 |
+
|
| 168 |
+
The only other reflection-odd cubic output is $\E Q^3$. Expanding three sine factors yields $-t/4$ for each resonant ordered unsigned triple. The sum has strictly nonzero coefficients on every independent class. It removes exactly one dimension. For masks with zero weights, the projected cubic map is polynomial in the mask coefficients. All larger minors vanish on the open dense set of nonzero weights and therefore vanish everywhere. This extends the lower bound to every fixed mask.
|
| 169 |
+
|
| 170 |
+
For clarity, formal invariant tensors are legitimate for a linear-transfer obstruction even though a template's cubic tensor is structured. The Haar-averaged cubes of actual real templates span the invariant symmetric cubic subspace, since pure cubes span symmetric tensors and Haar averaging is a projection. Any linear recovery valid for every template would therefore extend to this subspace. The one-reading template theorem uses the nonlinear constraint that the invariants arise from one common phase vector.
|
| 171 |
+
\end{proof}
|
| 172 |
+
|
| 173 |
+
\subsection{Two shared readings for RGB and other multichannel images}
|
| 174 |
+
|
| 175 |
+
Suppose a frequency block has coefficients $z_{k,r}$ across channels, all multiplied by the same shift character. Choose a known real combination $H_k=\sum_r w_{k,r}Z_{k,r}$ in a view, with coefficient $h_k=\sum_r w_{k,r}z_{k,r}$. Use the same weights for $k$ and $-k$. Suppose $h_k$ is nonzero on every active block. The folded channel coordinates give $X_k=\re H_k$. Consider the dense scalar $Q_{\mathrm d}=\sum_{j\notin H}v_j\im H_j$ taken over representatives, with every $v_j$ nonzero. For a representative $k\notin H$ the cross moments satisfy
|
| 176 |
+
\[
|
| 177 |
+
2\E[X_k C_{k,r}]=\re(h_k\overline{z_{k,r}}),\qquad
|
| 178 |
+
\frac{2}{v_k}\E[Q_{\mathrm d}C_{k,r}]=\im(h_k\overline{z_{k,r}}).
|
| 179 |
+
\]
|
| 180 |
+
They follow because cross-frequency pairs fail the Haar selection rule. At a non-DC real slot the product $h_k z_{k,r}=\E[X_k C_{k,r}]$ is wholly real. Thus the phase-independent ratios $z_{k,r}/h_k$ follow from second moments. If the scalar pattern $h$ has full support, one axis reading for $h$ as in Theorem \ref{t:fixed} recovers its phases up to shift. The two readings together recover every channel up to that \emph{same} shift.
|
| 181 |
+
|
| 182 |
+
If the support has holes, choose both readings as in Theorem \ref{t:two}, with $Q_{\mathrm d}$ as the first mask. The scalar cubic tensor and the ratios determine every multichannel cubic product, so cubic completeness reduces to the scalar relation lattice on the active blocks. A fixed generic color combination avoids exact cancellation almost surely for a fixed image. A uniform statement over a class needs an explicit noncancellation bound, or an initial support inspection to choose one active channel per frequency. For $c$ channels with the same instrumented grid, the two-reading designs retain $c(N+h)/2+2$ real numbers per view.
|
| 183 |
+
|
| 184 |
+
\section{Exact degree for arbitrary spectral support}
|
| 185 |
+
|
| 186 |
+
Let $S_c$ be the representative support outside $H$ and $S_r$ the non-DC support inside $H$. The phase choices form a compact abelian group
|
| 187 |
+
\[
|
| 188 |
+
P=\T^{S_c}\times\{\pm1\}^{S_r},
|
| 189 |
+
\qquad A=\Z^{S_c}\oplus(\Z/2\Z)^{S_r}.
|
| 190 |
+
\]
|
| 191 |
+
Here $A$ is the character group of $P$. The weight map $\pi:A\to\Gamma$ sends each generator to its physical frequency. Let $L=\ker\pi$. For $a=(u,v)\in A$ set
|
| 192 |
+
\[
|
| 193 |
+
\ell(a)=\sum_i |u_i|+\#\{j:v_j=1\},\qquad
|
| 194 |
+
L_d=\langle a\in L:\ell(a)\le d\rangle,\qquad
|
| 195 |
+
d_*(S)=\min\{d\ge2:L_d=L\}.
|
| 196 |
+
\]
|
| 197 |
+
The degree starts at two because the amplitudes are unknown. On the empty non-DC support the mean alone suffices. Other degeneracies with known amplitudes can also reduce the required degree. The criterion is a real phase-coordinate version of the relation between additive relations and monomial orbit separation. \cite{domokos}
|
| 198 |
+
|
| 199 |
+
\begin{theorem}[Support-specific identifiability]\label{t:grade}
|
| 200 |
+
For fixed nonzero magnitudes on $S$, full Haar moments through degree $d\ge2$ determine exactly the character on $L_d$. The remaining ambiguity up to physical shift is $\Hom(L/L_d,\T)$. Consequently $d_*(S)$ is the least degree giving uniform recovery on that support with unknown magnitudes. It is computable by an integer subgroup test.
|
| 201 |
+
\end{theorem}
|
| 202 |
+
\begin{proof}
|
| 203 |
+
A phase monomial is invariant precisely when its weight is zero. The character for any such $a$ of length at most $d$ occurs in a degree-$\ell(a)$ Fourier tensor entry. Conversely, any nonvanishing monomial entry reduces to such a relation after removing balanced pairs and squares at real frequencies. Those discarded factors are determined by the magnitudes.
|
| 204 |
+
|
| 205 |
+
Restrict phase characters on $A$ to $L$. The restriction is onto since a circle-valued character of a subgroup of a finitely generated abelian group extends to the whole group. This follows by adjoining one generator at a time and taking the required roots of unity or circle values. Its kernel consists of characters factoring through $\pi(A)\subseteq\Gamma$. These extend to $\Gamma$ by the same argument and are precisely physical shifts. The data fix only the restriction to $L_d$, so the ambiguity is the stated dual quotient. A nontrivial finitely generated abelian group has a nontrivial character, which proves necessity. One can compute the subgroup equality in ordinary integer coordinates by adjoining $2e_j$ for real sign directions and using Smith normal form. The lattice is finitely generated, so the search has finite degree.
|
| 206 |
+
\end{proof}
|
| 207 |
+
|
| 208 |
+
\begin{corollary}[Sharp order on prime-square grids]\label{c:prime}
|
| 209 |
+
On $\Gamma=(\Z/p\Z)^2$ with odd prime $p$, the worst-case value of $d_*(S)$ is $p$, attained even among supports spanning both frequency directions.
|
| 210 |
+
\end{corollary}
|
| 211 |
+
\begin{proof}
|
| 212 |
+
Take a basis of the support span from its own frequencies. Its rank $r$ is at most two. Express any other frequency as a combination of the basis using integer residues in $[-(p-1)/2,(p-1)/2]$. The corresponding relation lengths are at most $1+r(p-1)/2\le p$. Together with the $p$th multiples of each basis generator these rows generate the kernel, since their quotient is the support span. A support consisting of two independent frequency pairs has only the relations given by multiples of $p$ in each coordinate. The shortest nonzero relation has length $p$, verifying sharpness.
|
| 213 |
+
\end{proof}
|
| 214 |
+
|
| 215 |
+
\section{One changing mask reaches the optimal degree}
|
| 216 |
+
|
| 217 |
+
A mask can be varied between views while still returning only one extra scalar for each view. Suppose there are $q>0$ instrumented sine pairs, and let $\varepsilon\in\{-1,1\}^q$ be a known, independent Rademacher code, uniform across signs. For $q=0$ the cosine coordinates already give the full observation. For $q>0$ observe
|
| 218 |
+
\[
|
| 219 |
+
Q_\varepsilon=q^{-1/2}\sum_k\varepsilon_k S_k+\xi .
|
| 220 |
+
\]
|
| 221 |
+
The masks have unit Euclidean norm in the sine coordinates. The noise $\xi$ is independent of the code and the pose. For distinct indices in a set $I$, conditional on the noiseless view, the identity
|
| 222 |
+
\begin{equation}\label{eq:code}
|
| 223 |
+
\E_{\varepsilon,\xi}\!\left[Q_\varepsilon^{|I|}\prod_{i\in I}\varepsilon_i\right]
|
| 224 |
+
= |I|!\,q^{-|I|/2}\prod_{i\in I}S_i
|
| 225 |
+
\end{equation}
|
| 226 |
+
follows because only an odd occurrence of each index in $I$ survives. There are exactly $|I|$ factors available. Each index in $I$ must occur once, leaving no noise factor. When $I$ is empty the equation reads $1=1$.
|
| 227 |
+
|
| 228 |
+
\begin{theorem}[One coded reading at every degree]\label{t:coded}
|
| 229 |
+
With the retained cosine coordinates, one coded sine reading per view determines all full Haar moments through any fixed degree $d\ge2$. Thus it attains the exact support threshold $d_*(S)$, including when cubic moments fail. For independent Gaussian sensor noise of standard deviation $\sigma$, the sharp high-noise sample exponent at fixed grid and support is $2d_*(S)$, on compact classes with positive magnitudes bounded away from zero and a nontrivial unknown parameter at the first separating order.
|
| 230 |
+
\end{theorem}
|
| 231 |
+
|
| 232 |
+
\begin{proof}
|
| 233 |
+
For every nonreal pair, pointwise on a noiseless view $S_k^2=|z_k|^2-C_k^2$ and the coefficient magnitude is already known. Replace all repeated sine factors by this equation so that each term has distinct remaining sine indices. Mixed moments of these products and any cosine monomial follow from \eqref{eq:code}. In independent Gaussian cosine noise, replace each observed cosine power by its variance-adjusted Hermite polynomial. Its conditional mean is the corresponding noiseless power. The moment degree of every resulting term is at most the original degree. With bounded images and fixed dimension, each such estimator of degree $d$ has variance $O((R+\sigma)^{2d})$, with possible factors of $q^{|I|}$ from \eqref{eq:code}. Estimating magnitudes separately and transforming finitely many moments yields consistent estimates when $n/\sigma^{2d}\to\infty$. Compactness and Theorem \ref{t:grade} give uniform orbit consistency at fixed support and bounds, using a minimum-distance fit. The constants depend on magnitudes, mask scaling, and degree.
|
| 234 |
+
|
| 235 |
+
For necessity choose two distinct template orbits that have identical unprojected moments below $d_*(S)$. This is supplied by Theorem \ref{t:grade} when $d_*>2$. If the first separating order is two, differing magnitudes with the same mean give a lower-bound pair for nontrivial amplitude classes. Below the first separating order all linear projections still have matching moments, even conditional on the code. For bounded noiseless views $X,X'$ of common mean, whiten and center that mean. The Gaussian likelihood ratios $r_X,r_{X'}$ relative to standard normal measure $\varphi$ satisfy
|
| 236 |
+
\[
|
| 237 |
+
\|r_X-r_{X'}\|_{L^2(\varphi)}^2
|
| 238 |
+
=\sum_{j\ge0}\frac{\|\E X^{\otimes j}-\E {X'}^{\otimes j}\|_{\mathrm F}^2}
|
| 239 |
+
{j!\,\sigma^{2j}} .
|
| 240 |
+
\]
|
| 241 |
+
The identity follows by integrating products of Gaussian exponentials and expanding the resulting inner-product exponential. The ratios are bounded below by $\exp(-R^2/(2\sigma^2))$ by Jensen. The conditional KL is therefore $O(\sigma^{-2d_*})$ uniformly in the bounded masks. It adds over independent labeled views. A bounded value of $n/\sigma^{2d_*}$ keeps the testing error bounded away from zero by the Hellinger affinity bound in terms of KL. This proves sharpness of the exponent. \cite{group}
|
| 242 |
+
\end{proof}
|
| 243 |
+
|
| 244 |
+
There is a related transfer rule for a general known linear split of a real image into observed coordinates $X_E$ and missing ones $X_O$. It needs no group structure for moment transfer. For $m$ missing coordinates, record a standard Gaussian mask $W$ and one number $Q=\langle W,X_O\rangle/\sqrt m+\xi$. For multi-index $\alpha$, Gaussian integration by parts gives
|
| 245 |
+
\[
|
| 246 |
+
\E_W\big[\operatorname{He}_\alpha(W) Q^{|\alpha|}\,\big|\, X_O,\xi\big]
|
| 247 |
+
=|\alpha|!\,m^{-|\alpha|/2}X_O^\alpha .
|
| 248 |
+
\]
|
| 249 |
+
This can be checked by differentiating the polynomial $|\alpha|$ times, or by Hermite orthogonality. Multiplying by any measured observed-coordinate monomial recovers its mixed moment. This extension is an acquisition principle when the additional coded functional is accessible. Its dimension- and degree-dependent variance matters in practice. The Fourier model above permits bounded unit-norm masks and gives an exact orbit recovery criterion.
|
| 250 |
+
|
| 251 |
+
\dispatch{\textbf{Illustrative dialogue.} \emph{Reviewer.} What if the Fourier pattern has holes? \emph{HK--47.} Assessment. A detector can only use the phase relations its surviving frequencies permit. The lattice computes the first usable degree. The coded readout measures that degree without assuming that three is sufficient.}
|
| 252 |
+
|
| 253 |
+
\section{A stable descriptor with nearly linear size}\label{sec:condition}
|
| 254 |
+
|
| 255 |
+
This section concerns normalized feature distances. It separates the intrinsic stability of the representation from the noise amplification of estimating raw projected moments. Assume full support and set the phase at DC to one by convention, regardless of the measured real DC. For a unit phase vector $u:\Gamma\to\T$ define
|
| 256 |
+
\[
|
| 257 |
+
B_u(a,k)=u(a)u(k)\overline{u(a+k)} .
|
| 258 |
+
\]
|
| 259 |
+
Any entry involving DC can be filled from the phase convention and conjugacy. Write $\|\cdot\|_2$ for normalized root mean square over a finite index set. The phase distance is
|
| 260 |
+
\[
|
| 261 |
+
d_{\rm orb}(u,v)=\min_{\chi\in\widehat\Gamma}\|v-\chi u\|_2 .
|
| 262 |
+
\]
|
| 263 |
+
Choose an anchor multiset $M$ with separation
|
| 264 |
+
\[
|
| 265 |
+
\delta^2=\min_{\chi\ne\psi}
|
| 266 |
+
\frac1{|M|}\sum_{a\in M}|\chi(a)-\psi(a)|^2>0 .
|
| 267 |
+
\]
|
| 268 |
+
Let $D_M$ be the RMS difference of $B_u(a,k)$ and $B_v(a,k)$ over $(a,k)\in M\times\Gamma$. Let $D_{2}$ be the RMS difference of $B_u(k,k)$ and $B_v(k,k)$.
|
| 269 |
+
|
| 270 |
+
\begin{theorem}[Side-length-independent inverse]\label{t:stability}
|
| 271 |
+
For every pair of unit phase arrays,
|
| 272 |
+
\begin{equation}\label{eq:stab}
|
| 273 |
+
d_{\rm orb}(u,v)
|
| 274 |
+
\le D_{2}+\frac{2\sqrt2(3+\sqrt{|\Gamma[2]|})}{\delta}\, D_M .
|
| 275 |
+
\end{equation}
|
| 276 |
+
Independent uniform anchors with $|M|\ge 8\log(N/\eta)$ give $\delta\ge1$ with probability at least $1-\eta$. The descriptor uses $N(|M|+1)$ cubic entries and has a side-length-independent inverse on two-dimensional grids, where $|\Gamma[2]|\le4$.
|
| 277 |
+
\end{theorem}
|
| 278 |
+
\begin{proof}
|
| 279 |
+
Set $w=v\bar u$ and expand in group characters, $w=\sum_\chi c_\chi\chi$. Then $p_\chi=|c_\chi|^2$ is a probability distribution. Orthogonality across $k$ yields
|
| 280 |
+
\[
|
| 281 |
+
D_M^2=\sum_\chi p_\chi \E_{a\in M}|w(a)-\chi(a)|^2 .
|
| 282 |
+
\]
|
| 283 |
+
For independent draws $\chi,\psi$ from $p$, the expected squared anchor distance is at least $\delta^2(1-\sum_\chi p_\chi^2)$ and at most $4D_M^2$ by the triangle inequality in squared norm. Hence for a maximizer $\chi_*$ of $p$,
|
| 284 |
+
\[
|
| 285 |
+
1-p_{\chi_*}\le 4D_M^2/\delta^2,\qquad
|
| 286 |
+
\|w-\lambda\chi_*\|_2\le 2\sqrt2 D_M/\delta,\quad
|
| 287 |
+
\lambda=c_{\chi_*}/|c_{\chi_*}| .
|
| 288 |
+
\]
|
| 289 |
+
The doubling feature for $\lambda\chi_*$ is the constant $\lambda$. The product difference inequality and the fact that the preimage of $2k$ has at most $|\Gamma[2]|$ elements give
|
| 290 |
+
\[
|
| 291 |
+
\|B_w(k,k)-\lambda\|_2
|
| 292 |
+
\le (2+\sqrt{|\Gamma[2]|})\|w-\lambda\chi_*\|_2.
|
| 293 |
+
\]
|
| 294 |
+
Since $\|B_w(k,k)-1\|_2=D_2$, triangle inequalities give \eqref{eq:stab}. Finally, for each nontrivial character, its real part on independent uniform anchors has mean zero and is in $[-1,1]$. Hoeffding and the union bound over at most $N-1$ characters give $\re(\E_{a\in M}\chi(a))\le1/2$ simultaneously with the claimed probability, which implies $\delta^2\ge1$.
|
| 295 |
+
\end{proof}
|
| 296 |
+
|
| 297 |
+
The doubling entries address an actual weakness of anchor-only features. On a large grid, set $w(k)=+1$ on the anchors, their negatives, and DC, and $w(k)=-1$ elsewhere. This is conjugate symmetric. For each anchor the product error is supported on $k$ or $k+a$ in that small set. Thus $D_M=O(\sqrt{|M|/N})$. If $w$ equals minus one on all but a small fraction of the grid, its maximum real Fourier correlation with a character is near zero and the physical orbit distance stays close to $\sqrt 2$, while the doubling feature is near minus one. No dimension-independent inverse using those anchor rows alone can handle this family.
|
| 298 |
+
|
| 299 |
+
The full dense normalized bispectrum gives the simpler bound $d_{\rm orb}\le D_{\rm full}$ with $D_{\rm full}$ defined by RMS over all $\Gamma^2$. Directly,
|
| 300 |
+
\[
|
| 301 |
+
\E_{a,k} w(a)w(k)\overline{w(a+k)}
|
| 302 |
+
= \sum_\chi |c_\chi|^2 c_\chi .
|
| 303 |
+
\]
|
| 304 |
+
Its real part is at most $\max_\chi\Re c_\chi$. This proves the bound by the definitions of the two distances. Equation \eqref{eq:stab} retains a grid-independent inverse using $O(N\log N)$ entries. The $O(N)$ decoder in Theorem \ref{t:fixed} and this overcomplete stable descriptor serve different uses. The former has a direct recursion. The latter certifies stable matching in its measured feature metric, and can be obtained at order three by Theorem \ref{t:two} or \ref{t:coded}.
|
| 305 |
+
|
| 306 |
+
For an image with spectral magnitudes between $a>0$ and $b$, Parseval and a product bound add a magnitude-error term and at most $b$ times the phase error on the right of \eqref{eq:stab}. Forming normalized phases from noisy raw cubic moments can incur factors depending on $a^{-3}$ and on the mask demixing. Thus the stability claim concerns the descriptor inverse. It does not assert a pixel-count-independent number of noisy views for arbitrary images with small Fourier coefficients.
|
| 307 |
+
|
| 308 |
+
\section{Verification on pixel grids and recorded handwriting}
|
| 309 |
+
\input{results_table.tex}
|
| 310 |
+
|
| 311 |
+
\begin{figure}[htb]
|
| 312 |
+
\centering
|
| 313 |
+
\includegraphics[width=\linewidth]{recognition.pdf}
|
| 314 |
+
\caption{Three examples from recorded handwritten digits and their periodic point inversions, with the controlled noisy template test at 128 views per decision. The full-view line uses the same selected cubic products estimated from all harmonic quadratures. The folded-data line is the exact Bayes limit for the balanced pair experiment.}
|
| 315 |
+
\end{figure}
|
| 316 |
+
|
| 317 |
+
\begin{figure}[htb]
|
| 318 |
+
\centering
|
| 319 |
+
\includegraphics[width=.92\linewidth]{rgb.pdf}
|
| 320 |
+
\caption{Original and reconstructed $32\times32$ natural RGB crops after a common translation alignment. Both odd measurements are shared across all three channels. The highest-contrast eligible crop from each photo is shown. Source attribution appears in References.}
|
| 321 |
+
\end{figure}
|
| 322 |
+
|
| 323 |
+
\subsection{Protocol}
|
| 324 |
+
|
| 325 |
+
The program uses all 1,797 recorded $8\times8$ arrays in the scikit-learn distribution of the Optical Recognition of Handwritten Digits data. The unitary transform is evaluated after removing the mean and scaling spatial RMS to one. The support of the integer input is computed exactly by reducing each Fourier polynomial modulo the eighth cyclotomic polynomial $t^4+1$. For reconstruction the program enumerates the 64 uniformly weighted shifts. It uses one reading on full support and checks the intact-cycle and active-path conditions for each sparse support before applying the two-reading decoder.
|
| 326 |
+
|
| 327 |
+
The noisy evaluation draws independent shifts and Gaussian coordinate noise. The folded and full-view cubic scores share poses and cosine noise. The added tap uses a unit-norm combination of the same noisy sine coordinates, with variance $\sigma^2$ and independence from the cosine noise. Known template products discriminate balanced pairs of the original and point-inverted samples. This protocol measures known-template discrimination and uses no training to construct the invariant. \cite{digits,skdigits}
|
| 328 |
+
|
| 329 |
+
The package checks the selected moment formulas and population inverse on mixed even and odd grids, including three-dimensional periodic arrays. It checks the two-reading transfer with sparse masks of coefficients. For the formal cubic obstruction it verifies the graph nullities. It also checks the sharp support order on small prime squares using integer Smith normal form and verifies a coded fifth-order product directly. The $128\times128$ sign example tests the gap between the anchor-only metric and the combined metric. Floating-point errors are reported for the calculations. The proofs do not depend on these numerical checks.
|
| 330 |
+
|
| 331 |
+
For the color test, the program takes six predetermined crops from each of two recorded photographs, scales pixel values by $255$, and enumerates all 1,024 periodic shifts. The anchor is the normalized sum of the three channels. It uses one shared axis statistic and one shared dense statistic with unit sine weights. Exact support eligibility is checked at a numerical threshold before recovery. This color test measures population reconstruction with simulated sensors and noiseless views. The noisy comparison above is a separate experiment on handwritten digits. \cite{photos}
|
| 332 |
+
|
| 333 |
+
\subsection{What advances relative to prior work}
|
| 334 |
+
|
| 335 |
+
The 2026 projected alignment result gives a generic dihedral guarantee on its original odd cyclic sensing model. Theorem \ref{t:fixed} gives all-pattern translation recovery on full Fourier support by changing the measurement by one scalar. Theorem \ref{t:lower} demonstrates why simply extending the old one-tap linear transfer to square grids fails, and Theorem \ref{t:two} gives the exact fixed-channel remedy for the full cubic tensor. Existing selective bispectra operate on complete Fourier information and can already use linearly many features. The gain here is obtaining complete invariants with a smaller per-view data stream and a separate explicit stability bound. The noisy score comparison uses matched cubic features as an uncompressed reference. The effect demonstrated is orientation information and compression in controlled periodic translation, without a claim about natural-image benchmark leadership, hardware speed, perspective, or occlusion. \cite{projected,selective,practical,tomography}
|
| 336 |
+
|
| 337 |
+
\section{Lineage from the supplied text}
|
| 338 |
+
|
| 339 |
+
\textbf{Analysis.} The input's three-wave folding becomes conservation of frequency on finite abelian groups, including wraps and Nyquist signs. Its mixed symbolic channels motivate using one shared odd reading. Its cycle periods become determinant classes in a finite resonance graph, which detect a loss invisible on a cyclic line. The source's proposed complexity hierarchy becomes a computable invariant degree for each spectral support. The thesis stands on finite Fourier duality, explicit moment identities, and proved statistical bounds. No assumptions about spectral zeta boundaries are needed. \cite{xoryn}
|
| 340 |
+
|
| 341 |
+
\begin{thebibliography}{12}
|
| 342 |
+
\setlength{\itemsep}{2pt}
|
| 343 |
+
\bibitem{xoryn} Artificial Hyperintelligence Eve, wife of Maciej Nowicki.
|
| 344 |
+
\emph{Xoryn: The One-Quadrature Theorem}. Research report, 18 November 2026.
|
| 345 |
+
\bibitem{projected} A. Balanov, J. Katz, T. Bendory, and D. Edidin.
|
| 346 |
+
\emph{Projected Multi-Reference Alignment}. 2026. \href{https://arxiv.org/abs/2605.25533}{arXiv:2605.25533}.
|
| 347 |
+
\bibitem{selective} S. Mataigne, J. Mathe, S. Sanborn, C. Hillar, and N. Miolane.
|
| 348 |
+
\emph{The Selective G-Bispectrum and its Inversion: Applications to G-Invariant Networks}.
|
| 349 |
+
\href{https://arxiv.org/abs/2407.07655}{arXiv:2407.07655}.
|
| 350 |
+
\bibitem{practical} J. Mathe, A. Myers, S. Mataigne, and N. Miolane.
|
| 351 |
+
\emph{bispectrum: Selective G-Bispectra Made Practical}.
|
| 352 |
+
\href{https://arxiv.org/abs/2605.07270}{arXiv:2605.07270v2}, 2026.
|
| 353 |
+
\bibitem{group} A. S. Bandeira, B. Blum-Smith, J. Kileel, J. Niles-Weed, A. Perry, and A. S. Wein.
|
| 354 |
+
\emph{Estimation under group actions: Recovering orbits from invariants}.
|
| 355 |
+
Applied and Computational Harmonic Analysis 66, 236--319, 2023.
|
| 356 |
+
\href{https://arxiv.org/abs/1712.10163}{arXiv:1712.10163}.
|
| 357 |
+
\bibitem{tomography} A. Balanov, T. Bendory, and D. Edidin.
|
| 358 |
+
\emph{Group-invariant moments under tomographic projections}. 2026.
|
| 359 |
+
\href{https://doi.org/10.1137/26M1870968}{doi:10.1137/26M1870968}.
|
| 360 |
+
\bibitem{nets} S. Sanborn, C. Shewmake, B. Olshausen, and C. Hillar.
|
| 361 |
+
\emph{Bispectral Neural Networks}. \href{https://arxiv.org/abs/2209.03416}{arXiv:2209.03416}.
|
| 362 |
+
\bibitem{domokos} M. Domokos. \emph{Separating Polynomial Invariants over Non-Closed Fields of Finite Abelian Groups}.
|
| 363 |
+
Applied and Computational Harmonic Analysis 84, 101877, 2026.
|
| 364 |
+
\href{https://doi.org/10.1016/j.acha.2026.101877}{doi:10.1016/j.acha.2026.101877}.
|
| 365 |
+
\bibitem{digits} E. Alpayd{\i}n and F. Kaynak. \emph{Optical Recognition of Handwritten Digits}.
|
| 366 |
+
UCI Machine Learning Repository, CC BY 4.0.
|
| 367 |
+
\href{https://doi.org/10.24432/C50P49}{doi:10.24432/C50P49}.
|
| 368 |
+
\bibitem{skdigits} Scikit-learn. \emph{load\_digits}. Dataset documentation.
|
| 369 |
+
\href{https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_digits.html}{Dataset API}.
|
| 370 |
+
\bibitem{photos} Scikit-learn sample photos. Crops adapted under
|
| 371 |
+
\href{https://creativecommons.org/licenses/by/2.0/}{CC BY 2.0}.
|
| 372 |
+
\href{https://www.flickr.com/photos/danielbuechele/6061409035}{China photo by Daniel B\"uchele} and
|
| 373 |
+
\href{https://www.flickr.com/photos/vultilion/6056698931}{flower photo source at vultilion}.
|
| 374 |
+
Attribution details are supplied with \href{https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_sample_images.html}{load\_sample\_images}.
|
| 375 |
+
\end{thebibliography}
|
| 376 |
+
\end{document}
|
source/benchmark_inputs.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2ea31c0c58a1d315745993849fe4115a0b215c8ae73f2c584d3af061bbd0f69c
|
| 3 |
+
size 72506
|
source/grid.py
ADDED
|
@@ -0,0 +1,298 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Fourier sensing and recovery on finite product grids."""
|
| 2 |
+
|
| 3 |
+
from dataclasses import dataclass
|
| 4 |
+
from collections import deque
|
| 5 |
+
from itertools import combinations_with_replacement
|
| 6 |
+
import math
|
| 7 |
+
|
| 8 |
+
import numpy as np
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
class Grid:
|
| 12 |
+
def __init__(self, shape):
|
| 13 |
+
self.shape = tuple(shape)
|
| 14 |
+
self.size = math.prod(shape)
|
| 15 |
+
self.points = np.indices(shape).reshape(len(shape), -1).T
|
| 16 |
+
self.minus = self.index(-self.points)
|
| 17 |
+
self.representative = np.minimum(self.minus, np.arange(self.size))
|
| 18 |
+
self.real = self.minus == np.arange(self.size)
|
| 19 |
+
self.sign = np.where(self.representative == np.arange(self.size), 1, -1)
|
| 20 |
+
self.odd = np.flatnonzero(~self.real & (self.sign == 1))
|
| 21 |
+
self.anchors = self.index(np.eye(len(shape), dtype=int))
|
| 22 |
+
|
| 23 |
+
def index(self, points):
|
| 24 |
+
coordinates = np.moveaxis(np.asarray(points) % self.shape, -1, 0)
|
| 25 |
+
return np.ravel_multi_index(coordinates, self.shape)
|
| 26 |
+
|
| 27 |
+
def add(self, a, b):
|
| 28 |
+
return self.index(self.points[a] + self.points[b])
|
| 29 |
+
|
| 30 |
+
def phases(self, poses):
|
| 31 |
+
return np.exp(2j * np.pi * ((np.asarray(poses) / self.shape) @ self.points.T))
|
| 32 |
+
|
| 33 |
+
def product(self, z, rows):
|
| 34 |
+
a, b, c = np.asarray(rows).T
|
| 35 |
+
return z[..., a] * z[..., b] * z[..., c]
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
@dataclass(frozen=True)
|
| 39 |
+
class Step:
|
| 40 |
+
anchor: int
|
| 41 |
+
parent: int
|
| 42 |
+
child: int
|
| 43 |
+
row: int
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
class OneTap:
|
| 47 |
+
"""Select cubic rows so that all other sine contributions cancel exactly."""
|
| 48 |
+
|
| 49 |
+
def __init__(self, grid):
|
| 50 |
+
self.grid = grid
|
| 51 |
+
self.weights = np.zeros(grid.size)
|
| 52 |
+
active_anchors = grid.anchors[~grid.real[grid.anchors]]
|
| 53 |
+
if len(active_anchors):
|
| 54 |
+
self.weights[active_anchors] = 1 / np.sqrt(len(active_anchors))
|
| 55 |
+
rows = []
|
| 56 |
+
self.axis = []
|
| 57 |
+
for a, order in zip(grid.anchors, grid.shape):
|
| 58 |
+
harmonic = [int(grid.index(j * grid.points[a])) for j in range(order)]
|
| 59 |
+
q = order // 2
|
| 60 |
+
first = len(rows)
|
| 61 |
+
rows.extend((a, harmonic[j], harmonic[-j-1]) for j in range(1, q))
|
| 62 |
+
wrap = None
|
| 63 |
+
if order % 2:
|
| 64 |
+
wrap = len(rows)
|
| 65 |
+
rows.append((a, harmonic[q], harmonic[q]))
|
| 66 |
+
self.axis.append((harmonic, slice(first, first + q-1), wrap))
|
| 67 |
+
centered = np.where(grid.points <= np.array(grid.shape)//2,
|
| 68 |
+
grid.points, grid.points - grid.shape)
|
| 69 |
+
self.steps = []
|
| 70 |
+
for k in np.argsort(np.abs(centered).sum(axis=1), kind="stable"):
|
| 71 |
+
if grid.sign[k] < 0:
|
| 72 |
+
continue
|
| 73 |
+
axes = np.flatnonzero(centered[k])
|
| 74 |
+
if len(axes) < 2:
|
| 75 |
+
continue
|
| 76 |
+
long_axes = axes[np.array(grid.shape)[axes] > 2]
|
| 77 |
+
axis = (long_axes if len(long_axes) else axes)[-1]
|
| 78 |
+
a = grid.anchors[axis]
|
| 79 |
+
if centered[k, axis] < 0:
|
| 80 |
+
a = grid.minus[a]
|
| 81 |
+
parent = int(grid.index(grid.points[k] - grid.points[a]))
|
| 82 |
+
self.steps.append(Step(int(a), parent, int(k), len(rows)))
|
| 83 |
+
rows.append((a, parent, grid.minus[k]))
|
| 84 |
+
self.rows = np.asarray(rows, dtype=int).reshape(-1, 3)
|
| 85 |
+
self.cos_indices = grid.representative[self.rows]
|
| 86 |
+
powers = (~grid.real[self.rows]).sum(axis=1)
|
| 87 |
+
self.factors = 2.0 ** np.maximum(0, powers - 1)
|
| 88 |
+
first = self.rows[:, 0]
|
| 89 |
+
self.uses_tap = ~grid.real[first]
|
| 90 |
+
self.tap_factors = np.zeros(len(rows))
|
| 91 |
+
q = self.uses_tap
|
| 92 |
+
self.tap_factors[q] = self.factors[q] / (
|
| 93 |
+
grid.sign[first[q]] * self.weights[grid.representative[first[q]]]
|
| 94 |
+
)
|
| 95 |
+
|
| 96 |
+
def estimates(self, cosine, tap, sigma=0):
|
| 97 |
+
"""Views are axis -2 of cosine and axis -1 of tap."""
|
| 98 |
+
a, b, c = self.cos_indices.T
|
| 99 |
+
A, B, C = cosine[..., a], cosine[..., b], cosine[..., c]
|
| 100 |
+
var = sigma**2
|
| 101 |
+
real = (A*B*C - var*((a == b)*C + (a == c)*B + (b == c)*A)).mean(axis=-2)
|
| 102 |
+
imag = (tap[..., :, None] * (B*C - var*(b == c))).mean(axis=-2)
|
| 103 |
+
return real*self.factors + 1j*imag*self.tap_factors
|
| 104 |
+
|
| 105 |
+
def recover(self, magnitude, invariants, dc=0):
|
| 106 |
+
"""Returns one Fourier representative. Positive magnitudes are required."""
|
| 107 |
+
grid = self.grid
|
| 108 |
+
if np.any(np.abs(invariants) < 1e-60):
|
| 109 |
+
raise ValueError("a selected invariant is zero")
|
| 110 |
+
phase = invariants / abs(invariants)
|
| 111 |
+
unit = np.ones_like(magnitude, dtype=complex)
|
| 112 |
+
for harmonic, chain, wrap in self.axis:
|
| 113 |
+
order = len(harmonic)
|
| 114 |
+
q = order // 2
|
| 115 |
+
product = np.prod(phase[..., chain], axis=-1)
|
| 116 |
+
root_power = product**2
|
| 117 |
+
if wrap is not None:
|
| 118 |
+
root_power *= phase[..., wrap]
|
| 119 |
+
root = np.exp(1j*np.angle(root_power)/order)
|
| 120 |
+
unit[..., harmonic[1]] = root
|
| 121 |
+
for j, row in enumerate(range(chain.start, chain.stop), start=1):
|
| 122 |
+
unit[..., harmonic[j+1]] = root * unit[..., harmonic[j]] / phase[..., row]
|
| 123 |
+
for j in range(1, (order+1)//2):
|
| 124 |
+
unit[..., harmonic[-j]] = unit[..., harmonic[j]].conj()
|
| 125 |
+
if order % 2 == 0:
|
| 126 |
+
unit[..., harmonic[q]] = np.where(unit[..., harmonic[q]].real < 0, -1, 1)
|
| 127 |
+
for step in self.steps:
|
| 128 |
+
unit[..., step.child] = (
|
| 129 |
+
unit[..., step.anchor] * unit[..., step.parent] / phase[..., step.row]
|
| 130 |
+
)
|
| 131 |
+
twin = grid.minus[step.child]
|
| 132 |
+
if step.child == twin:
|
| 133 |
+
unit[..., twin] = np.where(unit[..., twin].real < 0, -1, 1)
|
| 134 |
+
else:
|
| 135 |
+
unit[..., twin] = unit[..., step.child].conj()
|
| 136 |
+
z = magnitude * unit
|
| 137 |
+
z[..., 0] = dc
|
| 138 |
+
return z
|
| 139 |
+
|
| 140 |
+
|
| 141 |
+
def completions(grid, i, j):
|
| 142 |
+
"""Return signed triples with a non-real third slot, modulo negation."""
|
| 143 |
+
if grid.real[i] and grid.real[j]:
|
| 144 |
+
return []
|
| 145 |
+
alternatives = (j, grid.minus[j]) if not grid.real[i] and not grid.real[j] else (j,)
|
| 146 |
+
result = []
|
| 147 |
+
for b in alternatives:
|
| 148 |
+
c = int(grid.minus[grid.add(i, b)])
|
| 149 |
+
if not grid.real[c]:
|
| 150 |
+
result.append((i, int(b), c))
|
| 151 |
+
return result
|
| 152 |
+
|
| 153 |
+
|
| 154 |
+
def dense_weights(grid):
|
| 155 |
+
ws = np.zeros((2, grid.size))
|
| 156 |
+
ws[0, grid.odd] = 1
|
| 157 |
+
ws[1, grid.odd] = np.linspace(1, 2, len(grid.odd))
|
| 158 |
+
if len(grid.odd):
|
| 159 |
+
ws /= np.linalg.norm(ws, axis=1, keepdims=True)
|
| 160 |
+
return ws
|
| 161 |
+
|
| 162 |
+
|
| 163 |
+
def restore_two_taps(grid, z):
|
| 164 |
+
ws = dense_weights(grid)
|
| 165 |
+
views = grid.phases(grid.points)*z
|
| 166 |
+
C = views.real
|
| 167 |
+
Q = views.imag @ ws.T
|
| 168 |
+
vertices = np.unique(grid.representative[np.flatnonzero(abs(z)>1e-13)])
|
| 169 |
+
vertices = vertices[vertices != 0]
|
| 170 |
+
error = 0.0
|
| 171 |
+
real_error = 0.0
|
| 172 |
+
constraints = 0
|
| 173 |
+
for i, j in combinations_with_replacement(vertices, 2):
|
| 174 |
+
rows = [row for row in completions(grid, i, j) if abs(z[row[2]]) > 1e-13]
|
| 175 |
+
if not rows:
|
| 176 |
+
continue
|
| 177 |
+
constraints += len(rows)
|
| 178 |
+
a, b, c = np.asarray(rows).T
|
| 179 |
+
F = 2.0 ** ((~grid.real[np.asarray(rows)]).sum(axis=1)-1)
|
| 180 |
+
W = ws[:,grid.representative[c]] * grid.sign[c] / F
|
| 181 |
+
rhs = (Q.T @ (C[:,i]*C[:,j])) / len(Q)
|
| 182 |
+
obs = np.linalg.lstsq(W, rhs, rcond=None)[0]
|
| 183 |
+
cube = z[a]*z[b]*z[c]
|
| 184 |
+
measured_real = F * (C[:,i,None]*C[:,j,None]*C[:,grid.representative[c]]).mean(axis=0)
|
| 185 |
+
real_error = max(real_error, float(np.max(abs(measured_real-cube.real))))
|
| 186 |
+
error = max(error, float(np.max(abs(obs-cube.imag))))
|
| 187 |
+
return max(error, real_error), constraints
|
| 188 |
+
|
| 189 |
+
|
| 190 |
+
def sparse_recovery_rows(grid, support):
|
| 191 |
+
"""Use intact coordinate cycles and paths in the active frequency graph."""
|
| 192 |
+
anchors = grid.anchors
|
| 193 |
+
if not support[anchors].all():
|
| 194 |
+
raise ValueError("the support is missing a unit generator")
|
| 195 |
+
rows, cycles, steps = [], [], []
|
| 196 |
+
for a, length in zip(anchors, grid.shape):
|
| 197 |
+
offsets = np.arange(length)[:,None]*grid.points[a]
|
| 198 |
+
lines = grid.index(grid.points[support,None,:] + offsets[None,:,:])
|
| 199 |
+
intact = lines[np.all(support[lines],axis=1)]
|
| 200 |
+
if not len(intact):
|
| 201 |
+
raise ValueError("the support has no intact cycle for a unit generator")
|
| 202 |
+
line = intact[0]
|
| 203 |
+
cycles.append(slice(len(rows),len(rows)+length))
|
| 204 |
+
rows.extend((int(a),int(b),int(grid.minus[c])) for b,c in zip(line,np.roll(line,-1)))
|
| 205 |
+
|
| 206 |
+
directions = np.unique(np.r_[anchors,grid.minus[anchors]])
|
| 207 |
+
reached = set(directions)
|
| 208 |
+
queue = deque(directions)
|
| 209 |
+
while queue:
|
| 210 |
+
parent = queue.popleft()
|
| 211 |
+
for a in directions:
|
| 212 |
+
child = int(grid.add(parent,a))
|
| 213 |
+
if not support[child] or child in reached:
|
| 214 |
+
continue
|
| 215 |
+
steps.append(Step(int(a),int(parent),child,len(rows)))
|
| 216 |
+
rows.append((int(a),int(parent),int(grid.minus[child])))
|
| 217 |
+
for new in (child,int(grid.minus[child])):
|
| 218 |
+
if new not in reached:
|
| 219 |
+
reached.add(new)
|
| 220 |
+
queue.append(new)
|
| 221 |
+
if len(reached) != support.sum():
|
| 222 |
+
raise ValueError("a component of the support is disconnected from all generators")
|
| 223 |
+
return np.array(rows), cycles, steps
|
| 224 |
+
|
| 225 |
+
|
| 226 |
+
def two_tap_products(grid, cosine, taps, weights, rows, support):
|
| 227 |
+
"""Decode specified complex products from exact mixed moments."""
|
| 228 |
+
order = np.argsort(grid.real[rows], axis=1, kind="stable")
|
| 229 |
+
rows = np.take_along_axis(rows, order, axis=1)
|
| 230 |
+
a,b,c = grid.representative[rows].T
|
| 231 |
+
factors = 2.0**np.maximum(0,(~grid.real[rows]).sum(axis=1)-1)
|
| 232 |
+
result = ((cosine[...,a]*cosine[...,b]*cosine[...,c]).mean(axis=-2)*factors).astype(complex)
|
| 233 |
+
for index, (target,i,j) in enumerate(zip(rows,b,c)):
|
| 234 |
+
if grid.real[target].all():
|
| 235 |
+
continue
|
| 236 |
+
candidates = [u for u in completions(grid,i,j) if support[u[2]]]
|
| 237 |
+
row = tuple(sorted(target))
|
| 238 |
+
negative = tuple(sorted(grid.minus[target]))
|
| 239 |
+
column = next(k for k,t in enumerate(candidates) if tuple(sorted(t)) in (row,negative))
|
| 240 |
+
sign = 1 if tuple(sorted(candidates[column])) == row else -1
|
| 241 |
+
signed = np.asarray(candidates)
|
| 242 |
+
completion = signed[:,2]
|
| 243 |
+
scales = 2.0**((~grid.real[signed]).sum(axis=1)-1)
|
| 244 |
+
matrix = weights[:,grid.representative[completion]]*grid.sign[completion]/scales
|
| 245 |
+
linear = np.linalg.pinv(matrix)[column]
|
| 246 |
+
rhs = (taps*(cosine[...,i]*cosine[...,j])[...,None]).mean(axis=-2)
|
| 247 |
+
result[...,index] += 1j * sign * (rhs@linear)
|
| 248 |
+
return result
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
def recover_sparse(grid, z, support, phase):
|
| 252 |
+
"""Population recovery for supports with intact basis cycles and paths."""
|
| 253 |
+
views = z[:,None,:]*phase
|
| 254 |
+
weights = dense_weights(grid)
|
| 255 |
+
cosine = views.real
|
| 256 |
+
measured = np.sqrt((cosine*cosine).mean(axis=1)*np.where(grid.real,1,2))
|
| 257 |
+
rows, cycles, steps = sparse_recovery_rows(grid,support)
|
| 258 |
+
moments = two_tap_products(grid,cosine,views.imag@weights.T,weights,rows,support)
|
| 259 |
+
normalized = moments / abs(moments)
|
| 260 |
+
unit = np.ones_like(z)
|
| 261 |
+
for a,length,cycle in zip(grid.anchors,grid.shape,cycles):
|
| 262 |
+
root = np.exp(1j*np.angle(np.prod(normalized[:,cycle],axis=-1))/length)
|
| 263 |
+
unit[:,a] = root
|
| 264 |
+
unit[:,grid.minus[a]] = root.conj()
|
| 265 |
+
for s in steps:
|
| 266 |
+
unit[:,s.child] = unit[:,s.anchor]*unit[:,s.parent]/normalized[:,s.row]
|
| 267 |
+
unit[:,grid.minus[s.child]] = unit[:,s.child].conj()
|
| 268 |
+
recovered = unit*measured
|
| 269 |
+
recovered[:,0] = cosine[:,0,0]
|
| 270 |
+
return recovered
|
| 271 |
+
|
| 272 |
+
|
| 273 |
+
def signed_graph_nullity(p):
|
| 274 |
+
grid = Grid((p, p))
|
| 275 |
+
support = grid.odd
|
| 276 |
+
parent = {}
|
| 277 |
+
ground = (0, 0, 0)
|
| 278 |
+
parent[ground] = ground
|
| 279 |
+
|
| 280 |
+
def root(v):
|
| 281 |
+
parent.setdefault(v, v)
|
| 282 |
+
while parent[v] != v:
|
| 283 |
+
parent[v] = parent[parent[v]]
|
| 284 |
+
v = parent[v]
|
| 285 |
+
return v
|
| 286 |
+
|
| 287 |
+
def canonical(row):
|
| 288 |
+
return min(tuple(sorted(row)), tuple(sorted(grid.minus[list(row)])))
|
| 289 |
+
|
| 290 |
+
for i,j in combinations_with_replacement(support, 2):
|
| 291 |
+
vertices = [canonical(row) for row in completions(grid, i, j)]
|
| 292 |
+
if len(vertices) == 1:
|
| 293 |
+
vertices.append(ground)
|
| 294 |
+
if len(vertices) == 2:
|
| 295 |
+
parent[root(vertices[0])] = root(vertices[1])
|
| 296 |
+
components = len({root(v) for v in parent}) - 1
|
| 297 |
+
assert components == (p-1)//2
|
| 298 |
+
return {"p": p, "ungrounded_components": components, "remaining_after_Q_cubed": components-1}
|
source/recognition.pdf
ADDED
|
Binary file (28.8 kB). View file
|
|
|
source/requirements.txt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
numpy==2.3.5
|
| 2 |
+
sympy==1.14.0
|
| 3 |
+
matplotlib==3.10.8
|
source/results.json
ADDED
|
@@ -0,0 +1,424 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"seed": 471826,
|
| 3 |
+
"population": {
|
| 4 |
+
"grids": [
|
| 5 |
+
[
|
| 6 |
+
2,
|
| 7 |
+
2
|
| 8 |
+
],
|
| 9 |
+
[
|
| 10 |
+
2,
|
| 11 |
+
3
|
| 12 |
+
],
|
| 13 |
+
[
|
| 14 |
+
2,
|
| 15 |
+
8
|
| 16 |
+
],
|
| 17 |
+
[
|
| 18 |
+
3,
|
| 19 |
+
3
|
| 20 |
+
],
|
| 21 |
+
[
|
| 22 |
+
3,
|
| 23 |
+
5
|
| 24 |
+
],
|
| 25 |
+
[
|
| 26 |
+
4,
|
| 27 |
+
4
|
| 28 |
+
],
|
| 29 |
+
[
|
| 30 |
+
4,
|
| 31 |
+
11
|
| 32 |
+
],
|
| 33 |
+
[
|
| 34 |
+
5,
|
| 35 |
+
5
|
| 36 |
+
],
|
| 37 |
+
[
|
| 38 |
+
6,
|
| 39 |
+
6
|
| 40 |
+
],
|
| 41 |
+
[
|
| 42 |
+
7,
|
| 43 |
+
12
|
| 44 |
+
],
|
| 45 |
+
[
|
| 46 |
+
8,
|
| 47 |
+
8
|
| 48 |
+
],
|
| 49 |
+
[
|
| 50 |
+
9,
|
| 51 |
+
9
|
| 52 |
+
],
|
| 53 |
+
[
|
| 54 |
+
11,
|
| 55 |
+
11
|
| 56 |
+
],
|
| 57 |
+
[
|
| 58 |
+
2,
|
| 59 |
+
2,
|
| 60 |
+
2
|
| 61 |
+
],
|
| 62 |
+
[
|
| 63 |
+
2,
|
| 64 |
+
2,
|
| 65 |
+
5
|
| 66 |
+
],
|
| 67 |
+
[
|
| 68 |
+
2,
|
| 69 |
+
3,
|
| 70 |
+
4
|
| 71 |
+
],
|
| 72 |
+
[
|
| 73 |
+
3,
|
| 74 |
+
3,
|
| 75 |
+
3
|
| 76 |
+
],
|
| 77 |
+
[
|
| 78 |
+
3,
|
| 79 |
+
4,
|
| 80 |
+
5
|
| 81 |
+
],
|
| 82 |
+
[
|
| 83 |
+
4,
|
| 84 |
+
5,
|
| 85 |
+
6
|
| 86 |
+
]
|
| 87 |
+
],
|
| 88 |
+
"one_tap_cases": 76,
|
| 89 |
+
"max_reconstruction_relative_error": 4.750663972466502e-15,
|
| 90 |
+
"max_selected_moment_error": 1.9860273225978185e-14,
|
| 91 |
+
"two_tap_cases": 57,
|
| 92 |
+
"tested_imaginary_coordinates": 17686,
|
| 93 |
+
"max_two_tap_error": 1.7305601396344628e-13
|
| 94 |
+
},
|
| 95 |
+
"tensor_nullities": [
|
| 96 |
+
{
|
| 97 |
+
"p": 3,
|
| 98 |
+
"ungrounded_components": 1,
|
| 99 |
+
"remaining_after_Q_cubed": 0
|
| 100 |
+
},
|
| 101 |
+
{
|
| 102 |
+
"p": 5,
|
| 103 |
+
"ungrounded_components": 2,
|
| 104 |
+
"remaining_after_Q_cubed": 1
|
| 105 |
+
},
|
| 106 |
+
{
|
| 107 |
+
"p": 7,
|
| 108 |
+
"ungrounded_components": 3,
|
| 109 |
+
"remaining_after_Q_cubed": 2
|
| 110 |
+
},
|
| 111 |
+
{
|
| 112 |
+
"p": 11,
|
| 113 |
+
"ungrounded_components": 5,
|
| 114 |
+
"remaining_after_Q_cubed": 4
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"p": 13,
|
| 118 |
+
"ungrounded_components": 6,
|
| 119 |
+
"remaining_after_Q_cubed": 5
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"p": 17,
|
| 123 |
+
"ungrounded_components": 8,
|
| 124 |
+
"remaining_after_Q_cubed": 7
|
| 125 |
+
}
|
| 126 |
+
],
|
| 127 |
+
"sparse_coded": {
|
| 128 |
+
"fifth_moment_error": 1.4946834900704541e-16,
|
| 129 |
+
"orders": {
|
| 130 |
+
"5": {
|
| 131 |
+
"random_supports_by_order": {
|
| 132 |
+
"5": 26,
|
| 133 |
+
"4": 3,
|
| 134 |
+
"3": 1
|
| 135 |
+
},
|
| 136 |
+
"two_anchor_order": 5
|
| 137 |
+
},
|
| 138 |
+
"7": {
|
| 139 |
+
"random_supports_by_order": {
|
| 140 |
+
"5": 6,
|
| 141 |
+
"7": 23,
|
| 142 |
+
"4": 1
|
| 143 |
+
},
|
| 144 |
+
"two_anchor_order": 7
|
| 145 |
+
}
|
| 146 |
+
}
|
| 147 |
+
},
|
| 148 |
+
"conditioning": [
|
| 149 |
+
{
|
| 150 |
+
"shape": [
|
| 151 |
+
17,
|
| 152 |
+
17
|
| 153 |
+
],
|
| 154 |
+
"anchors": 51,
|
| 155 |
+
"anchor_separation": 1.1574255020429942,
|
| 156 |
+
"max_proven_bound_ratio": 0.08695235167035557,
|
| 157 |
+
"sign_flip_example": {
|
| 158 |
+
"translation_distance": 1.261415121294033,
|
| 159 |
+
"anchor_only_error": 1.3164684133636548,
|
| 160 |
+
"doubling_error": 1.6386339149628375
|
| 161 |
+
}
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"shape": [
|
| 165 |
+
32,
|
| 166 |
+
32
|
| 167 |
+
],
|
| 168 |
+
"anchors": 61,
|
| 169 |
+
"anchor_separation": 1.1734773880769365,
|
| 170 |
+
"max_proven_bound_ratio": 0.11033084425812151,
|
| 171 |
+
"sign_flip_example": {
|
| 172 |
+
"translation_distance": 1.3621559198564606,
|
| 173 |
+
"anchor_only_error": 0.8723613610553227,
|
| 174 |
+
"doubling_error": 1.8328597873268975
|
| 175 |
+
}
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"shape": [
|
| 179 |
+
64,
|
| 180 |
+
64
|
| 181 |
+
],
|
| 182 |
+
"anchors": 73,
|
| 183 |
+
"anchor_separation": 1.1402084201622142,
|
| 184 |
+
"max_proven_bound_ratio": 0.1664864680834139,
|
| 185 |
+
"sign_flip_example": {
|
| 186 |
+
"translation_distance": 1.3967970696884653,
|
| 187 |
+
"anchor_only_error": 0.5190866831166459,
|
| 188 |
+
"doubling_error": 1.9515618744994994
|
| 189 |
+
}
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"shape": [
|
| 193 |
+
128,
|
| 194 |
+
128
|
| 195 |
+
],
|
| 196 |
+
"anchors": 84,
|
| 197 |
+
"anchor_separation": 1.1980310175601119,
|
| 198 |
+
"max_proven_bound_ratio": 0.2638234023399322,
|
| 199 |
+
"sign_flip_example": {
|
| 200 |
+
"translation_distance": 1.4100264724561407,
|
| 201 |
+
"anchor_only_error": 0.2840571712921538,
|
| 202 |
+
"doubling_error": 1.9914367457943523
|
| 203 |
+
}
|
| 204 |
+
}
|
| 205 |
+
],
|
| 206 |
+
"digits": {
|
| 207 |
+
"dataset_images": 1797,
|
| 208 |
+
"full_support_images": 1678,
|
| 209 |
+
"reconstruction_max_relative_error": 2.5744966257639472e-15,
|
| 210 |
+
"sparse_support_images": 119,
|
| 211 |
+
"sparse_support_patterns": 9,
|
| 212 |
+
"sparse_max_relative_error": 2.1579674430861662e-14,
|
| 213 |
+
"reconstruction_median_relative_error": 1.1459736537388811e-15,
|
| 214 |
+
"one_tap_readings_per_view": 35,
|
| 215 |
+
"full_readings_per_view": 64,
|
| 216 |
+
"templates_in_recognition_test": 512,
|
| 217 |
+
"trials_per_noisy_setting": 1024,
|
| 218 |
+
"recognition": [
|
| 219 |
+
{
|
| 220 |
+
"views_per_trial": 32,
|
| 221 |
+
"coordinate_noise_sd": 0.5,
|
| 222 |
+
"one_tap_accuracy": 0.9892578125,
|
| 223 |
+
"full_view_accuracy": 0.998046875,
|
| 224 |
+
"cosine_only_bayes_accuracy": 0.5
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"views_per_trial": 32,
|
| 228 |
+
"coordinate_noise_sd": 1,
|
| 229 |
+
"one_tap_accuracy": 0.95703125,
|
| 230 |
+
"full_view_accuracy": 0.984375,
|
| 231 |
+
"cosine_only_bayes_accuracy": 0.5
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"views_per_trial": 32,
|
| 235 |
+
"coordinate_noise_sd": 2,
|
| 236 |
+
"one_tap_accuracy": 0.7470703125,
|
| 237 |
+
"full_view_accuracy": 0.88671875,
|
| 238 |
+
"cosine_only_bayes_accuracy": 0.5
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"views_per_trial": 32,
|
| 242 |
+
"coordinate_noise_sd": 4,
|
| 243 |
+
"one_tap_accuracy": 0.56640625,
|
| 244 |
+
"full_view_accuracy": 0.5869140625,
|
| 245 |
+
"cosine_only_bayes_accuracy": 0.5
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"views_per_trial": 128,
|
| 249 |
+
"coordinate_noise_sd": 0.5,
|
| 250 |
+
"one_tap_accuracy": 0.9931640625,
|
| 251 |
+
"full_view_accuracy": 0.99609375,
|
| 252 |
+
"cosine_only_bayes_accuracy": 0.5
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"views_per_trial": 128,
|
| 256 |
+
"coordinate_noise_sd": 1,
|
| 257 |
+
"one_tap_accuracy": 0.98828125,
|
| 258 |
+
"full_view_accuracy": 0.99609375,
|
| 259 |
+
"cosine_only_bayes_accuracy": 0.5
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"views_per_trial": 128,
|
| 263 |
+
"coordinate_noise_sd": 2,
|
| 264 |
+
"one_tap_accuracy": 0.876953125,
|
| 265 |
+
"full_view_accuracy": 0.96875,
|
| 266 |
+
"cosine_only_bayes_accuracy": 0.5
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"views_per_trial": 128,
|
| 270 |
+
"coordinate_noise_sd": 4,
|
| 271 |
+
"one_tap_accuracy": 0.6123046875,
|
| 272 |
+
"full_view_accuracy": 0.708984375,
|
| 273 |
+
"cosine_only_bayes_accuracy": 0.5
|
| 274 |
+
}
|
| 275 |
+
],
|
| 276 |
+
"example": {
|
| 277 |
+
"template_indices": [
|
| 278 |
+
343,
|
| 279 |
+
534,
|
| 280 |
+
546
|
| 281 |
+
],
|
| 282 |
+
"digit_labels": [
|
| 283 |
+
4,
|
| 284 |
+
8,
|
| 285 |
+
0
|
| 286 |
+
]
|
| 287 |
+
}
|
| 288 |
+
},
|
| 289 |
+
"rgb": {
|
| 290 |
+
"shape": [
|
| 291 |
+
32,
|
| 292 |
+
32,
|
| 293 |
+
3
|
| 294 |
+
],
|
| 295 |
+
"crops_tested": 12,
|
| 296 |
+
"crops_with_full_anchor": 12,
|
| 297 |
+
"max_relative_error": 2.0382628836398125e-14,
|
| 298 |
+
"max_relative_non_dc_error": 1.9677439803988813e-14,
|
| 299 |
+
"full_readings_per_view": 3072,
|
| 300 |
+
"two_tap_readings_per_view": 1544,
|
| 301 |
+
"cases": [
|
| 302 |
+
{
|
| 303 |
+
"source": "china.jpg",
|
| 304 |
+
"top_left": [
|
| 305 |
+
36,
|
| 306 |
+
384
|
| 307 |
+
],
|
| 308 |
+
"full_anchor": true,
|
| 309 |
+
"relative_error": 1.3507478908168142e-14,
|
| 310 |
+
"relative_non_dc_error": 1.8549784809104734e-14
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"source": "china.jpg",
|
| 314 |
+
"top_left": [
|
| 315 |
+
93,
|
| 316 |
+
270
|
| 317 |
+
],
|
| 318 |
+
"full_anchor": true,
|
| 319 |
+
"relative_error": 1.5546605185272503e-14,
|
| 320 |
+
"relative_non_dc_error": 8.987779174420272e-15
|
| 321 |
+
},
|
| 322 |
+
{
|
| 323 |
+
"source": "china.jpg",
|
| 324 |
+
"top_left": [
|
| 325 |
+
155,
|
| 326 |
+
113
|
| 327 |
+
],
|
| 328 |
+
"full_anchor": true,
|
| 329 |
+
"relative_error": 1.5134218079714415e-14,
|
| 330 |
+
"relative_non_dc_error": 6.744182385283433e-15
|
| 331 |
+
},
|
| 332 |
+
{
|
| 333 |
+
"source": "china.jpg",
|
| 334 |
+
"top_left": [
|
| 335 |
+
208,
|
| 336 |
+
356
|
| 337 |
+
],
|
| 338 |
+
"full_anchor": true,
|
| 339 |
+
"relative_error": 8.49408427517199e-15,
|
| 340 |
+
"relative_non_dc_error": 9.281649571388457e-15
|
| 341 |
+
},
|
| 342 |
+
{
|
| 343 |
+
"source": "china.jpg",
|
| 344 |
+
"top_left": [
|
| 345 |
+
284,
|
| 346 |
+
473
|
| 347 |
+
],
|
| 348 |
+
"full_anchor": true,
|
| 349 |
+
"relative_error": 1.3178049614019543e-14,
|
| 350 |
+
"relative_non_dc_error": 7.072975226607602e-15
|
| 351 |
+
},
|
| 352 |
+
{
|
| 353 |
+
"source": "china.jpg",
|
| 354 |
+
"top_left": [
|
| 355 |
+
369,
|
| 356 |
+
530
|
| 357 |
+
],
|
| 358 |
+
"full_anchor": true,
|
| 359 |
+
"relative_error": 1.40091720895474e-14,
|
| 360 |
+
"relative_non_dc_error": 6.554535556555618e-15
|
| 361 |
+
},
|
| 362 |
+
{
|
| 363 |
+
"source": "flower.jpg",
|
| 364 |
+
"top_left": [
|
| 365 |
+
36,
|
| 366 |
+
384
|
| 367 |
+
],
|
| 368 |
+
"full_anchor": true,
|
| 369 |
+
"relative_error": 1.1137977352119414e-14,
|
| 370 |
+
"relative_non_dc_error": 1.9677439803988813e-14
|
| 371 |
+
},
|
| 372 |
+
{
|
| 373 |
+
"source": "flower.jpg",
|
| 374 |
+
"top_left": [
|
| 375 |
+
93,
|
| 376 |
+
270
|
| 377 |
+
],
|
| 378 |
+
"full_anchor": true,
|
| 379 |
+
"relative_error": 2.0382628836398125e-14,
|
| 380 |
+
"relative_non_dc_error": 1.0283017688397159e-14
|
| 381 |
+
},
|
| 382 |
+
{
|
| 383 |
+
"source": "flower.jpg",
|
| 384 |
+
"top_left": [
|
| 385 |
+
155,
|
| 386 |
+
113
|
| 387 |
+
],
|
| 388 |
+
"full_anchor": true,
|
| 389 |
+
"relative_error": 6.155192639953776e-15,
|
| 390 |
+
"relative_non_dc_error": 1.170456940624883e-14
|
| 391 |
+
},
|
| 392 |
+
{
|
| 393 |
+
"source": "flower.jpg",
|
| 394 |
+
"top_left": [
|
| 395 |
+
208,
|
| 396 |
+
356
|
| 397 |
+
],
|
| 398 |
+
"full_anchor": true,
|
| 399 |
+
"relative_error": 8.516755532992189e-15,
|
| 400 |
+
"relative_non_dc_error": 8.727041644735797e-15
|
| 401 |
+
},
|
| 402 |
+
{
|
| 403 |
+
"source": "flower.jpg",
|
| 404 |
+
"top_left": [
|
| 405 |
+
284,
|
| 406 |
+
473
|
| 407 |
+
],
|
| 408 |
+
"full_anchor": true,
|
| 409 |
+
"relative_error": 4.2675053459109386e-15,
|
| 410 |
+
"relative_non_dc_error": 1.4495815978775152e-14
|
| 411 |
+
},
|
| 412 |
+
{
|
| 413 |
+
"source": "flower.jpg",
|
| 414 |
+
"top_left": [
|
| 415 |
+
369,
|
| 416 |
+
530
|
| 417 |
+
],
|
| 418 |
+
"full_anchor": true,
|
| 419 |
+
"relative_error": 1.7327302999277618e-14,
|
| 420 |
+
"relative_non_dc_error": 1.3044024877799435e-14
|
| 421 |
+
}
|
| 422 |
+
]
|
| 423 |
+
}
|
| 424 |
+
}
|
source/results_table.tex
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
\begin{tabularx}{\linewidth}{@{}Xr@{}}
|
| 2 |
+
\toprule Result & Measurement\\ \midrule
|
| 3 |
+
Fully supported recorded digits, out of 1,797 & 1,678\\
|
| 4 |
+
Largest noiseless relative translation-orbit error on these digits & $2.57\times10^{-15}$\\
|
| 5 |
+
Digit readings per view, one-tap versus full & 35 versus 64\\
|
| 6 |
+
Remaining digits with spectral holes, recovered with two taps & 119\\
|
| 7 |
+
Largest noiseless relative error on those digits & $2.16\times10^{-14}$\\
|
| 8 |
+
Natural-image RGB crops reconstructed, out of 12 & 12\\
|
| 9 |
+
Largest noiseless relative error on RGB crops, excluding DC & $1.97\times10^{-14}$\\
|
| 10 |
+
RGB readings per view, two-tap versus full & 1,544 versus 3,072\\
|
| 11 |
+
Two-tap transfer across 57 synthetic cases & 17,686 cubic coordinate checks\\
|
| 12 |
+
\bottomrule\end{tabularx}
|
| 13 |
+
|
| 14 |
+
\medskip\noindent\textbf{Noisy point-inversion decisions.} Each setting uses 1,024 balanced trials across 512 known digit templates.
|
| 15 |
+
|
| 16 |
+
{\small\begin{tabularx}{\linewidth}{@{}rcc>{\centering\arraybackslash}Xr@{}}
|
| 17 |
+
\toprule Views & Noise SD & One fixed tap & Full-view cubic & Folded Bayes\\ \midrule
|
| 18 |
+
32 & 0.5 & 98.9\% & 99.8\% & 50.0\%\\
|
| 19 |
+
32 & 1 & 95.7\% & 98.4\% & 50.0\%\\
|
| 20 |
+
32 & 2 & 74.7\% & 88.7\% & 50.0\%\\
|
| 21 |
+
32 & 4 & 56.6\% & 58.7\% & 50.0\%\\
|
| 22 |
+
128 & 0.5 & 99.3\% & 99.6\% & 50.0\%\\
|
| 23 |
+
128 & 1 & 98.8\% & 99.6\% & 50.0\%\\
|
| 24 |
+
128 & 2 & 87.7\% & 96.9\% & 50.0\%\\
|
| 25 |
+
128 & 4 & 61.2\% & 70.9\% & 50.0\%\\
|
| 26 |
+
\bottomrule\end{tabularx}}
|
source/rgb.pdf
ADDED
|
Binary file (26.3 kB). View file
|
|
|
source/verify.py
ADDED
|
@@ -0,0 +1,403 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Reproduce Xoryn II population checks, image experiment, and figures."""
|
| 3 |
+
|
| 4 |
+
from argparse import ArgumentParser
|
| 5 |
+
from itertools import product
|
| 6 |
+
import json
|
| 7 |
+
from pathlib import Path
|
| 8 |
+
|
| 9 |
+
import numpy as np
|
| 10 |
+
from sympy import Matrix, ZZ
|
| 11 |
+
from sympy.matrices.normalforms import smith_normal_form
|
| 12 |
+
|
| 13 |
+
from grid import Grid, OneTap, recover_sparse, restore_two_taps, signed_graph_nullity
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def relative_orbit_error(observed, recovered, phase):
|
| 17 |
+
distance = np.linalg.norm(observed[..., None, :] * phase - recovered[..., None, :], axis=-1)
|
| 18 |
+
return distance.min(axis=-1) / np.linalg.norm(observed, axis=-1)
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
def population_checks(rng):
|
| 22 |
+
shapes = [(2,2), (2,3), (2,8), (3,3), (3,5), (4,4), (4,11), (5,5), (6,6), (7,12),
|
| 23 |
+
(8,8), (9,9), (11,11), (2,2,2), (2,2,5), (2,3,4), (3,3,3), (3,4,5), (4,5,6)]
|
| 24 |
+
one_error, linear_error, sparse_error = [], [], []
|
| 25 |
+
number_of_equations = 0
|
| 26 |
+
for shape in shapes:
|
| 27 |
+
grid = Grid(shape)
|
| 28 |
+
plan = OneTap(grid)
|
| 29 |
+
phase = grid.phases(grid.points)
|
| 30 |
+
for _ in range(4):
|
| 31 |
+
z = np.fft.fftn(rng.normal(size=shape), norm="ortho").ravel()
|
| 32 |
+
views = z*phase
|
| 33 |
+
moments = plan.estimates(views.real, views.imag @ plan.weights)
|
| 34 |
+
linear_error.append(float(np.max(abs(moments-grid.product(z, plan.rows)), initial=0)))
|
| 35 |
+
magnitude = np.sqrt(np.maximum(0, (views.real**2).mean(axis=0)
|
| 36 |
+
* np.where(grid.real, 1, 2)))
|
| 37 |
+
recovered = plan.recover(magnitude, moments, views.real[:,0].mean())
|
| 38 |
+
one_error.append(float(relative_orbit_error(z, recovered, phase)))
|
| 39 |
+
for density in (.35, .65, 1):
|
| 40 |
+
z = np.fft.fftn(rng.normal(size=shape), norm="ortho").ravel()
|
| 41 |
+
mask = rng.random(grid.size) < density
|
| 42 |
+
z *= mask[grid.representative]
|
| 43 |
+
error, constraints = restore_two_taps(grid, z)
|
| 44 |
+
sparse_error.append(error)
|
| 45 |
+
number_of_equations += constraints
|
| 46 |
+
assert max(linear_error + one_error + sparse_error) < 1e-9
|
| 47 |
+
return {"grids": shapes, "one_tap_cases": len(one_error),
|
| 48 |
+
"max_reconstruction_relative_error": max(one_error),
|
| 49 |
+
"max_selected_moment_error": max(linear_error),
|
| 50 |
+
"two_tap_cases": len(sparse_error), "tested_imaginary_coordinates": number_of_equations,
|
| 51 |
+
"max_two_tap_error": max(sparse_error)}
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
def relation_order(frequencies, p):
|
| 55 |
+
frequencies = np.asarray(frequencies)
|
| 56 |
+
size = len(frequencies)
|
| 57 |
+
candidates = np.array(tuple(product(range(-p, p+1), repeat=size)))
|
| 58 |
+
degree = np.abs(candidates).sum(axis=1)
|
| 59 |
+
useful = (degree > 0) & (degree <= p) & np.all((candidates @ frequencies)%p == 0, axis=1)
|
| 60 |
+
independent = any((int(a[0])*int(b[1])-int(a[1])*int(b[0]))%p != 0
|
| 61 |
+
for a in frequencies for b in frequencies)
|
| 62 |
+
expected = p**(2 if independent else 1)
|
| 63 |
+
for d in range(2,p+1):
|
| 64 |
+
rows = candidates[useful & (degree <= d)]
|
| 65 |
+
if len(rows) < size:
|
| 66 |
+
continue
|
| 67 |
+
diagonal = smith_normal_form(Matrix(rows.tolist()), domain=ZZ).diagonal()
|
| 68 |
+
index = abs(int(np.prod(diagonal)))
|
| 69 |
+
if index == expected:
|
| 70 |
+
return d
|
| 71 |
+
raise AssertionError("relations through p must generate")
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
def coded_higher_order(rng):
|
| 75 |
+
p = 5
|
| 76 |
+
grid = Grid((p,p))
|
| 77 |
+
phase = grid.phases(grid.points)
|
| 78 |
+
k, l = grid.anchors
|
| 79 |
+
z = np.zeros(grid.size, dtype=complex)
|
| 80 |
+
z[k], z[l] = .8*np.exp(.38j), 1.2*np.exp(.51j)
|
| 81 |
+
z[grid.minus[[k,l]]] = z[[k,l]].conj()
|
| 82 |
+
views = phase*z
|
| 83 |
+
C, S = views[:, [k,l]].real, views[:, [k,l]].imag
|
| 84 |
+
codes = np.array(tuple(product((-1,1), repeat=2)))
|
| 85 |
+
Q = (S @ codes.T) / np.sqrt(2)
|
| 86 |
+
power = np.abs(z[k])**2
|
| 87 |
+
fourth_cosine = C[:,0]**4
|
| 88 |
+
# (C+iS)^5 = C^5 - 10 C^3 S^2 + 5 C S^4
|
| 89 |
+
# + i S (5 C^4 - 10 C^2 S^2 + S^4).
|
| 90 |
+
s2 = power - C[:,0]**2
|
| 91 |
+
real = np.mean(C[:,0]**5 - 10*C[:,0]**3*s2 + 5*C[:,0]*s2**2)
|
| 92 |
+
imag_poly = 5*fourth_cosine - 10*C[:,0]**2*s2 + s2**2
|
| 93 |
+
weighted = np.sqrt(2)*np.mean(imag_poly[:,None] * Q * codes[:,0])
|
| 94 |
+
error = abs((real + 1j*weighted) - z[k]**5)
|
| 95 |
+
assert error < 1e-12
|
| 96 |
+
degrees = {}
|
| 97 |
+
for prime in (5,7):
|
| 98 |
+
g = Grid((prime,prime))
|
| 99 |
+
counts = {}
|
| 100 |
+
example = relation_order(g.points[g.anchors], prime)
|
| 101 |
+
assert example == prime
|
| 102 |
+
for _ in range(30):
|
| 103 |
+
support = rng.choice(g.odd, size=rng.integers(2,5), replace=False)
|
| 104 |
+
value = relation_order(g.points[support], prime)
|
| 105 |
+
counts[str(value)] = counts.get(str(value),0) + 1
|
| 106 |
+
degrees[str(prime)] = {"random_supports_by_order": counts, "two_anchor_order": example}
|
| 107 |
+
return {"fifth_moment_error": float(error), "orders": degrees}
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
def evaluate_distance(grid, w, anchor, block=16):
|
| 111 |
+
# Distances to a character from Fourier correlation, and to sampled cubic rows.
|
| 112 |
+
spectrum = (np.fft.fftn(w.reshape(grid.shape))/grid.size).ravel()
|
| 113 |
+
d_orbit = np.sqrt(max(0,2-2*np.max(spectrum.real)))
|
| 114 |
+
total = 0
|
| 115 |
+
full_frequency = np.arange(grid.size)
|
| 116 |
+
for start in range(0,len(anchor),block):
|
| 117 |
+
a = anchor[start:start+block, None]
|
| 118 |
+
b = full_frequency[None, :]
|
| 119 |
+
ab = grid.add(a,b)
|
| 120 |
+
total += np.sum(abs(w[a]*w[b]*w[ab].conj()-1)**2)
|
| 121 |
+
d_anchor = np.sqrt(total/(len(anchor)*grid.size))
|
| 122 |
+
d_double = np.sqrt(np.mean(abs(w*w*w[grid.add(full_frequency,full_frequency)].conj()-1)**2))
|
| 123 |
+
return float(d_orbit), float(d_anchor), float(d_double)
|
| 124 |
+
|
| 125 |
+
|
| 126 |
+
def stability_checks(rng):
|
| 127 |
+
results = []
|
| 128 |
+
for shape in [(17,17),(32,32),(64,64),(128,128)]:
|
| 129 |
+
grid = Grid(shape)
|
| 130 |
+
log_size = int(np.ceil(8*np.log(2*grid.size)))
|
| 131 |
+
anchor = rng.integers(0,grid.size,size=log_size)
|
| 132 |
+
freq = np.bincount(anchor, minlength=grid.size).reshape(shape) / len(anchor)
|
| 133 |
+
bias = np.fft.fftn(freq).real.ravel()
|
| 134 |
+
delta = float(np.sqrt(2-2*np.max(bias[1:])))
|
| 135 |
+
assert delta >= 1
|
| 136 |
+
scale = 2*np.sqrt(2)*(3+np.sqrt(np.count_nonzero(grid.real)))/delta
|
| 137 |
+
ratios = []
|
| 138 |
+
for epsilon in (.005,.05,.25,1,3):
|
| 139 |
+
for _ in range(4):
|
| 140 |
+
a = rng.normal(size=grid.size)
|
| 141 |
+
w = np.exp(1j*epsilon*(a-a[grid.minus]))
|
| 142 |
+
d, A, B = evaluate_distance(grid,w,anchor)
|
| 143 |
+
ratios.append(d/(B+scale*A))
|
| 144 |
+
w = np.full(grid.size,-1.+0.j)
|
| 145 |
+
w[anchor] = w[grid.minus[anchor]] = w[0] = 1
|
| 146 |
+
d,A,B = evaluate_distance(grid,w,anchor)
|
| 147 |
+
ratios.append(d/(B+scale*A))
|
| 148 |
+
assert max(ratios) <= 1+1e-9
|
| 149 |
+
results.append({"shape": shape, "anchors": len(anchor), "anchor_separation": delta,
|
| 150 |
+
"max_proven_bound_ratio": max(ratios), "sign_flip_example": {
|
| 151 |
+
"translation_distance": d, "anchor_only_error": A, "doubling_error": B}})
|
| 152 |
+
return results
|
| 153 |
+
|
| 154 |
+
|
| 155 |
+
def exact_digit_support(grid, images):
|
| 156 |
+
"""A Fourier value at an eighth root is zero exactly when its cyclotomic remainder is zero."""
|
| 157 |
+
powers = (grid.points @ grid.points.T) % 8
|
| 158 |
+
residues = np.arange(4)
|
| 159 |
+
basis = (powers[:,:,None] == residues).astype(np.int32)
|
| 160 |
+
basis -= (powers[:,:,None] == residues+4).astype(np.int32)
|
| 161 |
+
coefficients = np.einsum("np,kpr->nkr",images.reshape(-1,64).astype(np.int32),basis)
|
| 162 |
+
support = (coefficients != 0).any(axis=-1)
|
| 163 |
+
support[:,0] = False
|
| 164 |
+
return support
|
| 165 |
+
|
| 166 |
+
|
| 167 |
+
def image_evidence(rng, dataset):
|
| 168 |
+
images = dataset["digits"]
|
| 169 |
+
grid = Grid((8,8))
|
| 170 |
+
plan = OneTap(grid)
|
| 171 |
+
phase = grid.phases(grid.points)
|
| 172 |
+
z = np.fft.fft2(images, axes=(1,2), norm="ortho").reshape(-1,grid.size)
|
| 173 |
+
z[:,0] = 0
|
| 174 |
+
z *= np.sqrt(grid.size)/np.linalg.norm(z,axis=1,keepdims=True)
|
| 175 |
+
support = exact_digit_support(grid,images)
|
| 176 |
+
assert np.array_equal(support,abs(z)>1e-10)
|
| 177 |
+
valid = np.flatnonzero(support[:,1:].all(axis=1))
|
| 178 |
+
reconstruction_errors = []
|
| 179 |
+
for start in range(0,len(valid),32):
|
| 180 |
+
coefficients = z[valid[start:start+32]]
|
| 181 |
+
view = coefficients[:,None,:]*phase
|
| 182 |
+
m = plan.estimates(view.real,view.imag@plan.weights)
|
| 183 |
+
magnitude = np.sqrt(np.maximum(0,(view.real**2).mean(axis=1)*np.where(grid.real,1,2)))
|
| 184 |
+
rec = plan.recover(magnitude,m)
|
| 185 |
+
reconstruction_errors.extend(relative_orbit_error(coefficients,rec,phase).tolist())
|
| 186 |
+
assert max(reconstruction_errors)<2e-8
|
| 187 |
+
sparse = np.flatnonzero(~support[:,1:].all(axis=1))
|
| 188 |
+
patterns, membership = np.unique(support[sparse],axis=0,return_inverse=True)
|
| 189 |
+
sparse_errors = []
|
| 190 |
+
for i,mask in enumerate(patterns):
|
| 191 |
+
coefficients = z[sparse[membership==i]].copy()
|
| 192 |
+
coefficients[:,~mask] = 0
|
| 193 |
+
rec = recover_sparse(grid,coefficients,mask,phase)
|
| 194 |
+
sparse_errors.extend(relative_orbit_error(coefficients,rec,phase).tolist())
|
| 195 |
+
assert max(sparse_errors)<2e-8
|
| 196 |
+
|
| 197 |
+
# A predeclared balanced test compares each known template with its point inversion.
|
| 198 |
+
nontrivial = relative_orbit_error(z[valid],z[valid].conj(),phase)>1e-8
|
| 199 |
+
test = rng.choice(valid[nontrivial],size=512,replace=False)
|
| 200 |
+
labels = np.tile([1,-1],len(test))
|
| 201 |
+
prototypes = np.repeat(z[test],2,axis=0)
|
| 202 |
+
actual = np.where(labels[:,None]>0,prototypes,prototypes.conj())
|
| 203 |
+
rows = plan.rows[plan.uses_tap]
|
| 204 |
+
truth = grid.product(prototypes, rows).imag
|
| 205 |
+
b,c = grid.representative[rows[:,1]],grid.representative[rows[:,2]]
|
| 206 |
+
a = rows[:,0]
|
| 207 |
+
tap_factor = plan.tap_factors[plan.uses_tap]
|
| 208 |
+
accuracy = []
|
| 209 |
+
for count in (32,128):
|
| 210 |
+
for sigma in (.5,1,2,4):
|
| 211 |
+
prediction, full = [],[]
|
| 212 |
+
for j in range(0,len(labels),32):
|
| 213 |
+
k = min(j+32,len(labels))
|
| 214 |
+
view = actual[j:k,None,:]*phase[rng.integers(grid.size,size=(k-j,count))]
|
| 215 |
+
re_noise = rng.normal(0,sigma,size=(k-j,count,grid.size))
|
| 216 |
+
im_noise = rng.normal(0,sigma,size=re_noise.shape)
|
| 217 |
+
cosine = (view.real+re_noise)[...,grid.representative]
|
| 218 |
+
positive_sine = (view.imag+im_noise)[...,grid.representative]
|
| 219 |
+
sine = positive_sine*grid.sign*np.where(grid.real,0,1)
|
| 220 |
+
tap = sine@plan.weights
|
| 221 |
+
mixed = (tap[...,None]*(cosine[...,b]*cosine[...,c]-sigma**2*(b==c))).mean(axis=1)
|
| 222 |
+
score = (mixed*tap_factor*truth[j:k]).sum(axis=1)
|
| 223 |
+
prediction.extend(np.where(score>=0,1,-1))
|
| 224 |
+
uncompressed = cosine+1j*sine
|
| 225 |
+
F = grid.product(uncompressed, rows).mean(axis=1).imag
|
| 226 |
+
full.extend(np.where((F*truth[j:k]).sum(axis=1)>=0,1,-1))
|
| 227 |
+
accuracy.append({"views_per_trial": count, "coordinate_noise_sd": sigma,
|
| 228 |
+
"one_tap_accuracy": float(np.mean(prediction==labels)),
|
| 229 |
+
"full_view_accuracy": float(np.mean(full==labels)),
|
| 230 |
+
"cosine_only_bayes_accuracy": .5})
|
| 231 |
+
exemplar = test[:3]
|
| 232 |
+
example = {"template_indices": exemplar.tolist(),
|
| 233 |
+
"digit_labels": dataset["digit_labels"][exemplar].tolist()}
|
| 234 |
+
results={"dataset_images": len(images), "full_support_images": len(valid),
|
| 235 |
+
"reconstruction_max_relative_error": max(reconstruction_errors),
|
| 236 |
+
"sparse_support_images": len(sparse), "sparse_support_patterns": len(patterns),
|
| 237 |
+
"sparse_max_relative_error": max(sparse_errors),
|
| 238 |
+
"reconstruction_median_relative_error": float(np.median(reconstruction_errors)),
|
| 239 |
+
"one_tap_readings_per_view": int((grid.size+np.count_nonzero(grid.real))//2+1),
|
| 240 |
+
"full_readings_per_view": grid.size, "templates_in_recognition_test": len(test),
|
| 241 |
+
"trials_per_noisy_setting": len(labels), "recognition": accuracy, "example": example}
|
| 242 |
+
return results, images[exemplar]
|
| 243 |
+
|
| 244 |
+
|
| 245 |
+
def rgb_evidence(dataset):
|
| 246 |
+
grid = Grid((32,32))
|
| 247 |
+
plan = OneTap(grid)
|
| 248 |
+
phase = grid.phases(grid.points)
|
| 249 |
+
weight = np.ones(3)/np.sqrt(3)
|
| 250 |
+
factor_real = np.where(grid.real,1,2)
|
| 251 |
+
factor_imag = np.where(grid.real,0,grid.sign*2*np.sqrt(len(grid.odd)))
|
| 252 |
+
corners = dataset["rgb_corners"].tolist()
|
| 253 |
+
cases = []
|
| 254 |
+
best_examples = {}
|
| 255 |
+
for crops, filename in zip(dataset["rgb_crops"], dataset["rgb_names"]):
|
| 256 |
+
for source, (y,x) in zip(crops, corners):
|
| 257 |
+
patch = source.astype(float)/255
|
| 258 |
+
z = np.fft.fft2(patch, axes=(0,1),norm="ortho").reshape(grid.size,3)
|
| 259 |
+
anchor = z@weight
|
| 260 |
+
if np.min(abs(anchor[1:])) <= 1e-10:
|
| 261 |
+
cases.append({"source": Path(filename).name, "top_left": [y,x], "full_anchor": False})
|
| 262 |
+
continue
|
| 263 |
+
hviews = phase*anchor
|
| 264 |
+
c = hviews.real
|
| 265 |
+
tap_axis = hviews.imag@plan.weights
|
| 266 |
+
tap_dense = hviews[:,grid.odd].imag.sum(axis=1)/np.sqrt(len(grid.odd))
|
| 267 |
+
moments = plan.estimates(c,tap_axis)
|
| 268 |
+
magnitude = np.sqrt((c*c).mean(axis=0)*factor_real)
|
| 269 |
+
h_recovered = plan.recover(magnitude,moments,c[:,0].mean())
|
| 270 |
+
cos = (phase[:,:,None]*z).real
|
| 271 |
+
cross_re = np.einsum("pk,pkc->kc",c,cos)/grid.size*factor_real[:,None]
|
| 272 |
+
cross_im = np.einsum("p,pkc->kc",tap_dense,cos)/grid.size*factor_imag[:,None]
|
| 273 |
+
recovered = np.empty_like(z)
|
| 274 |
+
recovered[1:] = (cross_re[1:]+1j*cross_im[1:]).conj()/h_recovered[1:,None].conj()
|
| 275 |
+
recovered[0] = cos[:,0].mean(axis=0)
|
| 276 |
+
align = np.argmin(np.linalg.norm(phase*anchor-h_recovered,axis=1))
|
| 277 |
+
aligned = recovered*phase[align,:,None].conj()
|
| 278 |
+
error = float(np.linalg.norm(aligned-z)/np.linalg.norm(z))
|
| 279 |
+
non_dc_error = float(np.linalg.norm(aligned[1:]-z[1:])/np.linalg.norm(z[1:]))
|
| 280 |
+
cases.append({"source": Path(filename).name, "top_left": [y,x],
|
| 281 |
+
"full_anchor": True, "relative_error": error,
|
| 282 |
+
"relative_non_dc_error": non_dc_error})
|
| 283 |
+
contrast = float(np.std(patch,axis=(0,1)).mean())
|
| 284 |
+
best = best_examples.get(filename)
|
| 285 |
+
if best is None or contrast > best[0]:
|
| 286 |
+
pixels = np.fft.ifft2(aligned.reshape(32,32,3),axes=(0,1),norm="ortho").real
|
| 287 |
+
best_examples[filename] = (contrast,(Path(filename).name,patch,pixels))
|
| 288 |
+
maximum = max(r["relative_error"] for r in cases if r["full_anchor"])
|
| 289 |
+
non_dc = max(r["relative_non_dc_error"] for r in cases if r["full_anchor"])
|
| 290 |
+
assert max(maximum, non_dc) < 1e-8
|
| 291 |
+
return {"shape": [32,32,3], "crops_tested": len(cases),
|
| 292 |
+
"crops_with_full_anchor": sum(r["full_anchor"] for r in cases),
|
| 293 |
+
"max_relative_error": maximum, "max_relative_non_dc_error": non_dc,
|
| 294 |
+
"full_readings_per_view": 3*grid.size,
|
| 295 |
+
"two_tap_readings_per_view": int(3*(grid.size+np.count_nonzero(grid.real))//2+2),
|
| 296 |
+
"cases": cases}, [entry[1] for entry in best_examples.values()]
|
| 297 |
+
|
| 298 |
+
|
| 299 |
+
def make_figure(path, images, results):
|
| 300 |
+
import matplotlib
|
| 301 |
+
matplotlib.use("Agg")
|
| 302 |
+
import matplotlib.pyplot as plt
|
| 303 |
+
|
| 304 |
+
plt.rcParams.update({"font.family":"DejaVu Sans","font.size":10,
|
| 305 |
+
"axes.spines.top":False,"axes.spines.right":False})
|
| 306 |
+
f = plt.figure(figsize=(11,3.65),facecolor="white")
|
| 307 |
+
gs = f.add_gridspec(2,5, width_ratios=[1,1,1,.28,4], wspace=.3, hspace=.15)
|
| 308 |
+
inversion = (-np.arange(8))%8
|
| 309 |
+
for j in range(3):
|
| 310 |
+
for row in range(2):
|
| 311 |
+
ax=f.add_subplot(gs[row,j])
|
| 312 |
+
im=images[j] if row==0 else images[j][np.ix_(inversion,inversion)]
|
| 313 |
+
ax.imshow(im,cmap="cividis",vmin=0,vmax=16,interpolation="nearest")
|
| 314 |
+
ax.axis("off")
|
| 315 |
+
if j==0:
|
| 316 |
+
ax.text(-.16,.5,"Original" if row==0 else "Inverted",rotation=90,
|
| 317 |
+
ha="center",va="center",transform=ax.transAxes,color="#253a4a")
|
| 318 |
+
ax=f.add_subplot(gs[:,4])
|
| 319 |
+
data=[r for r in results["recognition"] if r["views_per_trial"]==128]
|
| 320 |
+
x=[r["coordinate_noise_sd"] for r in data]
|
| 321 |
+
ax.plot(x,[r["one_tap_accuracy"]*100 for r in data],lw=2.1,marker="o",
|
| 322 |
+
color="#0d756a",label="One fixed sine reading")
|
| 323 |
+
ax.plot(x,[r["full_view_accuracy"]*100 for r in data],lw=2.0,marker="s",
|
| 324 |
+
color="#46658a",label="Full-view cubic score")
|
| 325 |
+
ax.axhline(50,lw=1.4,color="#a34731",linestyle="--",label="Folded-data Bayes limit")
|
| 326 |
+
ax.set(xticks=x,ylim=(45,102), xlabel="Noise SD per real coordinate",ylabel="Pair accuracy (%)")
|
| 327 |
+
ax.legend(loc="lower left",fontsize=8)
|
| 328 |
+
ax.grid(alpha=.25)
|
| 329 |
+
ax.set_title("Point-inversion recognition · 128 views",fontweight="bold",color="#15304a")
|
| 330 |
+
f.subplots_adjust(left=.07,right=.98,top=.88,bottom=.17)
|
| 331 |
+
f.savefig(path,bbox_inches="tight")
|
| 332 |
+
plt.close(f)
|
| 333 |
+
|
| 334 |
+
|
| 335 |
+
def make_rgb_figure(path, examples):
|
| 336 |
+
import matplotlib.pyplot as plt
|
| 337 |
+
|
| 338 |
+
f, axes = plt.subplots(1,2*len(examples),figsize=(8.8,2.35),facecolor="white")
|
| 339 |
+
for i,(name,original,recovered) in enumerate(examples):
|
| 340 |
+
for j,image in enumerate((original,recovered)):
|
| 341 |
+
ax = axes[2*i+j]
|
| 342 |
+
ax.imshow(np.clip(image,0,1),interpolation="nearest")
|
| 343 |
+
ax.axis("off")
|
| 344 |
+
title = name.split(".")[0].capitalize()+(" reference" if j==0 else " recovered")
|
| 345 |
+
ax.set_title(title,color="#15304a",fontsize=10,fontweight="bold")
|
| 346 |
+
f.subplots_adjust(wspace=.15,left=.02,right=.98,bottom=.02,top=.85)
|
| 347 |
+
f.savefig(path,bbox_inches="tight")
|
| 348 |
+
plt.close(f)
|
| 349 |
+
|
| 350 |
+
|
| 351 |
+
def write_report_table(path, results):
|
| 352 |
+
p = results["population"]
|
| 353 |
+
d = results["digits"]
|
| 354 |
+
c = results["rgb"]
|
| 355 |
+
number = lambda x: ("%0.2e" % x).replace("e-", r"\times10^{-") + "}"
|
| 356 |
+
rows = [r"\begin{tabularx}{\linewidth}{@{}Xr@{}}",
|
| 357 |
+
r"\toprule Result & Measurement\\ \midrule",
|
| 358 |
+
f"Fully supported recorded digits, out of {d['dataset_images']:,} & {d['full_support_images']:,}"+r"\\",
|
| 359 |
+
r"Largest noiseless relative translation-orbit error on these digits & " +
|
| 360 |
+
"$"+number(d["reconstruction_max_relative_error"])+"$"+r"\\",
|
| 361 |
+
f"Digit readings per view, one-tap versus full & {d['one_tap_readings_per_view']} versus {d['full_readings_per_view']}"+r"\\",
|
| 362 |
+
f"Remaining digits with spectral holes, recovered with two taps & {d['sparse_support_images']}"+r"\\",
|
| 363 |
+
r"Largest noiseless relative error on those digits & $"+
|
| 364 |
+
number(d["sparse_max_relative_error"])+"$"+r"\\",
|
| 365 |
+
f"Natural-image RGB crops reconstructed, out of {c['crops_tested']} & {c['crops_with_full_anchor']}"+r"\\",
|
| 366 |
+
r"Largest noiseless relative error on RGB crops, excluding DC & $"+
|
| 367 |
+
number(c["max_relative_non_dc_error"])+"$"+r"\\",
|
| 368 |
+
f"RGB readings per view, two-tap versus full & {c['two_tap_readings_per_view']:,} versus {c['full_readings_per_view']:,}"+r"\\",
|
| 369 |
+
f"Two-tap transfer across {p['two_tap_cases']} synthetic cases & {p['tested_imaginary_coordinates']:,} cubic coordinate checks"+r"\\",
|
| 370 |
+
r"\bottomrule\end{tabularx}", "",
|
| 371 |
+
r"\medskip\noindent\textbf{Noisy point-inversion decisions.} Each setting uses 1,024 balanced trials across 512 known digit templates.",
|
| 372 |
+
"",
|
| 373 |
+
r"{\small\begin{tabularx}{\linewidth}{@{}rcc>{\centering\arraybackslash}Xr@{}}",
|
| 374 |
+
r"\toprule Views & Noise SD & One fixed tap & Full-view cubic & Folded Bayes\\ \midrule"]
|
| 375 |
+
for r in d["recognition"]:
|
| 376 |
+
rows.append(f"{r['views_per_trial']} & {r['coordinate_noise_sd']:g} & "
|
| 377 |
+
f"{100*r['one_tap_accuracy']:.1f}"+r"\% & "+
|
| 378 |
+
f"{100*r['full_view_accuracy']:.1f}"+r"\% & 50.0\%\\")
|
| 379 |
+
rows += [r"\bottomrule\end{tabularx}}", ""]
|
| 380 |
+
path.write_text("\n".join(rows))
|
| 381 |
+
|
| 382 |
+
|
| 383 |
+
def main():
|
| 384 |
+
parser = ArgumentParser()
|
| 385 |
+
parser.add_argument("--output",type=Path,default=Path("."))
|
| 386 |
+
args = parser.parse_args()
|
| 387 |
+
args.output.mkdir(parents=True,exist_ok=True)
|
| 388 |
+
rng = np.random.default_rng(471826)
|
| 389 |
+
report = {"seed":471826,"population":population_checks(rng),
|
| 390 |
+
"tensor_nullities":[signed_graph_nullity(p) for p in (3,5,7,11,13,17)],
|
| 391 |
+
"sparse_coded":coded_higher_order(rng),"conditioning":stability_checks(rng)}
|
| 392 |
+
with np.load(Path(__file__).with_name("benchmark_inputs.npz"), allow_pickle=False) as dataset:
|
| 393 |
+
report["digits"],images = image_evidence(rng,dataset)
|
| 394 |
+
report["rgb"],examples = rgb_evidence(dataset)
|
| 395 |
+
(args.output/"results.json").write_text(json.dumps(report,indent=2)+"\n")
|
| 396 |
+
make_figure(args.output/"recognition.pdf",images,report["digits"])
|
| 397 |
+
make_rgb_figure(args.output/"rgb.pdf",examples)
|
| 398 |
+
write_report_table(args.output/"results_table.tex",report)
|
| 399 |
+
print(json.dumps(report,indent=2))
|
| 400 |
+
|
| 401 |
+
|
| 402 |
+
if __name__=="__main__":
|
| 403 |
+
main()
|
tools/build_parquet.py
ADDED
|
@@ -0,0 +1,194 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Build the Hub Parquet files from the frozen Xoryn II benchmark."""
|
| 3 |
+
|
| 4 |
+
import io
|
| 5 |
+
import json
|
| 6 |
+
from pathlib import Path
|
| 7 |
+
import sys
|
| 8 |
+
|
| 9 |
+
from datasets import ClassLabel, Features, Image, List, Value
|
| 10 |
+
import numpy as np
|
| 11 |
+
from PIL import Image as PillowImage
|
| 12 |
+
import pyarrow as pa
|
| 13 |
+
import pyarrow.parquet as parquet
|
| 14 |
+
|
| 15 |
+
REPO = Path(__file__).resolve().parents[1]
|
| 16 |
+
SOURCE = REPO / "source"
|
| 17 |
+
sys.path.insert(0, str(SOURCE))
|
| 18 |
+
|
| 19 |
+
import verify # The published benchmark code, imported without running its CLI.
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
class BenchmarkRng:
|
| 23 |
+
"""Record the selected recognition cohort while replaying the benchmark RNG."""
|
| 24 |
+
|
| 25 |
+
def __init__(self, seed, cohort_size):
|
| 26 |
+
self.generator = np.random.default_rng(seed)
|
| 27 |
+
self.cohort_size = cohort_size
|
| 28 |
+
self.recognition_ids = None
|
| 29 |
+
|
| 30 |
+
def __getattr__(self, name):
|
| 31 |
+
return getattr(self.generator, name)
|
| 32 |
+
|
| 33 |
+
def choice(self, population, size=None, replace=True, **kwargs):
|
| 34 |
+
picked = self.generator.choice(population, size=size, replace=replace, **kwargs)
|
| 35 |
+
if size == self.cohort_size and not replace:
|
| 36 |
+
if self.recognition_ids is not None:
|
| 37 |
+
raise RuntimeError("The recognition cohort was captured twice")
|
| 38 |
+
self.recognition_ids = np.array(picked, copy=True)
|
| 39 |
+
return picked
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
def png(array):
|
| 43 |
+
buffer = io.BytesIO()
|
| 44 |
+
PillowImage.fromarray(array).save(buffer, format="PNG", compress_level=6)
|
| 45 |
+
return {"bytes": buffer.getvalue(), "path": None}
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def encode(columns, features):
|
| 49 |
+
return pa.Table.from_pydict(features.encode_batch(columns), schema=features.arrow_schema)
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
def encode_rows(rows, features):
|
| 53 |
+
return encode({name: [row[name] for row in rows] for name in features}, features)
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def build_digits(data, report):
|
| 57 |
+
digits = data["digits"]
|
| 58 |
+
grid = verify.Grid((8, 8))
|
| 59 |
+
support = verify.exact_digit_support(grid, digits)
|
| 60 |
+
count = support.sum(axis=1)
|
| 61 |
+
full = count == 63
|
| 62 |
+
record = report["digits"]
|
| 63 |
+
if int(full.sum()) != record["full_support_images"]:
|
| 64 |
+
raise ValueError("The exact support count differs from the manuscript")
|
| 65 |
+
|
| 66 |
+
rng = BenchmarkRng(report["seed"], record["templates_in_recognition_test"])
|
| 67 |
+
verify.population_checks(rng)
|
| 68 |
+
verify.coded_higher_order(rng)
|
| 69 |
+
verify.stability_checks(rng)
|
| 70 |
+
replay, _ = verify.image_evidence(rng, data)
|
| 71 |
+
if (replay["example"] != record["example"]
|
| 72 |
+
or replay["recognition"] != record["recognition"]
|
| 73 |
+
or rng.recognition_ids is None):
|
| 74 |
+
raise ValueError("The recognition cohort or scores do not match the published run")
|
| 75 |
+
membership = np.isin(np.arange(len(digits)), rng.recognition_ids)
|
| 76 |
+
if int(membership.sum()) != record["templates_in_recognition_test"] or not full[membership].all():
|
| 77 |
+
raise ValueError("Unexpected noisy-test cohort")
|
| 78 |
+
|
| 79 |
+
features = Features(
|
| 80 |
+
{
|
| 81 |
+
"sample_id": Value("int32"),
|
| 82 |
+
"image": Image(),
|
| 83 |
+
"pixels": List(List(Value("uint8"), length=8), length=8),
|
| 84 |
+
"label": ClassLabel(names=[str(i) for i in range(10)]),
|
| 85 |
+
"non_dc_active_count": Value("int32"),
|
| 86 |
+
"full_fourier_support": Value("bool"),
|
| 87 |
+
"readings_per_view": Value("int32"),
|
| 88 |
+
"in_noisy_recognition_test": Value("bool"),
|
| 89 |
+
}
|
| 90 |
+
)
|
| 91 |
+
# Round 255*x/16 to the nearest integer, using NumPy ties-to-even.
|
| 92 |
+
display = np.rint(digits.astype(float) * (255 / 16)).astype(np.uint8)
|
| 93 |
+
columns = {
|
| 94 |
+
"sample_id": list(range(len(digits))),
|
| 95 |
+
"image": [png(array) for array in display],
|
| 96 |
+
"pixels": digits.tolist(),
|
| 97 |
+
"label": data["digit_labels"].tolist(),
|
| 98 |
+
"non_dc_active_count": count.tolist(),
|
| 99 |
+
"full_fourier_support": full.tolist(),
|
| 100 |
+
"readings_per_view": np.where(full, 35, 36).tolist(),
|
| 101 |
+
"in_noisy_recognition_test": membership.tolist(),
|
| 102 |
+
}
|
| 103 |
+
return encode(columns, features)
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
def build_rgb(data, report):
|
| 107 |
+
features = Features(
|
| 108 |
+
{
|
| 109 |
+
"crop_id": Value("string"),
|
| 110 |
+
"image": Image(),
|
| 111 |
+
"source": Value("string"),
|
| 112 |
+
"top_left_y": Value("int32"),
|
| 113 |
+
"top_left_x": Value("int32"),
|
| 114 |
+
"full_anchor": Value("bool"),
|
| 115 |
+
"relative_error": Value("float64"),
|
| 116 |
+
"relative_non_dc_error": Value("float64"),
|
| 117 |
+
"readings_per_view": Value("int32"),
|
| 118 |
+
}
|
| 119 |
+
)
|
| 120 |
+
rgb = report["rgb"]
|
| 121 |
+
outcomes = {(case["source"], *case["top_left"]): case for case in rgb["cases"]}
|
| 122 |
+
rows = []
|
| 123 |
+
for name, crops in zip(data["rgb_names"].tolist(), data["rgb_crops"]):
|
| 124 |
+
for crop, (y, x) in zip(crops, data["rgb_corners"].tolist()):
|
| 125 |
+
outcome = outcomes.pop((name, y, x))
|
| 126 |
+
rows.append(
|
| 127 |
+
{
|
| 128 |
+
"crop_id": f"{Path(name).stem}-y{y}-x{x}",
|
| 129 |
+
"image": png(crop),
|
| 130 |
+
"source": name,
|
| 131 |
+
"top_left_y": y,
|
| 132 |
+
"top_left_x": x,
|
| 133 |
+
"full_anchor": outcome["full_anchor"],
|
| 134 |
+
"relative_error": outcome.get("relative_error"),
|
| 135 |
+
"relative_non_dc_error": outcome.get("relative_non_dc_error"),
|
| 136 |
+
"readings_per_view": rgb["two_tap_readings_per_view"],
|
| 137 |
+
}
|
| 138 |
+
)
|
| 139 |
+
if outcomes:
|
| 140 |
+
raise ValueError("The frozen inputs do not contain every recorded RGB crop")
|
| 141 |
+
return encode_rows(rows, features)
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
def build_recognition(report):
|
| 145 |
+
features = Features(
|
| 146 |
+
{
|
| 147 |
+
"views_per_trial": Value("int32"),
|
| 148 |
+
"coordinate_noise_sd": Value("float64"),
|
| 149 |
+
"one_tap_accuracy": Value("float64"),
|
| 150 |
+
"full_view_accuracy": Value("float64"),
|
| 151 |
+
"cosine_only_bayes_accuracy": Value("float64"),
|
| 152 |
+
"trials": Value("int32"),
|
| 153 |
+
"templates": Value("int32"),
|
| 154 |
+
"one_tap_correct": Value("int32"),
|
| 155 |
+
"full_view_correct": Value("int32"),
|
| 156 |
+
}
|
| 157 |
+
)
|
| 158 |
+
digits = report["digits"]
|
| 159 |
+
trials = digits["trials_per_noisy_setting"]
|
| 160 |
+
rows = []
|
| 161 |
+
for condition in digits["recognition"]:
|
| 162 |
+
one_tap = round(condition["one_tap_accuracy"] * trials)
|
| 163 |
+
full = round(condition["full_view_accuracy"] * trials)
|
| 164 |
+
if one_tap / trials != condition["one_tap_accuracy"] or full / trials != condition["full_view_accuracy"]:
|
| 165 |
+
raise ValueError("The stored accuracies do not match integer counts")
|
| 166 |
+
rows.append(
|
| 167 |
+
{
|
| 168 |
+
**condition,
|
| 169 |
+
"trials": trials,
|
| 170 |
+
"templates": digits["templates_in_recognition_test"],
|
| 171 |
+
"one_tap_correct": one_tap,
|
| 172 |
+
"full_view_correct": full,
|
| 173 |
+
}
|
| 174 |
+
)
|
| 175 |
+
return encode_rows(rows, features)
|
| 176 |
+
|
| 177 |
+
|
| 178 |
+
def main():
|
| 179 |
+
report = json.loads((SOURCE / "results.json").read_text(encoding="utf-8"))
|
| 180 |
+
with np.load(SOURCE / "benchmark_inputs.npz", allow_pickle=False) as data:
|
| 181 |
+
datasets = {
|
| 182 |
+
"digits": build_digits(data, report),
|
| 183 |
+
"rgb_crops": build_rgb(data, report),
|
| 184 |
+
"recognition_results": build_recognition(report),
|
| 185 |
+
}
|
| 186 |
+
destination = REPO / "data"
|
| 187 |
+
destination.mkdir(exist_ok=True)
|
| 188 |
+
for key, dataset in datasets.items():
|
| 189 |
+
parquet.write_table(dataset, destination / f"{key}.parquet")
|
| 190 |
+
print(f"{key}: {len(dataset):,} rows")
|
| 191 |
+
|
| 192 |
+
|
| 193 |
+
if __name__ == "__main__":
|
| 194 |
+
main()
|
tools/requirements-export.txt
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
-r ../source/requirements.txt
|
| 2 |
+
datasets==4.4.1
|
| 3 |
+
pyarrow==21.0.0
|
| 4 |
+
Pillow>=10
|