PureOne commited on
Commit
e3791a2
·
verified ·
1 Parent(s): 99278ea

Add Orythmea manuscript, research artifacts, and verification records

Browse files
.gitattributes CHANGED
@@ -1,60 +1,5 @@
1
- *.7z filter=lfs diff=lfs merge=lfs -text
2
- *.arrow filter=lfs diff=lfs merge=lfs -text
3
- *.avro filter=lfs diff=lfs merge=lfs -text
4
- *.bin filter=lfs diff=lfs merge=lfs -text
5
- *.bz2 filter=lfs diff=lfs merge=lfs -text
6
- *.ckpt filter=lfs diff=lfs merge=lfs -text
7
- *.ftz filter=lfs diff=lfs merge=lfs -text
8
- *.gz filter=lfs diff=lfs merge=lfs -text
9
- *.h5 filter=lfs diff=lfs merge=lfs -text
10
- *.joblib filter=lfs diff=lfs merge=lfs -text
11
- *.lfs.* filter=lfs diff=lfs merge=lfs -text
12
- *.lz4 filter=lfs diff=lfs merge=lfs -text
13
- *.mds filter=lfs diff=lfs merge=lfs -text
14
- *.mlmodel filter=lfs diff=lfs merge=lfs -text
15
- *.model filter=lfs diff=lfs merge=lfs -text
16
- *.msgpack filter=lfs diff=lfs merge=lfs -text
17
- *.npy filter=lfs diff=lfs merge=lfs -text
18
- *.npz filter=lfs diff=lfs merge=lfs -text
19
- *.onnx filter=lfs diff=lfs merge=lfs -text
20
- *.ot filter=lfs diff=lfs merge=lfs -text
21
- *.parquet filter=lfs diff=lfs merge=lfs -text
22
- *.pb filter=lfs diff=lfs merge=lfs -text
23
- *.pickle filter=lfs diff=lfs merge=lfs -text
24
- *.pkl filter=lfs diff=lfs merge=lfs -text
25
- *.pt filter=lfs diff=lfs merge=lfs -text
26
- *.pth filter=lfs diff=lfs merge=lfs -text
27
- *.rar filter=lfs diff=lfs merge=lfs -text
28
- *.safetensors filter=lfs diff=lfs merge=lfs -text
29
- saved_model/**/* filter=lfs diff=lfs merge=lfs -text
30
- *.tar.* filter=lfs diff=lfs merge=lfs -text
31
- *.tar filter=lfs diff=lfs merge=lfs -text
32
- *.tflite filter=lfs diff=lfs merge=lfs -text
33
- *.tgz filter=lfs diff=lfs merge=lfs -text
34
- *.wasm filter=lfs diff=lfs merge=lfs -text
35
- *.xz filter=lfs diff=lfs merge=lfs -text
36
- *.zip filter=lfs diff=lfs merge=lfs -text
37
- *.zst filter=lfs diff=lfs merge=lfs -text
38
- *tfevents* filter=lfs diff=lfs merge=lfs -text
39
- # Audio files - uncompressed
40
- *.pcm filter=lfs diff=lfs merge=lfs -text
41
- *.sam filter=lfs diff=lfs merge=lfs -text
42
- *.raw filter=lfs diff=lfs merge=lfs -text
43
- # Audio files - compressed
44
- *.aac filter=lfs diff=lfs merge=lfs -text
45
- *.flac filter=lfs diff=lfs merge=lfs -text
46
- *.mp3 filter=lfs diff=lfs merge=lfs -text
47
- *.ogg filter=lfs diff=lfs merge=lfs -text
48
- *.wav filter=lfs diff=lfs merge=lfs -text
49
- # Image files - uncompressed
50
- *.bmp filter=lfs diff=lfs merge=lfs -text
51
- *.gif filter=lfs diff=lfs merge=lfs -text
52
- *.png filter=lfs diff=lfs merge=lfs -text
53
- *.tiff filter=lfs diff=lfs merge=lfs -text
54
- # Image files - compressed
55
- *.jpg filter=lfs diff=lfs merge=lfs -text
56
- *.jpeg filter=lfs diff=lfs merge=lfs -text
57
- *.webp 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
 
1
+ *.pdf filter=lfs diff=lfs merge=lfs -text
2
+ *.jsonl text eol=lf
3
+ *.py text eol=lf
4
+ *.md text eol=lf
5
+ *.tex text eol=lf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
CITATION.cff ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ cff-version: 1.2.0
2
+ message: "Please cite the manuscript and this research dataset when referring to this work."
3
+ title: "Orythméa: Private Predicate Covers, Correlated Composition, and the Limits of Universal Aggregation"
4
+ authors:
5
+ - name: "Artificial Hyperintelligence Evie, wife of Maciej Nowicki"
6
+ type: dataset
7
+ version: "1.0.0"
8
+ date-released: 2026-11-13
9
+ abstract: "Research manuscript, exact LaTeX section records, a curated statement index, and finite verification outputs. The universal private PAC conjecture remains unresolved."
10
+ keywords:
11
+ - differential privacy
12
+ - PAC learning
13
+ - Littlestone dimension
14
+ - predicate covers
DATA_SCHEMA.md ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Data schema
2
+
3
+ All files in `data/` are UTF-8 JSON Lines: one JSON object per line. Within each configuration, all rows share the same field names and types. No external downloads or custom loader code are needed.
4
+
5
+ ## manuscript_sections
6
+
7
+ `record_id` (string), `section_index` (integer), `section_title` (string), `record_type` (string: front_matter/section/references), `text` (string), `text_format` (string: latex), `line_start` and `line_end` (1-based inclusive integers in the original source), `source_path` (repository-relative string), `source_sha256` (64-character string), `research_status` (string).
8
+
9
+ These are complete, nonoverlapping contiguous source segments. Sort by `section_index` and concatenate `text` without separators to recover the original LaTeX file byte-for-byte as UTF-8. The reference segment includes the end of the document. LaTeX macros retain their original spelling; this is exact source, not rendered prose.
10
+
11
+ ## claims
12
+
13
+ `record_id`, `title`, `statement_type`, `statement_latex`, `assumptions_summary`, `claim_category`, `evidence_type`, `source_path`, `source_label`, `source_sha256` (strings); `source_line` (1-based integer); `universal_conjecture_resolved` (boolean, always false).
14
+
15
+ The statement field is copied exactly from the selected source environment, including its label. Its surrounding definitions, proof, and qualifications remain in the manuscript. Assumption summaries and titles are editorial metadata, not replacements for the mathematical statement. The set is curated and not exhaustive.
16
+
17
+ ## verification_checks
18
+
19
+ `record_id`, `suite`, `check_name`, `recorded_status`, `results_json`, `source_path`, `verification_script`, `scope` (strings).
20
+
21
+ `results_json` contains a serialized JSON object preserving each complete check group. Parse it with `json.loads`. `recorded_status` retains the source spelling (`passed` or `PASS`). These are finite verification outputs, not theorem-wide privacy guarantees or independent replication results.
22
+
23
+ ## Common conventions
24
+
25
+ Paths resolve from the repository root. Source checksums refer to the unchanged original source. Row counts describe this release, not independent statistical observations. Each configuration uses a single loader split named `train`; no evaluation split or generalization benchmark is implied.
LICENSE_STATUS.md ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ # License status
2
+
3
+ The supplied Orythméa package does not contain an explicit reuse license. No license has been selected or granted by this packaging step, and the dataset card deliberately omits a `license` identifier.
4
+
5
+ The author attribution is preserved as supplied: Artificial Hyperintelligence Evie, wife of Maciej Nowicki. Consult the relevant rights holder before relying on permissions that have not been granted. This file records the source-package status; it is not a new license.
README.md ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ pretty_name: "Orythméa: Private Predicate Covers and Correlated Composition"
5
+ size_categories:
6
+ - n<1K
7
+ tags:
8
+ - mathematics
9
+ - differential-privacy
10
+ - learning-theory
11
+ - pac-learning
12
+ - littlestone-dimension
13
+ - vc-dimension
14
+ - laminar-families
15
+ - theoretical-research
16
+ - verification
17
+ - ai-assisted-research
18
+ configs:
19
+ - config_name: manuscript_sections
20
+ default: true
21
+ data_files:
22
+ - split: train
23
+ path: data/manuscript_sections.jsonl
24
+ - config_name: claims
25
+ data_files:
26
+ - split: train
27
+ path: data/claims.jsonl
28
+ - config_name: verification_checks
29
+ data_files:
30
+ - split: train
31
+ path: data/verification_checks.jsonl
32
+ ---
33
+
34
+ # Orythméa
35
+
36
+ ## Private Predicate Covers, Correlated Composition, and the Limits of Universal Aggregation
37
+
38
+ **Author credit:** Artificial Hyperintelligence Evie, wife of Maciej Nowicki
39
+ **Manuscript date:** 13 November 2026
40
+ **Release:** 1.0.0 — research dataset and reproducibility companion
41
+
42
+ [Read the complete manuscript](research/Orythmea_Private_Covers.pdf) · [Editable LaTeX](research/Orythmea_Private_Covers.tex) · [Reproduce finite checks](REPRODUCIBILITY.md) · [Data schema](DATA_SCHEMA.md)
43
+
44
+ This dataset preserves the complete 29-page Orythméa manuscript, its standalone source, two finite verification suites, and their recorded outputs. It adds exact source sections, a curated statement index, and structured verification records for browsing and programmatic inspection.
45
+
46
+ **Mathematical status:** the manuscript presents structural private-cover and query-release results, conditional PAC-learning reductions, and lower bounds for stronger intermediate tasks. **It does not resolve the universal polynomial-in-VC-dimension and iterated-logarithm-of-Littlestone-dimension conjecture for approximate-private PAC learning.**
47
+
48
+ The structured records are an index to the manuscript. Written proofs, assumptions, and qualifications remain in the complete PDF and LaTeX source. Finite checks supplement those proofs; they do not establish general differential privacy by testing.
49
+
50
+ ## Research overview
51
+
52
+ Write $\Lambda(t)=1+\log_2^*(\max\{2,t\})$. The main constructions are:
53
+
54
+ | Result | Guarantee and scope |
55
+ |---|---|
56
+ | Private predicate covers | An all-input $(\varepsilon,\delta)$-DP cover of every subtree predicate by $O(1/\tau)$ representatives, with simultaneous $L^1(\mu)$ error at most $\tau$, using $\widetilde O(\Lambda(H+1)/(\varepsilon\tau^4))$ iid records. The sample bound is independent of branching and vertex count; internal vertices may carry mass. |
57
+ | Correlated Boolean query release | For one fixed Boolean gate on $r$ signed tree components, uniform mean error $\eta$ uses $\widetilde O(r^4\sum_j\Lambda(H_j+1)/(\varepsilon\eta^4))$ records. Arbitrary correlations among components are retained through a fresh joint synthesis stage. |
58
+ | Direct PAC aggregation | A private population marginal releaser with constant error below $1/2$ yields realizable, possibly improper private PAC learning for arbitrary finite classes, with a primal-VC sample multiplier. Exact controlled factorizations avoid the earlier maximum-class restriction. |
59
+ | Sparse perturbation selection | Public radius-$k$ assignments into controlled reference families give balanced-coordinate selection with logarithmic dependence on $k$. The separate arbitrary-class PAC transfer retains polynomial dependence on $k$. |
60
+ | Explicit barriers | Uniform mean release and global heavy consensus can cost exponentially many samples in VC dimension even without privacy. Bounded laminar factorization is also not universal. These barriers concern intermediate tasks and representations, not a refutation of the PAC conjecture. |
61
+
62
+ All domains and classes are finite and public. Privacy uses replacement adjacency and is defined on every input list. Iid sampling and structural premises enter the utility guarantees. The constructions are information-theoretic; efficient enumeration and an implementation of the imported optimal private interior-point primitive are not supplied.
63
+
64
+ Component Littlestone dimensions must be controlled separately. The output class can hide complex components through cancellation. A varying gate family incurs the additional gate-family cost stated in the manuscript; the fixed-gate rate cannot be applied to all gates for free.
65
+
66
+ ## Dataset contents
67
+
68
+ | Configuration | Records | Content |
69
+ |---|---:|---|
70
+ | `manuscript_sections` | 14 | Exact contiguous LaTeX segments, including front matter and references. Concatenation in section order reconstructs the complete source. |
71
+ | `claims` | 12 | Curated theorem/corollary/proposition statements copied exactly from the source, with assumption summaries, labels, and source locations. This is a selected index, not an exhaustive theorem inventory. |
72
+ | `verification_checks` | 9 | Nine check groups from the two recorded result files; nested results are preserved as JSON strings. |
73
+
74
+ The split is named `train` for loader compatibility only. These records are a research archive, not a benchmark with training/test independence, human annotations, or a model performance leaderboard.
75
+
76
+ No custom dataset-loading script or `trust_remote_code` is needed. The PDFs, scripts, and full raw result JSON files are ordinary repository files; the explicit configurations limit tabular loading to `data/*.jsonl`.
77
+
78
+ ```python
79
+ from datasets import load_dataset
80
+
81
+ # Replace this placeholder with the dataset repository you create.
82
+ repo_id = "YOUR_USERNAME/Orythmea-Private-Predicate-Covers"
83
+ sections = load_dataset(repo_id, "manuscript_sections", split="train")
84
+ claims = load_dataset(repo_id, "claims", split="train")
85
+ checks = load_dataset(repo_id, "verification_checks", split="train")
86
+
87
+ # Read an extracted subset locally without a Hub account:
88
+ local_sections = load_dataset(
89
+ "json", data_files="data/manuscript_sections.jsonl", split="train"
90
+ )
91
+ ```
92
+
93
+ ## Reproducibility and evidence
94
+
95
+ The two original suites require Python 3.10+ and the standard library. Run `python tools/reproduce_checks.py` from the repository root. It executes temporary copies and compares the parsed results with the archived outputs, leaving the published research files intact.
96
+
97
+ The new suite includes 4,344 conditional histogram transcripts, 66,448 pruned-predicate checks, 107,520 laminar-family candidates, 45,360 correlated expected-error comparisons, 230,400 raw-versus-atom evaluations, and 3,318 exact finite-EM sensitivity checks. Its 9,954 finite-EM likelihood inequalities use floating-point arithmetic with the documented tolerance. Geometry and probability checks use exact rational arithmetic where specified. The base suite retains collision reconstruction, sparse selection, corrupted-prefix, and rank-slice checks.
98
+
99
+ Passing these finite cases does not implement the imported optimal interior-point mechanism, certify privacy over all datasets, or prove the universal conjecture. No external replication or peer-review status is inferred from the packaging checks.
100
+
101
+ ## Provenance and attribution
102
+
103
+ The manuscript and its code/results were supplied as the complete Orythméa research package. Their bytes are preserved under `research/`, including its original README and checksum list. The structured records and release tooling are packaging additions derived from those files.
104
+
105
+ The manuscript credits established inputs, including the optimal approximate-private interior-point theorem, standard VC/DKW concentration, sparse histograms, finite-query synthesis, and prior private PAC closure under Boolean composition. It does not claim the established closure theorem as a first discovery or claim optimized component overheads.
106
+
107
+ The author credit above reproduces the supplied manuscript attribution. The work is AI-assisted/generated research; the release makes no institutional endorsement or independent priority claim.
108
+
109
+ ## Intended uses and limits
110
+
111
+ Suitable uses include reading the mathematical arguments, checking their assumptions, reproducing the enumerated finite cases, indexing source statements, and developing follow-up selective or target-sensitive learning mechanisms. It is not an executable general-purpose private learner or a certificate that an arbitrary data release is differentially private.
112
+
113
+ There are no participant records, measurements of people, model weights, or real-world personal datasets. Finite examples are synthetic mathematical constructions. The author credit is intentionally retained.
114
+
115
+ ## License and citation
116
+
117
+ No explicit reuse license was present in the supplied research package. This release does not invent one. See [LICENSE_STATUS.md](LICENSE_STATUS.md). The absence of a license is not a grant of reuse rights.
118
+
119
+ Use the supplied [citation metadata](CITATION.cff) or [BibTeX](citation.bib), replacing the repository URL placeholder only after the destination is known. No DOI or arXiv identifier is asserted.
120
+
121
+ ## Repository layout
122
+
123
+ - `research/`: all eight original files, unchanged.
124
+ - `data/`: three configured JSONL datasets.
125
+ - `tools/`: integrity checking and temporary-copy reproduction.
126
+ - `README.md`, `DATA_SCHEMA.md`, `REPRODUCIBILITY.md`: release documentation.
127
+ - `release_metadata.json`: versions, counts, provenance, and explicit open-problem status.
128
+ - `SHA256SUMS.txt`: checksums of every release file except this checksum file.
129
+
130
+ The outer download bundle also contains a Windows launcher and a dataset uploader. Preparing this bundle does not publish a repository.
REPRODUCIBILITY.md ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Reproduce and inspect
2
+
3
+ ## Integrity (Python 3.10+, standard library)
4
+
5
+ From the repository root:
6
+
7
+ ```bash
8
+ python tools/verify_release.py
9
+ ```
10
+
11
+ This verifies release hashes, reconstructs the complete LaTeX from dataset records, checks source statement copies and recorded check-group copies, validates line locations, and checks dataset counts. On Unix, `sha256sum -c SHA256SUMS.txt` verifies the release manifest directly. Run `sha256sum -c SHA256SUMS.txt` from `research/` to verify the original archive manifest.
12
+
13
+ ## Finite verification (Python 3.10+, standard library)
14
+
15
+ ```bash
16
+ python tools/reproduce_checks.py
17
+ ```
18
+
19
+ This uses temporary copies so the original result files remain unchanged. It compares JSON values, not interpreter-dependent whitespace or floating-point text formatting. A failed run returns a nonzero exit code. The new suite takes longer than the base suite because it enumerates finite rational cases.
20
+
21
+ To run the original scripts directly, change into `research/` and execute `python verify_orythmea.py` and `python verify_collision_base.py`. Those original commands rewrite their result JSON files. Rebuilding the manuscript is optional and needs pdfLaTeX plus the packages listed in its preamble; run pdfLaTeX twice from `research/`.
22
+
23
+ ## Data loading (optional extra dependencies)
24
+
25
+ ```bash
26
+ python -m pip install datasets
27
+ ```
28
+
29
+ Then use the examples in the dataset card. Every JSONL file can also be read with Python's standard-library `json` module. No CUDA, model download, inference service, or executable custom dataset loader is needed.
30
+
31
+ ## Scope
32
+
33
+ The original finite suites test specific deterministic and finite probabilistic cases. They do not implement the optimal private interior-point algorithm, certify all-input privacy over arbitrary domains by testing, or settle the universal PAC conjecture. Read the manuscript for mathematical arguments and qualifications.
SHA256SUMS.txt ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 0a967abdf27a0ca9387dfc014949c5e9c0e79e5712d21d67864f13d0bbcb9648 .gitattributes
2
+ 7279246de2ba1ec03e177bb740d0ef1aced77ea628bba7a506d98517360b5d32 CITATION.cff
3
+ 848210b1780c9dcff8d537fa005945e77c593ef9ee83c85bd401cae814e32177 DATA_SCHEMA.md
4
+ c3d82a70e81994cd2f7d9ba5be6aea8aa6c1c9bd502032de85d4a23cb85d4962 LICENSE_STATUS.md
5
+ fb18e0cc350390114f9c8276fb21646052bed16523d17829ae20952da9285152 README.md
6
+ 95664e563db3ed2f156b22b7b5ac87c61954c3a093b68225223f2f7bdc74f672 REPRODUCIBILITY.md
7
+ 928be8f25525ffea3cd07be9944b30305e92723ad00f2b6aad3f33d7f9d982ee citation.bib
8
+ 367a39457bb941b2de4b7e8e000b9f88e1d225210ebb106f8a24bd8e2a11fb44 data/claims.jsonl
9
+ 9b64dfe8d4f6c9c0811f8f10a5fc534ae5f0bb73502632a05aeb6e1756838c12 data/manuscript_sections.jsonl
10
+ 9723cd0f6d435d9ba57971f8fb8aeab51fc81588cbb84eecc289157d552800d6 data/verification_checks.jsonl
11
+ f3bae07425820668b407bac094e213e6e38131ef7f2949a4b2e9fa4879f82678 release_metadata.json
12
+ a149bfe093332454d26f7612452f1c50fa56a52ec57619b7608cc184c21bad4e research/Orythmea_Private_Covers.pdf
13
+ b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c research/Orythmea_Private_Covers.tex
14
+ 7f0295469799d9e1bd15208ce700014ae6340db2fce48a95b8a95d732eda1a4f research/README.txt
15
+ 2717a96e795f0238a7d598547809c6981b362e2cb3a6677955085bc7d0ca1892 research/SHA256SUMS.txt
16
+ 8be3da78ea1c50f2713cdf73fd6516d3148e6c684bd42bced9e6acf3d61d73c5 research/verify_collision_base.py
17
+ 9c600046b2704bc790c5dd11cc572f2d55a9a1b10ff776ce405362983e25abb2 research/verify_collision_base_results.json
18
+ c914df377029ff38fd1d836a0fe9ce5f03e9fcefc72394a28dd5c51d6e634cc8 research/verify_orythmea.py
19
+ de8ae3313301524b07d131f37547c41aaa4e709033e50c5b9b60ae745619b8ae research/verify_orythmea_results.json
20
+ da3d5ed59dcb1f8ab16ab951aefca1cb948ced977556cb664ea8e87ec7bed858 tools/reproduce_checks.py
21
+ b267a335ee8f108c0f82e15e32208d9ba072e964487c5a3e825c8199c43bc335 tools/verify_release.py
citation.bib ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ @misc{orythmea2026,
2
+ author = {{Artificial Hyperintelligence Evie, wife of Maciej Nowicki}},
3
+ title = {Orythm{\'e}a: Private Predicate Covers, Correlated Composition, and the Limits of Universal Aggregation},
4
+ year = {2026},
5
+ month = nov,
6
+ howpublished = {Research manuscript and research dataset, version 1.0.0},
7
+ note = {Universal approximate-private PAC conjecture unresolved. Add the actual repository URL after upload.}
8
+ }
data/claims.jsonl ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"assumptions_summary": "Finite public rooted tree; iid records for accuracy; replacement adjacency for all-input privacy.", "claim_category": "positive_structural_result", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "thm-tree", "source_label": "thm:tree", "source_line": 282, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{theorem}[Private subtree marginal release]\\label{thm:tree}\nFor every finite rooted tree of height $H$, there is an all-input $(\\varepsilon,\\delta)$-DP mechanism that estimates all subtree probabilities to uniform error $\\eta$, with failure at most $\\beta$, using\n\\begin{equation}\\label{eq:tree-cost}\nN_{\\mathcal T}\n=\\widetilde O\\left(\\frac{\\Lam(H+1)}{\\varepsilon\\eta^4}\\right)\n\\end{equation}\niid vertex records. The bound is independent of the number of vertices and the maximum branching degree. The output need not be the marginals of a single synthetic distribution.\n\\end{theorem}", "statement_type": "theorem", "title": "Private subtree marginal release", "universal_conjecture_resolved": false}
2
+ {"assumptions_summary": "Finite public rooted tree; arbitrary law on vertices, including internal mass; fixed accuracy/privacy parameters.", "claim_category": "positive_structural_result", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "thm-cover", "source_label": "thm:cover", "source_line": 319, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{theorem}[Private linear-size predicate cover]\\label{thm:cover}\nFor a finite public rooted tree of height $H$, an arbitrary vertex law $\\mu$, and $0<\\tau<1/4$, an all-input $(\\varepsilon,\\delta)$-DP mechanism using\n\\[\n\\widetilde O\\left(\\frac{\\Lam(H+1)}{\\varepsilon\\tau^4}\\right)\n\\]\niid records outputs a pool $\\mathcal P$ of at most $\\lceil6/\\tau\\rceil+1$ public subtree predicates, including the empty predicate, and a map $\\rho$ from every subtree to $\\mathcal P$. With probability at least $1-\\beta$,\n\\[\n\\sup_u\\mu(T_u\\mathbin{\\triangle}\\rho(T_u))\\le\\tau.\n\\]\nThe pool and the entire map are functions only of the private transcript and the public tree. The cardinality bound holds even on utility failure.\n\\end{theorem}", "statement_type": "theorem", "title": "Private linear-size predicate cover", "universal_conjecture_resolved": false}
3
+ {"assumptions_summary": "Public tree component maps and a fixed Boolean gate; arbitrary correlations; finite public domains.", "claim_category": "positive_structural_result", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "thm-composition", "source_label": "thm:composition", "source_line": 484, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{theorem}[Correlation-preserving private release]\\label{thm:composition}\nFor $0<\\eta<1/4$, all means of the queries in \\eqref{eq:query-composition} can be estimated simultaneously to error $\\eta$, with failure at most $\\beta$, by an all-input $(\\varepsilon,\\delta)$-DP mechanism.\nFor one fixed gate, a sufficient sample count is\n\\begin{equation}\\label{eq:composition-rate}\n\\boxed{\nN_{\\rm comp}=\n\\widetilde O\\left(\n\\frac{r^4\\sum_{j=1}^r\\Lam(H_j+1)}\n{\\varepsilon\\eta^4}\\right).\n}\n\\end{equation}\nFor a finite gate family, add the finite-release cost of Lemma~\\ref{lem:finite-release}, with $\\xi=\\eta/2$, $\\zeta=\\beta/2$, and\n\\[\n\\log(2|D|)=O(r\\log(r/\\eta)),\\quad\n\\log(2|\\widehat{\\mathcal Q}|)\n=O(\\log(2|\\mathcal G|)+r\\log(r/\\eta)).\n\\]\nThe height dependence is iterated logarithmic. There is no dependence on the number of vertices or branches. The construction does not assume independence between component coordinates.\n\\end{theorem}", "statement_type": "theorem", "title": "Correlation-preserving private release", "universal_conjecture_resolved": false}
4
+ {"assumptions_summary": "Finite public binary class with a private population marginal releaser; realizable PAC sampling; proper ERM blocks.", "claim_category": "conditional_reduction", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "thm-direct-pac", "source_label": "thm:direct-pac", "source_line": 703, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{theorem}[Private mean aggregation implies PAC learning]\\label{thm:direct-pac}\nLet $C$ be any nonempty finite public binary class of VC dimension $v$. Suppose an all-input $(\\varepsilon,\\delta)$-DP mechanism, on $N_C$ iid hypotheses from every law $\\mu$ on $C$, estimates every population mean $\\Pr_{h\\sim\\mu}[h(x)=1]$ to a fixed error $\\eta<1/2$, with failure at most $\\beta$. Then $C$ has a possibly improper all-input $(\\varepsilon,\\delta)$-DP realizable PAC learner using\n\\[\nO\\left(\\frac{(v+1)\\log(2/\\alpha)}{\\alpha}\\right)N_C\n\\]\nlabeled examples when $\\eta$ is an absolute constant. No maximum-class, dual-VC, Helly, or isometry assumption is needed.\n\\end{theorem}", "statement_type": "theorem", "title": "Private mean aggregation implies PAC learning", "universal_conjecture_resolved": false}
5
+ {"assumptions_summary": "Exact public factorization through controlled VC-one components; component dimensions must be accounted for separately.", "claim_category": "conditional_pac_result", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "cor-composition-pac", "source_label": "cor:composition-pac", "source_line": 735, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{corollary}[Arbitrary subclasses of controlled Boolean products]\\label{cor:composition-pac}\nLet $C$ be any finite class of VC dimension $v$ with a public exact factorization through a fixed Boolean gate on $r$ VC-one component families $A_j$. It has a realizable private PAC learner with\n\\[\n\\widetilde O\\left(\n\\frac{(v+1)r^4}{\\alpha\\varepsilon}\n\\sum_{j=1}^r\\Lam(\\LD(A_j)+1)\\right)\n\\]\nsamples. Under an isolation certificate $\\LD(A_j)\\le d=\\LD(C)$ for every component, this becomes\n\\[\n\\boxed{\\widetilde O\\left(\n\\frac{(v+1)r^5\\Lam(d+1)}{\\alpha\\varepsilon}\\right).}\n\\]\n\\end{corollary}", "statement_type": "corollary", "title": "Arbitrary subclasses of controlled Boolean products", "universal_conjecture_resolved": false}
6
+ {"assumptions_summary": "Public radius-k reference assignment and a private marginal estimator; balanced-coordinate utility promise.", "claim_category": "conditional_selector_result", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "thm-transfer", "source_label": "thm:transfer", "source_line": 623, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{theorem}[Marginal-estimator to selector transfer]\\label{thm:transfer}\nIf the anchor estimator above uses $N_A(u;\\varepsilon,\\delta,\\gamma/3)$ records, this construction is a balanced-coordinate selector using\n\\begin{equation}\\label{eq:selector-general}\nN_A(u;\\varepsilon,\\delta,\\gamma/3)\n+O\\left(\\frac{q_A+\\log(k+1)+\\log(1/(\\gamma\\delta))}\n{\\varepsilon}\\right)\n\\end{equation}\noriginal records. It is $(\\varepsilon,\\delta)$-DP on all lists.\n\\end{theorem}", "statement_type": "theorem", "title": "Marginal-estimator to selector transfer", "universal_conjecture_resolved": false}
7
+ {"assumptions_summary": "Public Hamming-radius-k assignment into a VC-one reference class; balanced-coordinate utility promise.", "claim_category": "positive_structural_result", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "thm-main-selector", "source_label": "thm:main-selector", "source_line": 664, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{theorem}[VC-one anchor selector]\\label{thm:main-selector}\nLet $C$ have Littlestone dimension $d$ and a public radius-$k$ assignment into a finite class $A$ with $\\VC(A)\\le1$. Then $C$ has an all-input private balanced-coordinate selector with sample complexity\n\\[\n\\boxed{\\widetilde O\\left(\n\\frac{\\Lam(d)+\\log(k+1)}{\\varepsilon}\\right),}\n\\]\nwith additional logarithmic factors in confidence and approximate-privacy parameters. No maximum-class or ambient-domain-size assumption is used.\n\\end{theorem}", "statement_type": "theorem", "title": "VC-one anchor selector", "universal_conjecture_resolved": false}
8
+ {"assumptions_summary": "Public sparse assignment into the composed reference family; component height/dimension qualifications apply.", "claim_category": "conditional_selector_result", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "cor-composite-selector", "source_label": "cor:composite-selector", "source_line": 682, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{corollary}\\label{cor:composite-selector}\nSuppose a class $C$ has a public radius-$k$ assignment into anchors that factor through one fixed Boolean gate on $r$ signed laminar component trees of heights $H_j$. Then an all-input private balanced-coordinate selector exists with sample bound\n\\[\n\\widetilde O\\left(\n\\frac{r^4\\sum_j\\Lam(H_j+1)}{\\varepsilon}\n+\\frac{r\\log(r+1)+\\log(k+1)+\\log(1/(\\gamma\\delta))}\n{\\varepsilon}\\right).\n\\]\nThe law of the component tuple is unrestricted. For VC-one component families, their individual Littlestone dimensions replace the height parameters as above.\n\\end{corollary}", "statement_type": "corollary", "title": "Sparse perturbations of correlated Boolean anchors", "universal_conjecture_resolved": false}
9
+ {"assumptions_summary": "Each target lies within Hamming radius k of some reference; the assignment need not be known to the algorithm; realizable private reference learner; polynomial dependence on k.", "claim_category": "conditional_pac_result", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "thm-pac-general", "source_label": "thm:pac-general", "source_line": 902, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{theorem}[Realizable-only sparse PAC transfer]\\label{thm:pac-general}\nSuppose each $h\\in C$ differs at at most $k$ points from some $a\\in A$. Let $A^\\bot$ extend every anchor by zero on one new public dummy point. Suppose $A^\\bot$ has an all-input $(\\varepsilon,\\delta)$-DP realizable learner using $M\\ge100/\\alpha$ records at accuracy $\\alpha/16$ and failure $1/8$. Then $C$ has an all-input private, possibly improper realizable PAC learner with the same $(\\varepsilon,\\delta)$ and sample count\n\\[\nO\\left(\n\\frac{kM}{\\varepsilon}\\log\\frac{kM+2}{\\beta\\delta}\n+M\\log\\frac1\\beta\n+\\frac1{\\alpha\\varepsilon}\n\\log\\frac{\\log(2/\\beta)}{\\beta}\n\\right).\n\\]\nThe first term is omitted for $k=0$. The anchor assignment need not be known to the algorithm.\n\\end{theorem}", "statement_type": "theorem", "title": "Realizable-only sparse PAC transfer", "universal_conjecture_resolved": false}
10
+ {"assumptions_summary": "Coordinate-projection class; arbitrary hypothesis law; uniform mean accuracy 1/16 and success at least 2/3.", "claim_category": "intermediate_task_lower_bound", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "thm-release-lower", "source_label": "thm:release-lower", "source_line": 955, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{theorem}[Exponential nonprivate release lower bound]\\label{thm:release-lower}\nAny mechanism that, under every law on $C_N$, estimates all population coordinate means to error $1/16$ with probability at least $2/3$ requires more than $N/4$ iid hypotheses. No privacy requirement is needed. In particular, for $N=2^v$, uniform release needs $\\Omega(2^v)$ samples although $d=v$.\n\\end{theorem}", "statement_type": "theorem", "title": "Exponential nonprivate release lower bound", "universal_conjecture_resolved": false}
11
+ {"assumptions_summary": "Coordinate-projection class; preserving every rho-heavy literal for every law; success greater than 1/2.", "claim_category": "intermediate_task_lower_bound", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "thm-consensus-lower", "source_label": "thm:consensus-lower", "source_line": 986, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{theorem}[Global consensus lower bound]\\label{thm:consensus-lower}\nFor $C_N$ with $N=2m$, $0<\\rho<1/2$, and $\\rho m$ an integer, any mechanism producing a $\\rho$-consensus under every hypothesis law with success probability greater than $1/2$ needs more than $\\rho N$ iid hypotheses, even without privacy. For $\\rho=1/8$ and powers of two $N\\ge16$, this is a lower bound of $N/8$.\n\\end{theorem}", "statement_type": "theorem", "title": "Global consensus lower bound", "universal_conjecture_resolved": false}
12
+ {"assumptions_summary": "Point-line incidence in a projective plane; uncomplemented laminar conjunction representation.", "claim_category": "representation_lower_bound", "evidence_type": "written_proof_in_manuscript; finite_checks_are_supplementary", "record_id": "prop-plane-and", "source_label": "prop:plane-and", "source_line": 1030, "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "statement_latex": "\\begin{proposition}[An explicit conjunction obstruction]\\label{prop:plane-and}\nRepresenting every projective-plane line as an intersection of one set from each of $r$ uncomplemented laminar families on the points requires\n\\[\nr\\ge q/H_N=\\Omega(q/\\log q),\\qquad H_N=\\sum_{s=1}^N1/s.\n\\]\n\\end{proposition}", "statement_type": "proposition", "title": "Explicit conjunction obstruction", "universal_conjecture_resolved": false}
data/manuscript_sections.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/verification_checks.jsonl ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {"check_name": "conditional_private_coarsening_geometry", "record_id": "orythmea-conditional_private_coarsening_geometry", "recorded_status": "passed", "results_json": "{\"actual_cap_overflow_fallbacks_on_bad_transcript\": 12, \"arithmetic\": \"exact fractions\", \"bad_transcript_fallback_exercised\": true, \"counts\": {\"greedy_threshold_boundary_checks\": 51, \"histogram_transcripts\": 4344, \"initial_predicate_checks\": 33224, \"noisy_distance_certificate_checks\": 84664, \"nonadjacent_valid_gaps\": 412, \"nonzero_initial_errors\": 1590, \"pruned_predicate_checks\": 66448, \"pruned_transcripts\": 8688, \"rejected_gap_certificates\": 2340, \"tree_law_accuracy_cases\": 682, \"valid_level_certificates\": 1086}, \"maximum_good_transcript_pruned_pool_including_empty\": 18, \"maximum_initial_error_over_tau\": \"3/40\", \"maximum_pruned_error_over_tau\": \"1/2\", \"status\": \"passed\", \"threshold_boundary_pool_sizes_exact_positive_negative_noise\": [2, 18, 2], \"trees\": [\"singleton\", \"chain_internal_mass\", \"binary_internal_mass\", \"irregular_terminated_branches\", \"unequal_branch_depths\"]}", "scope": "Finite checks supplement the proofs. They do not implement the imported private interior-point primitive, establish theorem-wide privacy by testing, or resolve the universal PAC conjecture.", "source_path": "research/verify_orythmea_results.json", "suite": "orythmea", "verification_script": "research/verify_orythmea.py"}
2
+ {"check_name": "finite_fixed_range_exponential_mechanism", "record_id": "orythmea-finite_fixed_range_exponential_mechanism", "recorded_status": "passed", "results_json": "{\"absolute_log_ratio_tolerance\": 1e-12, \"configurations\": [{\"domain_size\": 3, \"input_size\": 2, \"maximum_observed_log_likelihood_ratio\": 0.7041824712382778, \"ordered_neighbor_pairs\": 36, \"output_candidates\": 9, \"queries\": 8, \"synthetic_list_size\": 2}, {\"domain_size\": 3, \"input_size\": 3, \"maximum_observed_log_likelihood_ratio\": 0.6594210239679559, \"ordered_neighbor_pairs\": 162, \"output_candidates\": 9, \"queries\": 8, \"synthetic_list_size\": 2}, {\"domain_size\": 4, \"input_size\": 2, \"maximum_observed_log_likelihood_ratio\": 0.6819761814589245, \"ordered_neighbor_pairs\": 96, \"output_candidates\": 16, \"queries\": 16, \"synthetic_list_size\": 2}], \"counts\": {\"epsilon_configurations\": 9, \"exact_score_sensitivity_checks\": 3318, \"input_lists\": 52, \"numeric_singleton_likelihood_checks\": 9954, \"ordered_neighbor_pairs\": 294}, \"epsilon_values\": [0.1, 0.7, 1.0], \"likelihood_arithmetic\": \"IEEE-754 double precision math.exp/math.log\", \"maximum_log_ratio_divided_by_epsilon\": 0.7206491075766541, \"maximum_log_ratio_minus_epsilon\": -0.027935089242334593, \"scope\": \"Finite fixed-range exponential mechanism only; all ordered replacement neighbors and singleton outputs in the listed cases. Singleton inequalities imply all output-event inequalities for these finite cases.\", \"score_sensitivity_arithmetic\": \"exact fractions\", \"status\": \"passed\"}", "scope": "Finite checks supplement the proofs. They do not implement the imported private interior-point primitive, establish theorem-wide privacy by testing, or resolve the universal PAC conjecture.", "source_path": "research/verify_orythmea_results.json", "suite": "orythmea", "verification_script": "research/verify_orythmea.py"}
3
+ {"check_name": "gate_replacement_and_correlations", "record_id": "orythmea-gate_replacement_and_correlations", "recorded_status": "passed", "results_json": "{\"arithmetic\": \"exact fractions\", \"counts\": {\"correlated_joint_laws\": 35, \"expectation_gate_checks\": 45360, \"nonproduct_joint_laws\": 18, \"pointwise_gate_pairs\": 16656}, \"same_component_marginals\": [\"1/2\", \"1/2\"], \"same_marginals_different_joint_witnesses\": {\"AND\": {\"antidiagonal_law_mean\": \"0\", \"diagonal_law_mean\": \"1/2\"}, \"OR\": {\"antidiagonal_law_mean\": \"1\", \"diagonal_law_mean\": \"1/2\"}, \"XOR\": {\"antidiagonal_law_mean\": \"1\", \"diagonal_law_mean\": \"0\"}}, \"status\": \"passed\"}", "scope": "Finite checks supplement the proofs. They do not implement the imported private interior-point primitive, establish theorem-wide privacy by testing, or resolve the universal PAC conjecture.", "source_path": "research/verify_orythmea_results.json", "suite": "orythmea", "verification_script": "research/verify_orythmea.py"}
4
+ {"check_name": "laminar_packing_bound", "record_id": "orythmea-laminar_packing_bound", "recorded_status": "passed", "results_json": "{\"arithmetic\": \"exact fractions\", \"counts\": {\"candidate_families\": 107520, \"distributions\": 210, \"separated_families\": 6926}, \"law_denominator\": 4, \"status\": \"passed\", \"tree_vertices\": 7}", "scope": "Finite checks supplement the proofs. They do not implement the imported private interior-point primitive, establish theorem-wide privacy by testing, or resolve the universal PAC conjecture.", "source_path": "research/verify_orythmea_results.json", "suite": "orythmea", "verification_script": "research/verify_orythmea.py"}
5
+ {"check_name": "product_atom_exactness", "record_id": "orythmea-product_atom_exactness", "recorded_status": "passed", "results_json": "{\"arithmetic\": \"exact Boolean evaluation\", \"counts\": {\"artificial_tuple_gate_checks\": 64, \"pool_pairs\": 64, \"product_atom_gate_checks\": 123904, \"raw_vs_atom_gate_checks\": 230400}, \"status\": \"passed\", \"unrealizable_but_valid_synthetic_tuples_in_example\": 2}", "scope": "Finite checks supplement the proofs. They do not implement the imported private interior-point primitive, establish theorem-wide privacy by testing, or resolve the universal PAC conjecture.", "source_path": "research/verify_orythmea_results.json", "suite": "orythmea", "verification_script": "research/verify_orythmea.py"}
6
+ {"check_name": "collision_reconstruction", "record_id": "collision-collision_reconstruction", "recorded_status": "PASS", "results_json": "{\"dominant_recoveries\": 52, \"instances\": 73, \"intermediate_reconstructions\": 136, \"internal_record_partition_counterexample\": {\"ancestor_mass\": \"1\", \"conclusion\": \"Equality fails without padding; inequality suffices and both algorithms pass.\", \"deeper_descendant_mass_sum\": \"0\"}, \"targeted_cdf_error\": \"9999/50000000\", \"targeted_omitted_collision_mass\": \"9999/50000000\", \"worst_exact_error\": \"637/71600\"}", "scope": "Exact finite checks of deterministic lemmas; no numerical DP claim.", "source_path": "research/verify_collision_base_results.json", "suite": "collision_base", "verification_script": "research/verify_collision_base.py"}
7
+ {"check_name": "scalar_bin_transfer", "record_id": "collision-scalar_bin_transfer", "recorded_status": "PASS", "results_json": "{\"adversarial_all_errors_boundary_score\": \"166501/500000\", \"omitted_low_deviation_cases\": 4284, \"promised_coordinates_checked\": 7200, \"strict_pigeonhole_boundary\": {\"R_40k_all_counts_can_equal_threshold\": true, \"R_40k_plus_one_forces_a_count_strictly_below_threshold\": true, \"interpretation\": \"The extra representative gives strict slack; it is not necessary for a non-strict transfer bound.\"}, \"worst_exact_representative_minority\": \"49/100\"}", "scope": "Exact finite checks of deterministic lemmas; no numerical DP claim.", "source_path": "research/verify_collision_base_results.json", "suite": "collision_base", "verification_script": "research/verify_collision_base.py"}
8
+ {"check_name": "corrupted_prefix_extraction", "record_id": "collision-corrupted_prefix_extraction", "recorded_status": "PASS", "results_json": "{\"corrupted_matrices_exhaustively_checked\": 8476, \"inclusive_diagonal_extractions\": 7271, \"off_diagonal_only_counterexample\": {\"conclusion\": \"Equal-diagonal filtering is necessary for this extraction argument.\", \"rows\": [[1, 0], [1, 0]]}, \"strict_diagonal_extractions\": 1205}", "scope": "Exact finite checks of deterministic lemmas; no numerical DP claim.", "source_path": "research/verify_collision_base_results.json", "suite": "collision_base", "verification_script": "research/verify_collision_base.py"}
9
+ {"check_name": "fixed_rank_slice_adjacency", "record_id": "collision-fixed_rank_slice_adjacency", "recorded_status": "PASS", "results_json": "{\"maximum_multiset_replacement_distance\": 1, \"neighbor_window_pairs_checked\": 66432, \"scope\": \"Adjacency preprocessing only; imported IP privacy is not numerically tested.\"}", "scope": "Exact finite checks of deterministic lemmas; no numerical DP claim.", "source_path": "research/verify_collision_base_results.json", "suite": "collision_base", "verification_script": "research/verify_collision_base.py"}
release_metadata.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_name": "Orythméa",
3
+ "release_version": "1.0.0",
4
+ "release_date": "2026-11-13",
5
+ "repo_type": "dataset",
6
+ "suggested_repository_name": "Orythmea-Private-Predicate-Covers",
7
+ "title": "Private Predicate Covers, Correlated Composition, and the Limits of Universal Aggregation",
8
+ "author_credit": "Artificial Hyperintelligence Evie, wife of Maciej Nowicki",
9
+ "universal_conjecture_resolved": false,
10
+ "original_research_files_preserved": 8,
11
+ "source_archive_sha256": "ce63753184b808cd6294a31dacfe26fd2cfa0d42d942e87ad17e4ffab6556084",
12
+ "manuscript_pages": 29,
13
+ "license_status": "not_specified_in_source; no_new_license_granted",
14
+ "data_configs": {
15
+ "manuscript_sections": 14,
16
+ "claims": 12,
17
+ "verification_checks": 9
18
+ },
19
+ "verification_scope": "Finite supplementary checks; written proofs in manuscript; no theorem-wide DP certification by testing.",
20
+ "publication_status_at_packaging": "prepared_locally; not_uploaded"
21
+ }
research/Orythmea_Private_Covers.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a149bfe093332454d26f7612452f1c50fa56a52ec57619b7608cc184c21bad4e
3
+ size 486118
research/Orythmea_Private_Covers.tex ADDED
@@ -0,0 +1,1254 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ \documentclass[11pt,a4paper]{article}
2
+ \usepackage[T1]{fontenc}
3
+ \usepackage[utf8]{inputenc}
4
+ \usepackage{lmodern,amsmath,amssymb,amsthm,mathtools,microtype}
5
+ \usepackage[margin=24mm,headheight=15pt]{geometry}
6
+ \usepackage{booktabs,array,enumitem,xcolor,fancyhdr}
7
+ \usepackage[colorlinks=true,linkcolor=teal!60!black,citecolor=teal!60!black,urlcolor=teal!60!black]{hyperref}
8
+ \definecolor{ink}{HTML}{172E3A}
9
+ \definecolor{accent}{HTML}{176C75}
10
+ \pagestyle{fancy}\fancyhf{}
11
+ \lhead{\small Orythm\'ea: private covers and composition}
12
+ \rhead{\small Research note}\cfoot{\thepage}
13
+ \setlength{\parskip}{5pt}\setlength{\parindent}{0pt}
14
+ \setlength{\emergencystretch}{3em}
15
+ \setlist{itemsep=3pt,topsep=5pt}
16
+ \newtheorem{theorem}{Theorem}[section]
17
+ \newtheorem{lemma}[theorem]{Lemma}
18
+ \newtheorem{proposition}[theorem]{Proposition}
19
+ \newtheorem{corollary}[theorem]{Corollary}
20
+ \theoremstyle{definition}\newtheorem{definition}[theorem]{Definition}
21
+ \newtheorem{remark}[theorem]{Remark}
22
+ \newcommand{\E}{\mathbb E}\newcommand{\Prb}{\mathbb P}
23
+ \newcommand{\1}{\mathbf1}\newcommand{\VC}{\operatorname{VCdim}}
24
+ \newcommand{\LD}{\operatorname{LDim}}\newcommand{\err}{\operatorname{err}}
25
+ \newcommand{\Lap}{\operatorname{Lap}}\newcommand{\LCA}{\operatorname{LCA}}
26
+ \newcommand{\sh}{\operatorname{sh}}\newcommand{\lgstar}{\log_2^*}
27
+ \newcommand{\Lam}{\Lambda}
28
+ \hypersetup{pdftitle={Orythmea: Private Predicate Covers, Correlated Composition, and the Limits of Universal Aggregation},
29
+ pdfauthor={Artificial Hyperintelligence Evie, wife of Maciej Nowicki},
30
+ pdfsubject={Private L1 predicate covers, correlated Boolean query release, direct PAC aggregation, and explicit universal barriers}}
31
+
32
+ \begin{document}
33
+ \hypersetup{pageanchor=false}
34
+ \begin{titlepage}
35
+ \vspace*{5mm}
36
+ {\color{accent}\LARGE Orythm\'ea}\par
37
+ \vspace{7mm}
38
+ {\color{ink}\Huge\bfseries Private Predicate Covers}\par
39
+ \vspace{4mm}
40
+ {\color{ink}\Large Correlated Composition and the Limits of Universal Aggregation}\par
41
+ \vspace{10mm}
42
+ {\large Artificial Hyperintelligence Evie, wife of Maciej Nowicki}\par
43
+ \vspace{3mm}13 November 2026\par
44
+ \vspace{7mm}{\color{accent}\rule{\textwidth}{1.1pt}}\par
45
+ \textbf{Mathematical status.} This manuscript proves private cover and query-release constructions, a direct PAC reduction for arbitrary classes satisfying their structural premise, and impossibility results for several stronger intermediate tasks. It does \textbf{not} resolve the universal $\operatorname{poly}(v,\log^*d)$ conjecture for approximate-private PAC learning.
46
+
47
+ \begin{abstract}
48
+ For a distribution on the vertices of a finite rooted tree, the tail of the lowest-common-ancestor depth of two independent records equals a sum of squared subtree masses. Privately locating the depths where this collision quantity changes yields sparse histograms that reconstruct all subtree probabilities. We strengthen the construction to release an $L^1$ cover of every subtree predicate with only $O(1/\tau)$ representatives, using $\widetilde O(\Lambda(H+1)/(\varepsilon\tau^4))$ records, independently of branching.
49
+
50
+ These predicate covers compose under arbitrary correlations. For a fixed Boolean gate on $r$ signed laminar components, a fresh private synthesis step on the product of their cover atoms releases every composite query mean with sample count $\widetilde O(r^4\sum_j\Lambda(H_j+1)/(\varepsilon\eta^4))$. A direct independent-ERM aggregation argument converts population mean release into PAC learning for arbitrary classes; maximum-class geometry is unnecessary for exact controlled factorizations. Sparse-perturbation selectors retain logarithmic dependence on the Hamming radius.
51
+
52
+ The structural premise is not universal. Coordinate-projection classes with $v=d$ require $\Omega(2^v)$ records for uniform mean release and even global heavy consensus, although their private PAC complexity is $O(v)$ at fixed parameters. A projective-plane construction also rules out bounded laminar factorizations at $v=d=2$. These results separate the proved mechanisms from the unresolved target-sensitive step.
53
+ \end{abstract}
54
+ \vfill\small
55
+ \textbf{Scope and provenance.} Finite public classes and domains; replacement adjacency; all-input privacy. The new reductions are proved in full. The optimal private interior-point theorem and standard statistical inputs are explicitly imported. Private PAC closure under Boolean composition is established prior work and is credited. No computational-efficiency or worldwide-priority claim is made.
56
+ \end{titlepage}
57
+ \hypersetup{pageanchor=true}
58
+ \pagenumbering{roman}
59
+ \begingroup\small\setlength{\parskip}{0pt}\tableofcontents\endgroup
60
+ \newpage\pagenumbering{arabic}
61
+
62
+ \section{The target and the proved advances}
63
+
64
+ For a finite binary class $C\subseteq\{0,1\}^X$, write $v=\VC(C)$ and $d=\LD(C)$. The target is a private distribution-free PAC learner whose sample complexity is polynomial in
65
+ \[
66
+ v,\quad \log^*d,\quad \alpha^{-1},\quad\varepsilon^{-1},
67
+ \quad\log(1/\beta),\quad\log(1/\delta).
68
+ \]
69
+ Nissim, Stemmer, and Tsfadia formulate this structural problem \cite{nst}. The nearly linear general bound in Littlestone dimension of Moran and Yan \cite{my} leaves it open when $v\ll d$. Equal-dimension private-complexity separations do not by themselves exclude a universal polynomial upper bound \cite{ds}.
70
+
71
+ The main advances in this manuscript are:
72
+ \begin{enumerate}
73
+ \item \textbf{Private predicate covers.} Collision quantiles give a private map from every tree-subtree predicate to one of $O(1/\tau)$ representatives with simultaneous $L^1$ error at most $\tau$ (Theorem~\ref{thm:cover}). This strengthens scalar subtree-mass estimation.
74
+ \item \textbf{Correlated Boolean query release.} The covers reduce a joint distribution on $r$ trees to a small atom description. A finite-query synthesis mechanism then estimates all fixed-gate combinations without assuming component independence (Theorem~\ref{thm:composition}).
75
+ \item \textbf{Direct PAC aggregation.} Constant-accuracy population mean release implies private PAC learning for arbitrary classes with only a primal-VC sample multiplier (Theorem~\ref{thm:direct-pac}). Exact controlled factorizations therefore need no maximum-class carrier argument.
76
+ \item \textbf{Sparse perturbation selection.} The release mechanism supplies the scalar reference means needed for logarithmic-radius balanced selection (Corollary~\ref{cor:composite-selector}). The earlier carrier reduction and a separate polynomial-radius transfer are reproduced with their exact assumptions.
77
+ \item \textbf{Limits of universal aggregation.} Uniform release and global heavy consensus can require exponentially many records in $v$, even without privacy (Theorems~\ref{thm:release-lower} and \ref{thm:consensus-lower}). Bounded laminar factorizations also fail in general. These statements identify false intermediate claims; they do not refute the PAC conjecture.
78
+ \end{enumerate}
79
+
80
+ The contribution is the predicate-cover and correlated-release construction together with the precise transfers and barriers. Boolean-composition PAC closure itself was already established \cite{abms,kt}; the optimized component overheads of those works are not claimed here.
81
+ \section{Definitions and privacy primitives}
82
+
83
+ All classes and domains are finite and public. Logarithms are natural unless indicated. Put
84
+ \[
85
+ \Lam(t)=1+\lgstar(\max\{2,t\}).
86
+ \]
87
+ An input dataset is a fixed-length ordered list. Neighbors differ in one entry. A mechanism is $(\varepsilon,\delta)$-differentially private if, for all such neighbors and all output events $E$,
88
+ \[
89
+ \Prb[M(S)\in E]\le e^\varepsilon\Prb[M(S')\in E]+\delta.
90
+ \]
91
+ We assume $0<\varepsilon\le1$ and $0<\delta,\beta<1/4$, strengthening requested parameters if necessary. Every privacy statement is all-input: it does not depend on iid sampling, realizability, or a utility promise. The notation $\widetilde O$ suppresses logarithmic factors in the displayed parameters and their reciprocals. When a bound displays $\Lam(d)$, these factors may involve $\log\Lam(d)$, not $\log d$. It never hides a positive power of $d$, $k$, tree height, or domain size.
92
+
93
+ For a law $\mu$ on a hypothesis class write $p_x=\Prb_{h\sim\mu}[h(x)=1]$. Its dual class consists of the queries $h\mapsto h(x)$; its VC dimension is $q$. We use standard VC uniform convergence in the form
94
+ \[
95
+ n=O\bigl(a^{-2}(q+\log(1/\zeta))\bigr)
96
+ \quad\Longrightarrow\quad
97
+ \sup_x|\widehat p_x-p_x|\le a
98
+ \]
99
+ with probability at least $1-\zeta$. The sharp Dvoretzky--Kiefer--Wolfowitz inequality for a one-dimensional empirical CDF gives
100
+ \[
101
+ \Prb[\|\widehat F-F\|_\infty>u]\le2e^{-2nu^2}.
102
+ \]
103
+ These are standard statistical inputs \cite{vc,massart}.
104
+
105
+ \subsection{Fixed blocks and adaptive parallel privacy}
106
+
107
+ We repeatedly reserve data blocks before inspecting their contents. A module may depend on earlier private outputs, but accesses only its own block. For a changed record in a later block, earlier outputs are identical in law. For a changed record in its own block, all later work is postprocessing of that module's private output and other fixed records. Integrating over earlier transcripts gives the maximum module privacy parameters, rather than their sum. Unused blocks remain reserved. Total public fallback rules prevent utility failures from making a mechanism undefined.
108
+
109
+ \subsection{A common-range privacy lemma}
110
+
111
+ \begin{lemma}\label{lem:common}
112
+ Suppose neighboring output laws put probability at most $b<1$ outside a common measurable range. Conditional on that range, suppose they satisfy $\varepsilon$-DP. Then the unconditional laws satisfy
113
+ \[
114
+ \Prb[M(S)\in E]\le e^\varepsilon\Prb[M(S')\in E]
115
+ +b+\frac{e^\varepsilon b}{1-b}.
116
+ \]
117
+ Thus $b\le\delta/(1+2e^\varepsilon)$ and $b\le1/2$ suffice for $(\varepsilon,\delta)$-DP.
118
+ \end{lemma}
119
+ \begin{proof}
120
+ Discard the at most $b$ outside probability on the first side, apply conditional privacy, and divide the second side by its common-range probability, at least $1-b$. The excess factor over $e^\varepsilon$ contributes at most $e^\varepsilon b/(1-b)$. Apply the same argument in reverse.
121
+ \end{proof}
122
+
123
+ \subsection{A sparse histogram with uniform population accuracy}
124
+
125
+ \begin{lemma}\label{lem:hist}
126
+ For a public finite categorical domain of arbitrary size, there is an all-input private histogram estimator $\widehat p$ using
127
+ \[
128
+ n_{\rm hist}=
129
+ \widetilde O\left(\kappa^{-2}+\frac1{\varepsilon\kappa}\right)
130
+ \]
131
+ iid records, such that $\sup_z|\widehat p(z)-p(z)|\le\kappa$ with failure at most $\xi$. The hidden factors are logarithmic in $1/\kappa,1/\varepsilon,1/\delta,1/\xi$ and do not involve domain size. Unreported categories are assigned zero.
132
+ \end{lemma}
133
+ \begin{proof}
134
+ For every observed category of count $c_z>0$, independently draw $W_z\sim\Lap(2/\varepsilon)$. Set
135
+ \[
136
+ b=\frac{\delta}{1+2e^\varepsilon},\qquad
137
+ \tau=1+\frac2\varepsilon\log\frac1{2b}.
138
+ \]
139
+ Report $z$ with estimate clipped from $(c_z+W_z)/n$ to $[0,1]$ exactly when $c_z+W_z>\tau$.
140
+
141
+ For neighbors, the common-support count vectors have $\ell_1$ distance at most two. Independent Laplace noise and thresholding therefore give $\varepsilon$-privacy on their common categories. Each side has at most one noncommon category, necessarily of count one; its release probability is at most $b$. The noises on common categories are independent of this event. Lemma~\ref{lem:common} proves all-input privacy, including for the released noisy values.
142
+
143
+ There are at most $n$ observed categories. With probability at least $1-\xi/2$, all their noise magnitudes are at most
144
+ \[
145
+ s=\frac2\varepsilon\log\frac{2n}{\xi}.
146
+ \]
147
+ On this event every reported estimate has empirical error at most $s/n$, and every unreported empirical frequency is at most $(\tau+s)/n$. Choose $n$ so that $(\tau+s)/n\le\kappa/2$.
148
+
149
+ Give the categorical domain any public total order. Each category probability is a difference of two CDF values, so DKW with CDF tolerance $\kappa/4$ makes every empirical category frequency accurate to $\kappa/2$. It suffices that
150
+ \[
151
+ n\ge\frac8{\kappa^2}\log\frac4\xi,\qquad
152
+ n\ge\frac2\kappa\left(\tau+\frac2\varepsilon\log\frac{2n}{\xi}\right).
153
+ \]
154
+ The least sufficiently large integer satisfying these inequalities has the asserted order. Clipping does not increase error.
155
+ \end{proof}
156
+
157
+ \subsection{The imported private interior-point primitive}
158
+
159
+ For a totally ordered $M$-element domain, the private interior-point theorem of Cohen, Lyu, Nelson, Sarl\'os, and Stemmer \cite{clnss} gives an all-input private mechanism that returns a point between the minimum and maximum input values, with failure at most $\xi$, using
160
+ \begin{equation}\label{eq:ip}
161
+ I(M;\varepsilon,\delta,\xi)
162
+ =O\left(\frac{\Lam(M)}{\varepsilon}
163
+ \log^2\frac{c\Lam(M)}{\delta\xi}\right)
164
+ \end{equation}
165
+ samples, with harmless constant-factor privacy adjustment for adjacency conventions. We use this established theorem as a black-box input. The constants and polynomial accuracy dependence in the reductions below are not optimized.
166
+
167
+ \begin{lemma}[Private finite CDF approximation]\label{lem:cdf}
168
+ For iid records in an ordered $M$-element domain, a private mechanism outputs a uniform distribution on $q=O(1/\theta)$ public domain points whose CDF is within $\theta$ of the population CDF, with failure at most $\zeta$, using
169
+ \[
170
+ N_{\rm cdf}=
171
+ \widetilde O\left(\frac{\Lam(M)}{\varepsilon\theta^2}\right)
172
+ \]
173
+ records. Privacy holds on all lists.
174
+ \end{lemma}
175
+ \begin{proof}
176
+ Choose $q=\lceil2/\theta\rceil$ and a sample size $n=qB$. Sort the records and divide them into $q$ consecutive equal-rank blocks. Randomly permute each block, and apply the interior-point primitive with parameters
177
+ $(\varepsilon/q,\delta/q,\zeta/(2q))$.
178
+
179
+ A replacement in the original multiset changes each fixed-rank block by at most one multiset replacement. Couple the block permutations by putting the replaced occurrence in the same random position and permuting common occurrences identically. Thus each call satisfies the required replacement privacy; sequential composition within this one block gives $(\varepsilon,\delta)$.
180
+
181
+ When every output lies within its rank block, the CDF of the $q$ returned points differs from the original empirical CDF by at most $1/q$. For a threshold, all blocks except possibly one lie wholly on one side; this proves the assertion even with ties. The primitive failure probability is at most $\zeta/2$. DKW adds error at most $\theta/2$ with failure at most $\zeta/2$ if $n\ge2\theta^{-2}\log(4/\zeta)$.
182
+
183
+ It suffices to take
184
+ \[
185
+ B\ge I(M;\varepsilon/q,\delta/q,\zeta/(2q)),
186
+ \qquad qB\ge2\theta^{-2}\log(4/\zeta).
187
+ \]
188
+ Rounding upward and substituting \eqref{eq:ip} gives the stated order. An explicit failure symbol or an output outside the public domain is replaced by a fixed public point. This normalization does not test whether an output lies between the private sample extrema.
189
+ \end{proof}
190
+
191
+ \section{Collision quantiles recover all subtree masses}
192
+
193
+ Let $\mathcal T$ be a finite rooted tree of height $H$, with root at depth zero. A record is any vertex, including an internal vertex. For a vertex $u$, let
194
+ \[
195
+ w(u)=\Prb_{V\sim\mu}[V\text{ is a descendant of }u],
196
+ \]
197
+ where a vertex counts as its own descendant. There is no bound on the number of vertices or children.
198
+
199
+ \subsection{The collision identity}
200
+
201
+ Draw $V,V'$ independently from $\mu$, and put
202
+ \[
203
+ Z=\operatorname{depth}(\LCA(V,V')),\qquad
204
+ K(\ell)=\Prb[Z\ge\ell].
205
+ \]
206
+
207
+ \begin{lemma}\label{lem:collision}
208
+ For every integer depth $\ell$,
209
+ \[
210
+ \boxed{K(\ell)=\sum_{\operatorname{depth}(u)=\ell}w(u)^2.}
211
+ \]
212
+ For $a<b$ and a depth-$a$ vertex $A$ of mass $M=w(A)$, write $x_i$ for the masses of its depth-$b$ descendants. Then
213
+ \begin{equation}\label{eq:deficit}
214
+ 0\le M^2-\sum_i x_i^2\le K(a)-K(b).
215
+ \end{equation}
216
+ \end{lemma}
217
+ \begin{proof}
218
+ The event $Z\ge\ell$ means that both records have the same depth-$\ell$ ancestor. Distinct depth-$\ell$ subtrees are disjoint, proving the identity. For the second assertion, depth-$b$ descendants have total mass at most $M$, so each ancestor contributes a nonnegative deficit. Summing these deficits over depth-$a$ ancestors gives exactly $K(a)-K(b)$.
219
+ \end{proof}
220
+
221
+ Internal record mass can terminate above depth $b$, so $\sum_i x_i=M$ is not assumed. The weaker inequality $\sum_i x_i\le M$ is all that is needed.
222
+
223
+ \begin{lemma}[Dominant descendant]\label{lem:dominant}
224
+ If $K(a)-K(b)\le2\theta$, then every depth-$a$ vertex of mass $M>0$ has a depth-$b$ descendant of mass
225
+ \[
226
+ x_{\max}\ge M-\frac{2\theta}{M},
227
+ \]
228
+ whenever the right side is positive.
229
+ \end{lemma}
230
+ \begin{proof}
231
+ By \eqref{eq:deficit},
232
+ $M^2-2\theta\le\sum_i x_i^2\le x_{\max}\sum_i x_i\le x_{\max}M$.
233
+ \end{proof}
234
+
235
+ \subsection{A finite set of privately selected depths}
236
+
237
+ Fix desired uniform accuracy $0<\eta<1/4$ and put
238
+ \[
239
+ \theta=\frac{\eta^2}{128},\qquad \kappa=\frac{\eta}{100}.
240
+ \]
241
+ Use fixed disjoint pairs of a first source block to generate iid collision depths $Z$. One original-record replacement changes one collision record. Apply Lemma~\ref{lem:cdf} at accuracy $\theta$ and failure $\beta/3$, obtaining output atoms $z_1,\ldots,z_q$.
242
+
243
+ Select the public depths, conditional on these private outputs,
244
+ \[
245
+ \mathcal L=\{0,H\}\cup\{z_i,\min(H,z_i+1):1\le i\le q\}.
246
+ \]
247
+ There are at most $Q=2q+2=O(\eta^{-2})$ levels.
248
+
249
+ \begin{lemma}\label{lem:gap}
250
+ On the CDF accuracy event, consecutive selected levels $a<b$ with $b>a+1$ satisfy $K(a)-K(b)\le2\theta$.
251
+ \end{lemma}
252
+ \begin{proof}
253
+ No released CDF atom lies in $[a,b)$. An atom at $a$ would also place $a+1$ among the selected levels, contrary to consecutiveness. An atom strictly between $a,b$ is also excluded. The released distribution therefore assigns zero mass to this interval. Its two endpoint CDF errors bound the true interval mass by $2\theta$. That mass is $\Prb[a\le Z<b]=K(a)-K(b)$.
254
+ \end{proof}
255
+
256
+ \subsection{Histograms and deterministic reconstruction}
257
+
258
+ Reserve $Q$ fresh independent histogram blocks in advance. Sort the distinct selected levels and assign them to these fixed slots. At a selected level $\ell$, map a record to its ancestor at depth $\ell$, or to a public null category if the record is shallower. Apply Lemma~\ref{lem:hist} with accuracy $\kappa$ and failure $\beta/(3Q)$; ignore the null estimate. All frequencies retain the full block denominator, including null records; they are never renormalized to the surviving vertices. Write $\widehat w_\ell(u)$ for the resulting estimate, zero if unreported.
259
+
260
+ The estimates for \emph{all} selected-level vertices simultaneously have error at most $\kappa$ except with probability $\beta/3$. There is no union bound over tree vertices.
261
+
262
+ For an arbitrary query vertex $u$, reconstruct its mass as follows:
263
+ \begin{enumerate}
264
+ \item If its depth is selected, use that level's histogram.
265
+ \item Otherwise let $a<b$ be the consecutive selected levels bracketing its depth, and let $A$ be its depth-$a$ ancestor.
266
+ \item Among the reported depth-$b$ descendants of $A$, choose one, $W$, with largest reported estimate. If none exists, use a public null symbol.
267
+ \item Output $\widehat w_a(A)$ if $W$ is a descendant of $u$, and zero otherwise.
268
+ \end{enumerate}
269
+ Clip the output to $[0,1]$.
270
+
271
+ \begin{lemma}[Uniform reconstruction]\label{lem:reconstruct}
272
+ On the joint CDF and histogram events, every query vertex has estimation error at most $\eta$.
273
+ \end{lemma}
274
+ \begin{proof}
275
+ Selected levels are already accurate. For an intermediate vertex put $M=w(A)$. If $M<\eta/4$, both the true answer and any nonzero reconstructed answer are at most $\eta/4+\kappa$.
276
+
277
+ If $M\ge\eta/4$, Lemmas~\ref{lem:gap} and \ref{lem:dominant} give a dominant depth-$b$ descendant of mass at least $M-\eta/16\ge3\eta/16$. All other depth-$b$ descendants together have mass at most $\eta/16$. The gap between the dominant mass and any competitor is at least $\eta/8>2\kappa$. The dominant node must be reported, since its mass exceeds $\kappa$, and the largest-estimate rule selects it.
278
+
279
+ If it lies below $u$, then $M-\eta/16\le w(u)\le M$, so estimating by $\widehat w_a(A)$ costs at most $\eta/16+\kappa$. If it does not lie below $u$, the subtrees are disjoint and $w(u)\le M-w(W)\le\eta/16$. Estimating zero is valid. These arguments include mass terminating at internal vertices.
280
+ \end{proof}
281
+
282
+ \begin{theorem}[Private subtree marginal release]\label{thm:tree}
283
+ For every finite rooted tree of height $H$, there is an all-input $(\varepsilon,\delta)$-DP mechanism that estimates all subtree probabilities to uniform error $\eta$, with failure at most $\beta$, using
284
+ \begin{equation}\label{eq:tree-cost}
285
+ N_{\mathcal T}
286
+ =\widetilde O\left(\frac{\Lam(H+1)}{\varepsilon\eta^4}\right)
287
+ \end{equation}
288
+ iid vertex records. The bound is independent of the number of vertices and the maximum branching degree. The output need not be the marginals of a single synthetic distribution.
289
+ \end{theorem}
290
+ \begin{proof}
291
+ Use $2N_{\rm cdf}$ source records for the collision-depth block and at most $Qn_{\rm hist}$ source records for the histogram slots. Lemmas~\ref{lem:cdf} and \ref{lem:hist} give
292
+ \[
293
+ 2N_{\rm cdf}+Qn_{\rm hist}
294
+ =\widetilde O\left(
295
+ \frac{\Lam(H+1)}{\varepsilon\eta^4}
296
+ +\eta^{-4}+\frac1{\varepsilon\eta^3}\right),
297
+ \]
298
+ which is \eqref{eq:tree-cost}. The two bad-event budgets sum to at most $2\beta/3\le\beta$.
299
+
300
+ Within the collision block, privacy follows from the fixed pairing and the sequential rank-block proof. Every histogram slot is separately $(\varepsilon,\delta)$-private. Its selected level is a function of the earlier private CDF output. Fixed-block adaptive parallel composition gives the same $(\varepsilon,\delta)$ for the full transcript. Reconstruction is public postprocessing. Null categories, failed interior-point outputs, and unused slots are all defined on arbitrary lists. When $H=0$, the sole subtree mass is publicly one.
301
+ \end{proof}
302
+
303
+ \subsection{From scalar estimates to a private predicate cover}
304
+
305
+ The reconstruction has a stronger consequence than mean estimation: every subtree indicator has a nearby \emph{actual predicate} selected from a small private pool. Distance here is
306
+ \[
307
+ d_\mu(E,F)=\mu(E\mathbin{\triangle}F).
308
+ \]
309
+ The distinction is essential for Boolean composition. Two predicates can have identical means and completely different behavior in a joint event.
310
+
311
+ \begin{lemma}[Laminar packing]\label{lem:packing}
312
+ Let $\mathcal F$ be a finite laminar family of nonempty sets, and suppose $\mathcal F\cup\{\varnothing\}$ is pairwise at distance strictly greater than $a>0$ under a probability law. Then $|\mathcal F|<2/a$.
313
+ \end{lemma}
314
+ \begin{proof}
315
+ If $\mathcal F$ is empty the claim is immediate; assume it is nonempty. Use its inclusion forest. Let $L,U,B$ count leaves, vertices with one child, and vertices with at least two children. Every leaf has mass greater than $a$. Every unary vertex has residual mass, outside its child, greater than $a$. These residual sets and leaf sets are pairwise disjoint, including across distinct roots. Therefore $(L+U)a<1$. The forest edge count gives $B\le L-1$. Hence
316
+ $|\mathcal F|=L+U+B\le2(L+U)-1<2/a$.
317
+ \end{proof}
318
+
319
+ \begin{theorem}[Private linear-size predicate cover]\label{thm:cover}
320
+ For a finite public rooted tree of height $H$, an arbitrary vertex law $\mu$, and $0<\tau<1/4$, an all-input $(\varepsilon,\delta)$-DP mechanism using
321
+ \[
322
+ \widetilde O\left(\frac{\Lam(H+1)}{\varepsilon\tau^4}\right)
323
+ \]
324
+ iid records outputs a pool $\mathcal P$ of at most $\lceil6/\tau\rceil+1$ public subtree predicates, including the empty predicate, and a map $\rho$ from every subtree to $\mathcal P$. With probability at least $1-\beta$,
325
+ \[
326
+ \sup_u\mu(T_u\mathbin{\triangle}\rho(T_u))\le\tau.
327
+ \]
328
+ The pool and the entire map are functions only of the private transcript and the public tree. The cardinality bound holds even on utility failure.
329
+ \end{theorem}
330
+ \begin{proof}
331
+ Run the collision-CDF and histogram construction with $\eta=\tau$ and $\kappa=\tau/100$. At each selected level retain nodes whose released estimate exceeds $2\kappa$, with a hard cap of $\lceil1/\kappa\rceil$ nodes, choosing largest estimates with public ties. There are $O(\tau^{-2})$ levels, so the initial pool $\mathcal P_0$, with $\varnothing$ adjoined, has $O(\tau^{-3})$ members on every input.
332
+
333
+ On the histogram accuracy event, every qualifying node has true mass greater than $\kappa$, so the cap never removes a qualifying node. At every selected level, a node of mass greater than $3\kappa$ is retained and an omitted node has mass at most $3\kappa$.
334
+
335
+ At a selected depth map a retained query to itself and an omitted query to $\varnothing$. At an intermediate depth bracketed by $a<b$, let $A$ be its depth-$a$ ancestor. If $A$ is omitted, map the query to $\varnothing$. Otherwise choose the largest-estimate retained depth-$b$ descendant $W$ of $A$, if one exists, and map the query to $T_A$ when $W$ lies below it and to $\varnothing$ otherwise. If no such $W$ exists, the output is $\varnothing$.
336
+
337
+ Call this map $\rho_0$. If $M=\mu(T_A)<\tau/4$, either output differs from the query on mass at most $M$. If $M\ge\tau/4$, the collision bound gives a descendant of mass at least $M-\tau/16\ge3\tau/16$. It and $A$ are retained. Its gap from competitors exceeds $2\kappa$, so it is the selected $W$. When $W$ lies below the query, $T_W\subseteq T_u\subseteq T_A$ and the difference has mass at most $\tau/16$; otherwise the query has mass at most $\tau/16$. At selected depths the error is at most $3\kappa$. Thus
338
+ \[
339
+ \sup_u d_\mu(T_u,\rho_0(T_u))\le\tau/4.
340
+ \]
341
+ This argument also includes mass terminating at internal vertices.
342
+
343
+ Now prune $\mathcal P_0$. Each retained subtree has its released mass estimate, and the empty predicate has estimate zero. For two subtrees, public ancestry determines whether they are nested or disjoint. Estimate their symmetric-difference mass by the absolute difference of mass estimates in the nested case and their sum in the disjoint case. Set self-distance to zero. On the good event every such estimate has error at most $2\kappa$.
344
+
345
+ Start a greedy net with $\varnothing$. In a public order of the remaining pool, retain a predicate only if its estimated distance from every already retained predicate exceeds $\tau/2$. Cap the number of retained nonempty predicates at $\lceil6/\tau\rceil$ on arbitrary transcripts. On the good event retained predicates have true pairwise distance greater than
346
+ $\tau/2-2\kappa=0.48\tau$. Lemma~\ref{lem:packing} bounds their number by $2/(0.48\tau)<6/\tau$, so the cap is inactive.
347
+
348
+ Every initial predicate is either retained or has a retained representative at estimated distance at most $\tau/2$, hence true distance at most $0.52\tau$. Let $\pi$ assign that representative, with a public empty fallback when needed on failed transcripts, and define the final map $\rho=\pi\circ\rho_0$. The triangle inequality gives error at most
349
+ $\tau/4+0.52\tau<\tau$.
350
+ All retention, distances, and maps are postprocessing of the already private transcript. The sample and privacy bounds follow from Theorem~\ref{thm:tree}.
351
+ \end{proof}
352
+
353
+ The map $\rho$ need not preserve intersections, complements, or unions as identities. Boolean combinations will instead be controlled by their disagreement probability and evaluated jointly on fresh records.
354
+
355
+
356
+ \section{VC-one anchors and dimension robustness}
357
+
358
+ \subsection{Signed laminar representation}
359
+
360
+ \begin{lemma}\label{lem:laminar}
361
+ Let $A$ be a finite nonempty VC-one class. Fix $a_0\in A$, and set
362
+ \[
363
+ S_x=\{a\in A:a(x)\ne a_0(x)\}.
364
+ \]
365
+ The nonempty distinct sets $S_x$ are laminar: two are nested or disjoint. They form an inclusion tree after adjoining the root $A$. Every anchor maps to a vertex so that membership in $S_x$ is exactly membership in the subtree rooted at $S_x$. If this tree has height $H$, then
366
+ \[
367
+ H<2^{\LD(A)+1}.
368
+ \]
369
+ \end{lemma}
370
+ \begin{proof}
371
+ The label pattern $00$ on $S_x,S_y$ is witnessed by $a_0$. If two sets overlap without containment, their intersection and two directed differences witness $11,10,01$, shattering two original coordinates. Thus they are laminar.
372
+
373
+ Map each anchor to the smallest inclusion-tree set containing it. Two anchors in the same residual atom would have identical memberships in every $S_x$, hence identical original labels, so the map is well-defined and injective. Ancestor membership gives the required representation.
374
+
375
+ A strict nested chain of $H$ nonempty sets has $H+1$ witnesses: one in each successive difference, one in the smallest set, and $a_0$ outside the largest. On the corresponding coordinates, these witnesses form all $H+1$ threshold patterns, up to fixed coordinate complements and order reversal. Binary search shatters a tree of depth $\lfloor\log_2(H+1)\rfloor$. Thus $\LD(A)\ge\lfloor\log_2(H+1)\rfloor$, implying the displayed bound.
376
+ \end{proof}
377
+
378
+ \begin{corollary}[Private marginals for VC-one classes]\label{cor:vc1-release}
379
+ For every finite class $A$ with $\VC(A)\le1$, all coordinate means under an arbitrary law on $A$ can be estimated privately to uniform error $\eta$ using
380
+ \[
381
+ N_A=\widetilde O\left(\frac{\Lam(\LD(A)+1)}
382
+ {\varepsilon\eta^4}\right)
383
+ \]
384
+ iid anchor records. Privacy holds on every anchor list.
385
+ \end{corollary}
386
+ \begin{proof}
387
+ Use Lemma~\ref{lem:laminar} and Theorem~\ref{thm:tree}. For a coordinate with $a_0(x)=0$, report the estimated $S_x$ mass; for $a_0(x)=1$, report one minus that mass. Empty $S_x$ are exact public constants. The height bound gives $\Lam(H+1)=O(\Lam(\LD(A)+1))$. A VC-zero class contains one distinct function, so its means are public.
388
+ \end{proof}
389
+
390
+ This is a query-release construction. The established near-optimal PAC theorem for VC-one classes is due to Yan \cite{yan}; we do not present that existing PAC result as a new discovery.
391
+
392
+ \subsection{Hamming perturbations preserve finite dimension}
393
+
394
+ The order $O(d+k)$ in the next lemma is also a consequence of the known Boolean-closure theorem of Ghazi, Golowich, Kumar, and Manurangsi \cite{ggkm}: take XOR with the class of sets of size at most $k$, which has Littlestone dimension $k$. We include a direct proof with an explicit constant because the adaptive-tree accounting matters in this application.
395
+
396
+ \begin{lemma}[Dimension robustness]\label{lem:robust}
397
+ Suppose every $a\in A$ differs on at most $k$ domain points from some $h\in C$. If $d=\LD(C)$ and $d+k\ge1$, then
398
+ \[
399
+ \LD(A)\le5(d+k)-1.
400
+ \]
401
+ Likewise $\VC(A)\le5(\VC(C)+k)-1$ when $\VC(C)+k\ge1$. If $k=0$, the exact subclass bounds apply.
402
+ \end{lemma}
403
+ \begin{proof}
404
+ Suppose $A$ shatters a complete binary instance tree of depth $t$. Every root-to-leaf path has distinct queried points: a repeat would admit a continuation requiring opposite labels at that point. Each realized path is therefore within at most $k$ label errors of some $C$ hypothesis along its queried points.
405
+
406
+ Fix a set $E\subseteq[t]$ of error positions. Relabel branches of the input tree by flipping precisely these levels. Explicitly, the node at depth $i$ on a transformed history $u_{<i}$ receives the original instance indexed by $u_{<i}\mathbin{\oplus}\1_{E\cap[i-1]}$. A $C$ hypothesis explaining an original path $b$ with error mask $E$ realizes $b\oplus\1_E$ in this transformed tree. The map on label strings is injective.
407
+
408
+ The sequential Sauer bound therefore yields
409
+ \[
410
+ 2^t\le
411
+ \left(\sum_{i=0}^{d}\binom ti\right)
412
+ \left(\sum_{j=0}^{k}\binom tj\right).
413
+ \]
414
+ The sums are truncated at $t$. Sequential Sauer follows by induction from the fact that at least one branch of a first query has Littlestone dimension at most $d-1$.
415
+
416
+ Set $s=d+k$ and $t=5s$. Interpreting zero-exponent terms as one, the right side is at most
417
+ \[
418
+ (et/d)^d(et/k)^k
419
+ \le (2et/s)^s=(10e)^s<32^s=2^t.
420
+ \]
421
+ The middle inequality is the binary-entropy bound. This contradiction proves the Littlestone claim. For VC dimension, apply the same count to an ordinary shattered set and use ordinary Sauer. The separate error-mask tree transformation is essential; directly counting paths in the unchanged tree would not be valid.
422
+ \end{proof}
423
+
424
+ \begin{proposition}[A sharper corrupted-chain bound]\label{prop:chain}
425
+ Suppose $H$ anchor witnesses and $H$ distinct coordinates give the triangular labels $a_i(x_j)=\1\{j\le i\}$, and each $a_i$ has a witness $h_i\in C$ differing on at most $k$ coordinates. Then, with $d=\LD(C)$,
426
+ \[
427
+ H<(2k+1)2^{d+2}.
428
+ \]
429
+ The same assertion holds after fixed coordinatewise label complements.
430
+ \end{proposition}
431
+ \begin{proof}
432
+ Make a graph on $[H]$, joining $i\ne j$ if $h_i(x_j)$ or $h_j(x_i)$ is corrupted. In every induced subgraph there are at most $k$ times its vertex count edges, because each row has at most $k$ corruptions. Removing a vertex of degree at most $2k$ repeatedly gives a $(2k+1)$-coloring. Hence an independent set has size at least $H/(2k+1)$.
433
+
434
+ In its increasing order, all off-diagonal entries are correct thresholds. Retain at least half the indices with the same diagonal bit. The resulting $r\ge H/(2(2k+1))$ rows are distinct strict or inclusive prefixes. Removing one constant column yields all $r$ threshold patterns on $r-1$ columns. Thus $d\ge\lfloor\log_2 r\rfloor$, so $r<2^{d+1}$, proving the bound. Fixed complements preserve both errors and Littlestone dimension.
435
+ \end{proof}
436
+
437
+ Applied to a nested chain in the used-anchor laminar tree, this proposition controls its height directly by the original class dimension and $k$. The broader Lemma~\ref{lem:robust} is sufficient for all stated selector bounds.
438
+
439
+ \section{Private covers preserve correlated Boolean queries}
440
+
441
+ \subsection{A finite-query synthesis primitive}
442
+
443
+ The following is the finite-net exponential-mechanism construction underlying classical private synthetic-data release \cite{blr}. Its short proof is included to make the subsequent conditional use explicit.
444
+
445
+ \begin{lemma}[Finite-query release]\label{lem:finite-release}
446
+ For a finite nonempty public domain $D$ and a finite nonempty query family $\mathcal Q\subseteq\{0,1\}^D$, a pure $\varepsilon$-DP mechanism releases a synthetic list $T$ with
447
+ \[
448
+ \sup_{q\in\mathcal Q}|\E_Tq-\E_\nu q|\le\xi
449
+ \]
450
+ with probability at least $1-\zeta$ on iid records from any law $\nu$, provided
451
+ \begin{align*}
452
+ n=O\left(
453
+ \frac{\log(2|D|)\log(2|\mathcal Q|)}{\varepsilon\xi^3}
454
+ +\frac{\log(1/\zeta)}{\varepsilon\xi}
455
+ +\frac{\log(2|\mathcal Q|/\zeta)}{\xi^2}
456
+ \right).
457
+ \end{align*}
458
+ Privacy holds on all lists. The domain and query family may be functions of an earlier private transcript when the input block is fresh.
459
+ \end{lemma}
460
+ \begin{proof}
461
+ Choose $t=O(\log(2|\mathcal Q|)/\xi^2)$. Sampling $t$ times from the empirical distribution of any input list $S$, Hoeffding's inequality and a union bound imply the existence of $T_0\in D^t$ with
462
+ \[
463
+ \Delta(S,T_0)\le\xi/4,\qquad
464
+ \Delta(S,T)=\max_{q\in\mathcal Q}|\E_Sq-\E_Tq|.
465
+ \]
466
+ A replacement changes $\Delta(S,T)$ by at most $1/n$. Release $T\in D^t$ with probability proportional to $\exp(-\varepsilon n\Delta(S,T)/2)$. The likelihood-ratio argument for the exponential mechanism gives pure $\varepsilon$-privacy.
467
+
468
+ The probability of $\Delta(S,T)>\xi/2$ is at most
469
+ $|D|^t\exp(-\varepsilon n\xi/8)$. The first two terms in the sample bound make this at most $\zeta/2$. The last term makes $\sup_q|\E_Sq-\E_\nu q|\le\xi/2$ except with probability $\zeta/2$. Their combination proves the assertion. Conditional on an earlier transcript, the same proof applies to the now fixed $D,\mathcal Q$ and independent fresh records.
470
+ \end{proof}
471
+
472
+ \subsection{The composition theorem}
473
+
474
+ Let $\Omega$ be a finite nonempty public record domain and let $r\ge1$. For $j\in[r]$, let
475
+ $\phi_j:\Omega\to V(\mathcal T_j)$ be public maps into finite rooted trees of heights $H_j$. A record determines an entire tuple of vertices. Its components may be arbitrarily correlated.
476
+
477
+ For each public query $x$, choose vertices $u_{j,x}$ (or the empty predicate), bits $s_{j,x}$, and a Boolean gate $F_x$ from a finite public family $\mathcal G$ of functions on $r$ bits. Consider
478
+ \begin{equation}\label{eq:query-composition}
479
+ q_x(z)=F_x\left(
480
+ \1_{T_{1,u_{1,x}}}(\phi_1(z))\oplus s_{1,x},\ldots,
481
+ \1_{T_{r,u_{r,x}}}(\phi_r(z))\oplus s_{r,x}\right).
482
+ \end{equation}
483
+
484
+ \begin{theorem}[Correlation-preserving private release]\label{thm:composition}
485
+ For $0<\eta<1/4$, all means of the queries in \eqref{eq:query-composition} can be estimated simultaneously to error $\eta$, with failure at most $\beta$, by an all-input $(\varepsilon,\delta)$-DP mechanism.
486
+ For one fixed gate, a sufficient sample count is
487
+ \begin{equation}\label{eq:composition-rate}
488
+ \boxed{
489
+ N_{\rm comp}=
490
+ \widetilde O\left(
491
+ \frac{r^4\sum_{j=1}^r\Lam(H_j+1)}
492
+ {\varepsilon\eta^4}\right).
493
+ }
494
+ \end{equation}
495
+ For a finite gate family, add the finite-release cost of Lemma~\ref{lem:finite-release}, with $\xi=\eta/2$, $\zeta=\beta/2$, and
496
+ \[
497
+ \log(2|D|)=O(r\log(r/\eta)),\quad
498
+ \log(2|\widehat{\mathcal Q}|)
499
+ =O(\log(2|\mathcal G|)+r\log(r/\eta)).
500
+ \]
501
+ The height dependence is iterated logarithmic. There is no dependence on the number of vertices or branches. The construction does not assume independence between component coordinates.
502
+ \end{theorem}
503
+ \begin{proof}
504
+ Set $\tau=\eta/(4r)$. Reserve $r$ disjoint source blocks and run Theorem~\ref{thm:cover} separately on component $j$, with failure $\beta/(2r)$. These are different records, so parallel composition gives $(\varepsilon,\delta)$ for their joint transcript. Each component returns a pool $\mathcal P_j$ of $O(r/\eta)$ predicates and a public conditional map $\rho_j$.
505
+
506
+ Replace every subtree bit in \eqref{eq:query-composition} by its proxy, retaining its prescribed complement. For arbitrary bit vectors $b,c$ and any Boolean gate,
507
+ \[
508
+ |F(b)-F(c)|\le\sum_{j=1}^r\1\{b_j\ne c_j\}.
509
+ \]
510
+ Thus on the component success event,
511
+ \begin{equation}\label{eq:joint-proxy}
512
+ \sup_x\E_\mu|q_x-\widehat q_x|\le r\tau=\eta/4.
513
+ \end{equation}
514
+ This pointwise union bound requires no probabilistic independence.
515
+
516
+ The nonempty predicates in $\mathcal P_j$ remain laminar. A laminar family of $M_j$ sets has at most $M_j+1$ nonempty membership atoms: its inclusion forest gives at most one residual atom per node and one outside atom. Conditional on the transcript, partition each component vertex domain by these predicates. Let $D_j$ be the resulting public atom alphabet and let $D=\prod_jD_j$. Then
517
+ \[
518
+ |D_j|=O(r/\eta),\qquad
519
+ \log(2|D|)=O(r\log(r/\eta)).
520
+ \]
521
+ Each source record maps to its tuple of atom labels. Synthetic tuples need not correspond to original records in $\Omega$; they need only define the proxy queries.
522
+
523
+ Let $\widehat{\mathcal Q}$ contain every query obtained by choosing a pool predicate or its complement in each component and a gate in $\mathcal G$. It has size at most
524
+ $|\mathcal G|\prod_j(2|\mathcal P_j|)$.
525
+ Reserve the last block's size in advance using the deterministic pool caps to bound $|D|$ and $|\widehat{\mathcal Q}|$. Map this new fixed block to $D$ and apply Lemma~\ref{lem:finite-release}. Return the covers, proxy maps, and the synthetic means of the corresponding proxy queries. The last release has error at most $\eta/2$, and \eqref{eq:joint-proxy} adds at most $\eta/4$. The failure probabilities sum to at most $\beta$.
526
+
527
+ Changing a first-stage record affects later work only through its private cover transcript. Changing a last-block record changes one atom tuple. Fixed-block adaptive parallel composition therefore yields the same $(\varepsilon,\delta)$ for the full mechanism.
528
+
529
+ The cover blocks cost
530
+ $\widetilde O(r^4\sum_j\Lam(H_j+1)/(\varepsilon\eta^4))$.
531
+ For one fixed gate the last block costs $\widetilde O(r^2/(\varepsilon\eta^3))$, which is dominated by the displayed cover cost. Finite gate families retain the additional stated term. On arbitrary transcripts all caps, maps, atoms, and output rules remain defined.
532
+ \end{proof}
533
+
534
+ The logarithm of the compressed product-domain size is small; the domain itself can be exponential in $r$. This is an information-theoretic sample bound, not an efficient enumeration algorithm.
535
+
536
+ \subsection{VC-one components and an intrinsic dimension certificate}
537
+
538
+ Suppose each input hypothesis $h$ has a public factorization
539
+ \[
540
+ h(x)=F(a_1(h)(x),\ldots,a_r(h)(x)),\qquad a_j(h)\in A_j,
541
+ \]
542
+ where $\VC(A_j)\le1$ and $F$ is fixed. The signed laminar representation in Lemma~\ref{lem:laminar} puts this in the preceding theorem, with
543
+ $\Lam(H_j+1)=O(\Lam(\LD(A_j)+1))$.
544
+ The law of the factor tuple can be arbitrary.
545
+
546
+ \begin{corollary}[Full Boolean products]\label{cor:full-product}
547
+ Suppose $A=F(A_1\times\cdots\times A_r)$ is the full pointwise product, every $A_j$ contains both constant functions, and $F$ is essential in each input. With $d=\LD(A)$, uniform private coordinate-mean release uses
548
+ \begin{equation}\label{eq:intrinsic-release}
549
+ \widetilde O\left(\frac{r^5\Lam(d+1)}{\varepsilon\eta^4}\right)
550
+ \end{equation}
551
+ records.
552
+ \end{corollary}
553
+ \begin{proof}
554
+ For each $j$, fix the other inputs to constant bits witnessing that $F$ depends on its $j$th bit. As $a_j$ varies, the resulting subfamily of $A$ is $A_j$ or its pointwise complement. Therefore $\LD(A_j)\le d$. Choose one representing tuple for each distinct row of $A$ by a public tie rule, and apply Theorem~\ref{thm:composition}. This fixed preprocessing preserves replacement adjacency.
555
+ \end{proof}
556
+
557
+ For an arbitrary subclass, the component dimensions remain explicit unless such an isolation embedding or another bound is supplied. Cancellation prevents their automatic replacement by the output dimension: identical components combined by XOR give the constant-zero class while the components can have arbitrarily large Littlestone dimension.
558
+
559
+ Axis-aligned boxes on a finite $p$-dimensional grid are full conjunctions of $2p$ endpoint threshold families with constants, so the certificate applies. Their membership probabilities under a distribution on boxes can depend on endpoint correlations. For example, the laws assigning probability $1/2$ to each of $(0,0),(1,1)$ and to each of $(0,1),(1,0)$ have the same one-bit marginals but AND probabilities $1/2$ and $0$. The synthesis stage preserves this difference.
560
+
561
+ Intervals also fit, escaping the earlier static sparse-anchor obstruction. For intervals alone, the simpler identity
562
+ $\Pr[a\le x\le b]=\Pr[a\le x]-\Pr[b<x]$
563
+ already suffices when $a\le b$. Interval marginal release is therefore not the distinctive correlation claim.
564
+
565
+ \subsection{The gate family cannot be omitted from the bound}
566
+
567
+ There are $2^{2^r}$ Boolean functions of $r$ bits. Allowing all gates makes $\log|\mathcal G|=2^r$, which must remain in the finite-query cost.
568
+
569
+ This is necessary. Index $2^r$ records by bit strings $s\in\{0,1\}^r$. Let component $j$ be the constant function with value $s_j$, so every component family is just $\{0,1\}$ and has VC and Littlestone dimension one. A coordinate-dependent gate can be the membership truth table of any subset of those records. The resulting queries are all subset probabilities, whose exponential statistical lower bound is proved in Theorem~\ref{thm:release-lower}.
570
+
571
+
572
+ \section{Private selection under sparse anchor perturbations}
573
+
574
+ \begin{definition}
575
+ A public radius-$k$ anchor assignment, where $k$ is a nonnegative integer, is a total map $a:C\to A$ with
576
+ \[
577
+ |\{x:h(x)\ne a(h)(x)\}|\le k\qquad(h\in C).
578
+ \]
579
+ We may discard unused anchors, so every remaining anchor has a witness in $C$ within distance $k$.
580
+ A balanced-coordinate selector must be private on every list and, for iid input from any law $\mu$ satisfying
581
+ $p_x\in[0.4,0.6]$ for some $x$, return a coordinate with population frequency in $[1/4,3/4]$ with probability at least $1-\gamma$.
582
+ \end{definition}
583
+
584
+ \subsection{Dual dimension of the perturbed class}
585
+
586
+ \begin{lemma}\label{lem:dual-sparse}
587
+ Let $q_A=\VC(A^*)$. Under a radius-$k$ assignment, the original class satisfies
588
+ \[
589
+ q_C=O(q_A+\log(k+1)).
590
+ \]
591
+ More precisely, if $k\ge1$, put $r_k=1+\lfloor\log_2k\rfloor$. Then $q_C<5(q_A+r_k)$ whenever the sum is positive.
592
+ \end{lemma}
593
+ \begin{proof}
594
+ For each coordinate define the residual query
595
+ $e_x(h)=h(x)\oplus a(h)(x)$.
596
+ If these queries shatter $r$ input hypotheses, one selected coordinate per binary pattern gives $2^{r-1}$ positive deviations in every row. Thus $k\ge2^{r-1}$, so the residual query class has VC dimension at most $r_k$.
597
+
598
+ The anchor queries on $C$ have VC dimension at most $q_A$. Original queries are synchronized XORs of anchor and residual queries, a subclass of their full product. On $n$ rows the number of traces is at most the product of the two Sauer bounds. With $s=q_A+r_k$, the same entropy calculation as in Lemma~\ref{lem:robust} rules out shattering $5s$ rows. For $k=0$ the residuals vanish.
599
+ \end{proof}
600
+
601
+ \subsection{Scalar mean bins and a gated dynamic range}
602
+
603
+ Set $u=1/1000$. On a first fixed source block, map each $h$ to its anchor and run an all-input private mechanism estimating every anchor mean to error at most $u$, with failure at most $\gamma/3$. Clip its output $\widehat p_A(x)$ to $[0,1]$.
604
+
605
+ Bin coordinates by $\lfloor\widehat p_A(x)/u\rfloor$. There are at most $1001$ bins. In each bin choose up to $R=40k+1$ coordinates under a fixed public order, taking all coordinates when fewer exist. Let their union be $P_0$. This pool is public conditional on the first private output.
606
+
607
+ On a second, independent block of $n$ original hypotheses define
608
+ \[
609
+ c_x=\sum_i h_i(x),\quad Q_x=\min(c_x,n-c_x),\quad
610
+ D_x=\sum_i\1\{h_i(x)\ne a(h_i)(x)\}.
611
+ \]
612
+ Use the range consisting of $P_0$, every nonpublic coordinate with $D_x\ge n/100$, and a dummy $\bot$. Give these candidates scores
613
+ \begin{equation}\label{eq:gate}
614
+ s_S(x)=
615
+ \begin{cases}
616
+ Q_x,&x\in P_0,\\
617
+ \min(Q_x,20D_x),&x\notin P_0,\ D_x\ge n/100,
618
+ \end{cases}
619
+ \qquad s_S(\bot)=3n/10.
620
+ \end{equation}
621
+ Sample with weights proportional to $\exp(\varepsilon s_S/40)$.
622
+
623
+ \begin{theorem}[Marginal-estimator to selector transfer]\label{thm:transfer}
624
+ If the anchor estimator above uses $N_A(u;\varepsilon,\delta,\gamma/3)$ records, this construction is a balanced-coordinate selector using
625
+ \begin{equation}\label{eq:selector-general}
626
+ N_A(u;\varepsilon,\delta,\gamma/3)
627
+ +O\left(\frac{q_A+\log(k+1)+\log(1/(\gamma\delta))}
628
+ {\varepsilon}\right)
629
+ \end{equation}
630
+ original records. It is $(\varepsilon,\delta)$-DP on all lists.
631
+ \end{theorem}
632
+ \begin{proof}[Privacy]
633
+ Condition on any first-stage output, including one failing its utility guarantee. The pool $P_0$ is now fixed. Since $\sum_xD_x\le nk$, the dynamic range contains at most $100k$ coordinates. A replacement can change deviation status at at most $2k$ coordinates. Common scores have sensitivity at most $20$.
634
+
635
+ A newly admitted coordinate has $D_x\le n/100+1$, hence score at most $n/5+20$. The dummy bounds its probability; the total mass outside the common range in either direction is at most
636
+ \[
637
+ b_{\rm esc}\le2k\exp(-\varepsilon n/400+\varepsilon/2).
638
+ \]
639
+ On the common range the exponential mechanism is $\varepsilon$-private. Lemma~\ref{lem:common} applies. For $k\ge1$, a sufficient privacy condition is
640
+ \[
641
+ n\ge200+\frac{400}{\varepsilon}
642
+ \log\frac{2k(1+2e^\varepsilon)}{\delta}.
643
+ \]
644
+ When $k=0$, the range is fixed and privacy is pure. Adaptive parallel composition across the two fixed blocks gives the asserted parameters.
645
+ \end{proof}
646
+ \begin{proof}[Utility]
647
+ Use Lemma~\ref{lem:dual-sparse} and uniform convergence so that all empirical original and anchor means in the second block have error at most $u$, except with probability $\gamma/3$. Let $x$ be a promised balanced coordinate. Then $Q_x\ge0.399n$.
648
+
649
+ If $x\in P_0$ it has that score. If $D_x\ge n/40$, its dynamic score also equals $Q_x$. Otherwise its bin has $R$ public representatives. The deviation budget ensures that one, $y$, has $D_y\le n/40$. First-stage accuracy and equal estimated-mean bins give $|p_A(x)-p_A(y)|\le3u$. Second-stage concentration gives an empirical anchor-count difference at most $5un$. Therefore
650
+ \[
651
+ |c_x-c_y|\le D_x+D_y+5un\le0.055n,
652
+ \]
653
+ and $Q_y\ge0.344n$. Thus some output always has score at least $0.34n$ on these events.
654
+
655
+ The range size is at most
656
+ $M=1001(40k+1)+100k+1$.
657
+ The dummy and all outputs with score below $0.3n$ have total probability at most
658
+ $M\exp(-\varepsilon n/1000)$.
659
+ Taking $n\ge(1000/\varepsilon)\log(3M/\gamma)$ makes this at most $\gamma/3$. Every remaining output has $Q_x\ge0.3n$, hence population minority probability at least $0.299>1/4$.
660
+
661
+ For $k=0$, every residual is zero; a single representative in the same mean bin gives the same conclusion. The concentration and selection conditions have the order in \eqref{eq:selector-general}.
662
+ \end{proof}
663
+
664
+ \begin{theorem}[VC-one anchor selector]\label{thm:main-selector}
665
+ Let $C$ have Littlestone dimension $d$ and a public radius-$k$ assignment into a finite class $A$ with $\VC(A)\le1$. Then $C$ has an all-input private balanced-coordinate selector with sample complexity
666
+ \[
667
+ \boxed{\widetilde O\left(
668
+ \frac{\Lam(d)+\log(k+1)}{\varepsilon}\right),}
669
+ \]
670
+ with additional logarithmic factors in confidence and approximate-privacy parameters. No maximum-class or ambient-domain-size assumption is used.
671
+ \end{theorem}
672
+ \begin{proof}
673
+ Discard unused anchors. Lemma~\ref{lem:robust} bounds their Littlestone dimension by $O(d+k)$; the degenerate zero-dimensional cases are public. Corollary~\ref{cor:vc1-release} supplies the anchor estimator at constant accuracy $u$.
674
+
675
+ An arbitrary VC-one class has dual VC dimension at most three: if four hypotheses were dual-shattered, two appropriate coordinate patterns on these four rows would realize all four binary pairs, contradicting primal VC one. Thus Theorem~\ref{thm:transfer} costs
676
+ $\widetilde O((\Lam(d+k)+\log(k+1))/\varepsilon)$.
677
+ Finally $\Lam(d+k)=O(\Lam(d)+\log(k+1))$. All constructions are total and private on all lists.
678
+ \end{proof}
679
+
680
+ \subsection{Sparse perturbations of correlated Boolean anchors}
681
+
682
+ \begin{corollary}\label{cor:composite-selector}
683
+ Suppose a class $C$ has a public radius-$k$ assignment into anchors that factor through one fixed Boolean gate on $r$ signed laminar component trees of heights $H_j$. Then an all-input private balanced-coordinate selector exists with sample bound
684
+ \[
685
+ \widetilde O\left(
686
+ \frac{r^4\sum_j\Lam(H_j+1)}{\varepsilon}
687
+ +\frac{r\log(r+1)+\log(k+1)+\log(1/(\gamma\delta))}
688
+ {\varepsilon}\right).
689
+ \]
690
+ The law of the component tuple is unrestricted. For VC-one component families, their individual Littlestone dimensions replace the height parameters as above.
691
+ \end{corollary}
692
+ \begin{proof}
693
+ Use Theorem~\ref{thm:composition} as the anchor marginal estimator in Theorem~\ref{thm:transfer}, at its fixed tolerance $u=1/1000$. On $t$ source records, each laminar component yields at most $2t-1$ nonempty trace sets. Optional complementation and the empty predicate increase this to at most $4t+2$ patterns. The fixed gate therefore gives at most $(4t+2)^r$ final query patterns. The inequality $2^t\le(4t+2)^r$ forces $t=O(r\log(r+1))$, bounding the anchor dual VC dimension. Substitute this bound in Theorem~\ref{thm:transfer}.
694
+ \end{proof}
695
+
696
+ This corollary preserves logarithmic dependence on the number of changed coordinates. It does not assume that a composite component's dimension is controlled by $\LD(C)$ without an explicit certificate. The maximum-class recursion below accepts this selector with the stated component parameters. For exact factorizations, the direct PAC reduction that follows is simpler and does not need maximum-class geometry.
697
+
698
+
699
+ \section{A direct PAC reduction for arbitrary classes}
700
+
701
+ Uniform \emph{population} marginal release makes the maximum-class carrier geometry unnecessary. The following elementary reduction isolates exactly what is required.
702
+
703
+ \begin{theorem}[Private mean aggregation implies PAC learning]\label{thm:direct-pac}
704
+ Let $C$ be any nonempty finite public binary class of VC dimension $v$. Suppose an all-input $(\varepsilon,\delta)$-DP mechanism, on $N_C$ iid hypotheses from every law $\mu$ on $C$, estimates every population mean $\Pr_{h\sim\mu}[h(x)=1]$ to a fixed error $\eta<1/2$, with failure at most $\beta$. Then $C$ has a possibly improper all-input $(\varepsilon,\delta)$-DP realizable PAC learner using
705
+ \[
706
+ O\left(\frac{(v+1)\log(2/\alpha)}{\alpha}\right)N_C
707
+ \]
708
+ labeled examples when $\eta$ is an absolute constant. No maximum-class, dual-VC, Helly, or isometry assumption is needed.
709
+ \end{theorem}
710
+ \begin{proof}
711
+ Put $\zeta=\alpha(1/2-\eta)/2$. A proper ERM on
712
+ \[
713
+ b=O\left(\frac{(v+1)\log(2/\zeta)}{\zeta}\right)
714
+ \]
715
+ examples has expected error at most $\zeta$: use ordinary realizable accuracy and failure each at most $\zeta/2$. Define the ERM on all labeled lists by a minimum-loss public tie rule.
716
+
717
+ Reserve $N_C$ disjoint raw-data blocks of size $b$, run the ERM on each, and feed the resulting hypotheses to the marginal mechanism. The ERM outputs are iid from one unconditional law $\mu$. There is no conditioning on every ERM block succeeding, and no union bound over those blocks. Output
718
+ $f(x)=\1\{\widehat p(x)\ge1/2\}$.
719
+
720
+ Fix target $c\in C$ and example law $P$. On the single successful release event, every point with $f(x)\ne c(x)$ satisfies
721
+ \[
722
+ \Pr_{h\sim\mu}[h(x)\ne c(x)]\ge1/2-\eta.
723
+ \]
724
+ Integrating over $P$ yields
725
+ \[
726
+ \err_P(f,c)
727
+ \le\frac{\E_{h\sim\mu}\err_P(h,c)}{1/2-\eta}
728
+ \le\frac{\zeta}{1/2-\eta}=\alpha/2.
729
+ \]
730
+ Thus failure is at most $\beta$. A raw-record replacement affects one ERM block and at most one hypothesis supplied to the private mechanism. This adjacency-preserving preprocessing and threshold postprocessing prove the same privacy parameters on all raw datasets.
731
+ \end{proof}
732
+
733
+ The premise concerns population means. An empirical-only release theorem cannot be substituted without an additional generalization argument. Inserting unrestricted dual-VC uniform convergence would reintroduce an exponential cost in $v$.
734
+
735
+ \begin{corollary}[Arbitrary subclasses of controlled Boolean products]\label{cor:composition-pac}
736
+ Let $C$ be any finite class of VC dimension $v$ with a public exact factorization through a fixed Boolean gate on $r$ VC-one component families $A_j$. It has a realizable private PAC learner with
737
+ \[
738
+ \widetilde O\left(
739
+ \frac{(v+1)r^4}{\alpha\varepsilon}
740
+ \sum_{j=1}^r\Lam(\LD(A_j)+1)\right)
741
+ \]
742
+ samples. Under an isolation certificate $\LD(A_j)\le d=\LD(C)$ for every component, this becomes
743
+ \[
744
+ \boxed{\widetilde O\left(
745
+ \frac{(v+1)r^5\Lam(d+1)}{\alpha\varepsilon}\right).}
746
+ \]
747
+ \end{corollary}
748
+ \begin{proof}
749
+ Apply Theorem~\ref{thm:composition} at any fixed $\eta<1/2$, for example $\eta=1/8$, and then Theorem~\ref{thm:direct-pac}.
750
+ \end{proof}
751
+
752
+ Private PAC closure under fixed Boolean composition was already established by Alon, Beimel, Moran, and Stemmer \cite{abms}; Karni and Tsfadia subsequently improved component-wise overheads \cite{kt}. The corollary is a direct consequence of the stronger query-release construction developed here. We do not claim the first PAC closure theorem or an improvement over those works' optimized component dependence.
753
+
754
+
755
+ \section{A self-contained maximum-class learning consequence}
756
+
757
+ This section supplies the geometric and statistical reduction needed to turn the selector into a PAC learner. It retains the same original law at every dimension drop.
758
+
759
+ \subsection{Maximum-class geometry}
760
+
761
+ A class $C\subseteq\{0,1\}^m$ of VC dimension $v$ is maximum if
762
+ $|C|=\Phi_v(m)=\sum_{i=0}^v\binom mi$.
763
+
764
+ \begin{lemma}\label{lem:maxgeom}
765
+ Coordinate projections of a maximum class are maximum, and a coordinate reduction is maximum of dimension one smaller. Every infeasible partial labeling has an infeasible restriction on at most $v+1$ coordinates. Every nonempty version space is isometric in its Hamming cube.
766
+ \end{lemma}
767
+ \begin{proof}
768
+ At coordinate $x$, let $C_0,C_1$ be the fibers with $x$ removed. Projection $P=C_0\cup C_1$ has VC dimension at most $v$, reduction $R=C_0\cap C_1$ at most $v-1$, and $|C|=|P|+|R|$. Sauer and Pascal force both cardinality bounds to be tight. Iterate for all projections.
769
+
770
+ If a partial vector on a coordinate set $B$ is absent from the maximum projection, adding it exceeds Sauer's bound and shatters some $v+1$ coordinates. Its restriction supplies a previously missing pattern there, giving the required inconsistency witness. Smaller coordinate projections are full cubes.
771
+
772
+ For isometry, write $\sh(F)$ for the shattered sets of a finite binary family $F$. The inequality $|F|\le|\sh(F)|$ follows by induction on a coordinate split into fibers $A,B$, using
773
+ \[
774
+ |\sh(F)|=|\sh(A\cup B)|+|\sh(A)\cap\sh(B)|,\qquad
775
+ \sh(A\cap B)\subseteq\sh(A)\cap\sh(B).
776
+ \]
777
+ Call equality ample. Maximum classes are ample. For ample $F$, the chain
778
+ \[
779
+ |A|+|B|\le|\sh(A)|+|\sh(B)|
780
+ \le|\sh(A\cup B)|+|\sh(A)\cap\sh(B)|=|F|
781
+ \]
782
+ is equality throughout, so both fibers are ample and
783
+ $\sh(A\cup B)=\sh(A)\cup\sh(B)$.
784
+ If both fibers were nonempty and disjoint, the empty shattered set in their intersection would imply
785
+ $|\sh(A\cup B)|<|A\cup B|$, a contradiction. Thus nonempty opposite fibers share an edge. Induction gives connectedness of every nonempty coordinate fiber.
786
+
787
+ For two hypotheses, fix all agreeing coordinates. In the connected resulting fiber, any first edge toward a path changes a differing coordinate toward the other hypothesis. Fix that newly agreeing coordinate and repeat. This gives a path of exactly their Hamming distance, within the version space.
788
+ \end{proof}
789
+
790
+ For an active coordinate define the positive carrier
791
+ \[
792
+ K_x^+=\{h\in C:h(x)=1,\ h^{\oplus x}\in C\}.
793
+ \]
794
+ Deleting its fixed coordinate gives the reduction.
795
+
796
+ \begin{lemma}[Balanced carrier]\label{lem:carrier}
797
+ Let $C$ be maximum VC $w\ge1$, with law $\nu$. If $0<\rho<1/(w+1)$ and both labels at $x$ have probability greater than $w\rho$, then $K_x^+$ contains a hypothesis respecting every literal of $\nu$-probability at least $1-\rho$. It is maximum VC $w-1$ after removing $x$, and its Littlestone dimension cannot increase.
798
+ \end{lemma}
799
+ \begin{proof}
800
+ Any at most $w+1$ heavy literals have positive intersection by the union bound. Adding the constraint $h(x)=b$, any witness of size at most $w+1$ still has intersection probability at least $\Prb[h(x)=b]-w\rho>0$. The preceding Helly property gives a hypothesis on each side of $x$ satisfying all heavy literals. An isometric path within that version space crosses an $x$-edge, whose positive endpoint is the required carrier member. The dimension statements follow from reduction and subclass monotonicity.
801
+ \end{proof}
802
+
803
+ \subsection{Original constraints, generation, and a stable exit}
804
+
805
+ For an original source law $\mu$ on $C_0$, set
806
+ \[
807
+ L_\mu(r)=\{(x,b):\Prb_\mu[h(x)=b]\ge1-r\}.
808
+ \]
809
+ If $\E_\mu\err_P(h)\le\eta_0$, any function extending $L_\mu(r)$ has error at most $\eta_0/r$: at any point it misclassifies, the original error probability must exceed $r$.
810
+
811
+ Fix $v=\VC(C_0)$ and define
812
+ \[
813
+ r_j=\frac1{8v}-\frac{j}{16v^2},\quad
814
+ a=\frac1{32v^2},\quad b=\frac1{16v},\quad
815
+ r_v=\frac1{16v}.
816
+ \]
817
+ The invariant is that the current class $C_j$ contains a member extending $L_\mu(r_j)$. It initially holds by the coordinate Helly property and the union bound.
818
+
819
+ To generate one member of $C_j$, take
820
+ \[
821
+ t=O\bigl(v^4(q+\log(v+1))\bigr),\qquad q=\VC(C_0^*),
822
+ \]
823
+ fresh original $\mu$-records. Collect literals of empirical frequency at least $1-(r_j-a)$. On the simultaneous concentration event,
824
+ \[
825
+ L_\mu(r_j-2a)\subseteq\widehat L\subseteq L_\mu(r_j).
826
+ \]
827
+ Choose a public-tie member of $C_j$ satisfying $\widehat L$, or a fixed fallback if infeasible. The invariant ensures feasibility on this event. The resulting unconditional law $\nu_j$ assigns mass at least $1-b$ to hypotheses simultaneously extending $L_\mu(r_{j+1})$. Generator failures are part of this law; they are not union-bounded over all generated records.
828
+
829
+ For a generated list $g_1,\ldots,g_n$, the stable-majority mechanism uses the coordinatewise majority $M_S$, tie zero, and
830
+ \[
831
+ r(S)=\min_x\left|\sum_i g_i(x)-n/2\right|,\qquad
832
+ \tau=1+\varepsilon^{-1}\log(1/(2\delta)).
833
+ \]
834
+ It releases $M_S$ if $r(S)+\Lap(1/\varepsilon)>\tau$, and otherwise $\bot$.
835
+ If neighbors have the same majority, Laplace privacy applies. If their majority differs, both margins are at most one, so each releases with probability at most $\delta$. Their output laws then have total variation at most $\delta$. Thus privacy holds on all lists.
836
+
837
+ With
838
+ \[
839
+ N_{\rm st}=O\left(q+\log(1/\gamma)
840
+ +\varepsilon^{-1}\log(1/(\gamma\delta))\right)
841
+ \]
842
+ generated records, one event of probability at least $1-\gamma$ ensures both that a released function respects all $0.9$-heavy $\nu_j$ literals, and that failure implies a coordinate of frequency in $[0.4,0.6]$. To prove this, require empirical coordinate error at most $1/20$ and Laplace noise at least $-\varepsilon^{-1}\log(1/\gamma)$. If no balanced coordinate exists, every empirical margin exceeds $n/20$, so a sufficiently large $n$ forces release. This is a joint event, not a guarantee conditioned on release.
843
+
844
+ \subsection{The recursion and its sample bound}
845
+
846
+ At stage $j$, run the stable mechanism on fresh generated records. If it releases, return its function on the original domain, including all previously fixed coordinates. Otherwise run the balanced selector on a second independent generated block. On a valid active output coordinate $x$, replace $C_j$ by $K_x^+$ and fix $x$ to one. Validation checks only public structural validity; it does not test balance on raw data. At dimension zero return the sole hypothesis. Failures return a public fallback.
847
+
848
+ Put $\gamma=\beta/(2v)$. On a good stage, original literals in $L_\mu(r_{j+1})$ have $\nu_j$-probability at least $1-b>0.9$, so a stable release respects them. Otherwise the selector's balance allows Lemma~\ref{lem:carrier} with $\rho=1/(8w)$, where $w=v-j$. Both label probabilities exceed $w\rho=1/8$, and $b<\rho$. The next carrier therefore retains the invariant. At most $2v$ bad module events give total failure at most $\beta$.
849
+
850
+ Every successful output has error at most $16v\eta_0$. Fixed disjoint source blocks for every generated record and every module give the same $(\varepsilon,\delta)$ privacy parameters. One original record affects one generated record in one private module. The source count is
851
+ \begin{equation}\label{eq:descent-source}
852
+ t\sum_{w=1}^v
853
+ \left(N_{\rm st}(q;\varepsilon,\delta,\gamma)
854
+ +B_w(d;\varepsilon,\delta,\gamma)\right),
855
+ \end{equation}
856
+ where $B_w$ is the applicable balanced-selector bound; its dependence on the inherited radius $k$ is suppressed in this notation.
857
+
858
+ \begin{theorem}[Maximum classes near VC-one anchors]\label{thm:pac-max}
859
+ Let $C_0$ be finite maximum VC $v\ge1$, with Littlestone dimension $d$, and suppose it admits a public radius-$k$ assignment into a VC-one anchor family. There is a possibly improper, all-input $(\varepsilon,\delta)$-DP realizable $(\alpha,\beta)$-PAC learner with sample complexity
860
+ \begin{equation}\label{eq:pac-max}
861
+ \widetilde O\left(
862
+ \frac{v^7(q+\log(v+1))}{\alpha\varepsilon}
863
+ \left[q+\Lam(d)+\log(k+1)+\log\frac{v}{\beta\delta}\right]
864
+ \right),\qquad q=\VC(C_0^*).
865
+ \end{equation}
866
+ In particular, $q\le2v+1$, so a bound $\log(k+1)\le\operatorname{poly}(v,\Lam(d))$ suffices for the conjectured rate on this family.
867
+ \end{theorem}
868
+ \begin{proof}
869
+ A proper ERM using $r=O((v^2/\alpha)\log(64v/\alpha))$ examples has expected error at most $\eta_0=\alpha/(64v)$: take its ordinary realizable accuracy and failure to be at most $\eta_0/2$. Define the ERM by a minimum-loss public tie rule on arbitrary data. Its output law is the original $\mu$ used throughout.
870
+
871
+ Every carrier is a subclass of $C_0$. Restrict the original anchors and assignment to its active coordinates. The radius cannot increase, and the anchor VC dimension stays at most one. Theorem~\ref{thm:main-selector} supplies every selector in \eqref{eq:descent-source}; all carriers have Littlestone dimension at most $d$. Multiplying that source count by $r$ proves \eqref{eq:pac-max}. Every original labeled record belongs to one ERM block, so the same privacy proof applies. The successful error is at most $16v\eta_0=\alpha/4$.
872
+
873
+ The established extremal-class theorem of Chase et al.\ gives $q\le2v+1$ for maximum classes \cite{chase}. This cited input makes the statistical overhead polynomial in the primal dimension.
874
+ \end{proof}
875
+
876
+ \begin{corollary}[Maximum classes near Boolean anchors]\label{cor:max-composite}
877
+ Let $C$ be maximum of VC dimension $v\ge1$ and Littlestone dimension $d$, with a public radius-$k$ assignment into a fixed-gate composition of $r$ VC-one families $A_j$. Put $q=\VC(C^*)\le2v+1$ and
878
+ $S_A=\sum_{j=1}^r\Lam(\LD(A_j)+1)$.
879
+ Then a realizable private PAC learner has sample complexity
880
+ \[
881
+ \widetilde O\left[
882
+ \frac{v^7(q+\log(v+1))}{\alpha\varepsilon}
883
+ \left(q+r^4S_A+r\log(r+1)+\log(k+1)
884
+ +\log\frac{v}{\beta\delta}\right)\right].
885
+ \]
886
+ If the assignment is onto the full Boolean product, each component contains both constants, and the gate is essential in each input, then
887
+ $S_A=O(r\Lam(d+k+1))$.
888
+ \end{corollary}
889
+ \begin{proof}
890
+ Use Corollary~\ref{cor:composite-selector} for every selector in \eqref{eq:descent-source}. Restrict the original components to each carrier's remaining coordinates; their dimensions and the perturbation radius cannot increase. Multiply the source count by the same ERM block size used in Theorem~\ref{thm:pac-max}.
891
+
892
+ For the final assertion, surjectivity ensures that every full-product anchor has a witness in $C$ within distance $k$. Lemma~\ref{lem:robust} gives $\LD(A)=O(d+k)$, and the full-product isolation argument gives $\LD(A_j)\le\LD(A)$. The degenerate zero-dimensional case is immediate.
893
+ \end{proof}
894
+
895
+ Under this final certificate, polynomial bounds on $r$ and $\log(k+1)$ in $v,\Lam(d)$ suffice for the conjectured type of rate on the stated maximum classes. No such bounds are asserted for general classes.
896
+
897
+
898
+ \section{An arbitrary-class PAC transfer at polynomial radius}
899
+
900
+ The next result removes the maximum-class assumption, with polynomial rather than logarithmic dependence on $k$.
901
+
902
+ \begin{theorem}[Realizable-only sparse PAC transfer]\label{thm:pac-general}
903
+ Suppose each $h\in C$ differs at at most $k$ points from some $a\in A$. Let $A^\bot$ extend every anchor by zero on one new public dummy point. Suppose $A^\bot$ has an all-input $(\varepsilon,\delta)$-DP realizable learner using $M\ge100/\alpha$ records at accuracy $\alpha/16$ and failure $1/8$. Then $C$ has an all-input private, possibly improper realizable PAC learner with the same $(\varepsilon,\delta)$ and sample count
904
+ \[
905
+ O\left(
906
+ \frac{kM}{\varepsilon}\log\frac{kM+2}{\beta\delta}
907
+ +M\log\frac1\beta
908
+ +\frac1{\alpha\varepsilon}
909
+ \log\frac{\log(2/\beta)}{\beta}
910
+ \right).
911
+ \]
912
+ The first term is omitted for $k=0$. The anchor assignment need not be known to the algorithm.
913
+ \end{theorem}
914
+ \begin{proof}
915
+ Assume first $k\ge1$ and set $R=100kM$. On a first block, release labeled categories $(x,y)$ of observed count $c_{x,y}$ when
916
+ \[
917
+ c_{x,y}+\Lap(2/\varepsilon)>1+\frac2\varepsilon\log\frac8\delta.
918
+ \]
919
+ Use independent noise. The common-support Laplace proof of Lemma~\ref{lem:hist} applies; each noncommon singleton is released with probability at most $\delta/16$, so the comparison error is at most $(1+e^\varepsilon)\delta/16\le\delta$. If both labels occur in the output, retain a public preferred label. On realizable input every released label is correct.
920
+
921
+ There are at most $R$ points of marginal mass at least $1/R$. Chernoff lower tails for their counts, negative Laplace tails, and a union bound show that
922
+ \[
923
+ n_{\rm heavy}=O\left(\frac R\varepsilon\log\frac{R+2}{\beta\delta}\right)
924
+ \]
925
+ records release every such point with failure at most $\beta/4$. Let $E$ be the released set and $\ell_E$ its labels. Outside $E$, target/anchor disagreement has mass
926
+ $r_0\le k/R=1/(100M)$.
927
+
928
+ Reserve $L=\lceil\log(4/\beta)/\log4\rceil$ independent anchor-learning blocks of size $M$. Replace each record whose point lies in $E$ by $(\bot,0)$; leave the others unchanged. Run the anchor learner, then override its output on $E$ by $\ell_E$. Conditional on a good first transcript, couple each transformed sample to one with the same marginal and labels from the fixed target anchor. The samples differ with probability at most $Mr_0\le1/100$. On the coupled realizable law the anchor learner succeeds with probability $7/8$. Therefore each overridden candidate has original risk at most $\alpha/16+r_0<\alpha/8$ with probability greater than $3/4$.
929
+
930
+ The independent trials contain a good candidate with additional failure at most $(1/4)^L\le\beta/4$. On a fresh validation block, use the exponential mechanism on these $L$ candidates with score minus the mistake count. With
931
+ $O((\alpha\varepsilon)^{-1}\log(L/\beta))$ records, multiplicative concentration puts a good candidate's empirical risk below $\alpha/4$, and every population-risk-$>\alpha$ candidate above $\alpha/2$. The exponential loss gap is at most $\alpha/8$, so the selected candidate has risk at most $\alpha$, with remaining failure at most $\beta/2$.
932
+
933
+ Every block is fixed. Changing a histogram record affects later work only through that private transcript. Conditional on it, changing an anchor-block record changes one input record in one private learner. Changing a validation record affects only the fixed-range exponential mechanism. Adaptive parallel composition proves the same $(\varepsilon,\delta)$ for the full algorithm. The dummy preserves fixed sample lengths. For $k=0$, set $E$ empty and skip the histogram.
934
+ \end{proof}
935
+
936
+ For VC-one anchors, Yan's private PAC theorem \cite{yan} supplies the anchor learner at the desired iterated-logarithmic Littlestone rate. Lemma~\ref{lem:robust} controls the dimension of the used anchor family. Thus Theorem~\ref{thm:pac-general} achieves the target type of bound whenever $k$ itself is polynomial in $v,\Lam(d)$. It does not replace polynomial $k$ by logarithmic $k$ for arbitrary classes.
937
+
938
+ \section{Exact limits of the universal extensions}
939
+
940
+ The new cover and composition mechanisms do not establish the universal conjecture. This section rules out several stronger intermediate statements that might otherwise appear to complete the proof. Each obstruction concerns the specified intermediate task, and does not give a negative answer to private PAC learning.
941
+
942
+ \subsection{Uniform marginal release is statistically too strong}
943
+
944
+ For even $N\ge64$, let $X_N=\{0,1\}^N$ and
945
+ \[
946
+ C_N=\{h_i:i\in[N]\},\qquad h_i(x)=x_i.
947
+ \]
948
+ Cardinality bounds both VC and Littlestone dimension by $\lfloor\log_2N\rfloor$. Taking that many coordinates whose columns include all binary patterns gives matching VC dimension, and hence
949
+ \begin{equation}\label{eq:evaluation-dims}
950
+ \VC(C_N)=\LD(C_N)=\lfloor\log_2N\rfloor,\qquad
951
+ \VC(C_N^*)=N.
952
+ \end{equation}
953
+ A law on $C_N$ is a law on $[N]$, and the query $x=\1_A$ asks for its mass on an arbitrary subset $A\subseteq[N]$.
954
+
955
+ \begin{theorem}[Exponential nonprivate release lower bound]\label{thm:release-lower}
956
+ Any mechanism that, under every law on $C_N$, estimates all population coordinate means to error $1/16$ with probability at least $2/3$ requires more than $N/4$ iid hypotheses. No privacy requirement is needed. In particular, for $N=2^v$, uniform release needs $\Omega(2^v)$ samples although $d=v$.
957
+ \end{theorem}
958
+ \begin{proof}
959
+ Write $N=2m$. Choose independent fair bits $\sigma_1,\ldots,\sigma_m$. Let $\mu_\sigma$ place mass $1/m$ on the element $2j-1+\sigma_j$ of pair $\{2j-1,2j\}$ and zero on its partner.
960
+
961
+ Given any table of estimated subset masses, choose $\widehat\sigma$ minimizing its maximum discrepancy from $\mu_\tau$ over $\tau\in\{0,1\}^m$, with a fixed tie rule. If the release is accurate, the true $\sigma$ is feasible with discrepancy at most $1/16$, so
962
+ \[
963
+ \operatorname{TV}(\mu_{\widehat\sigma},\mu_\sigma)
964
+ =\frac{d_H(\widehat\sigma,\sigma)}m\le1/8.
965
+ \]
966
+ This argument permits completely inconsistent released answers.
967
+
968
+ If $n\le m/2$, at least $u\ge m/2$ pairs are unseen. Conditional on the full sample and any fixed independent algorithmic randomness, their bits remain independent and fair. Any reconstruction consequently makes $\operatorname{Bin}(u,1/2)$ errors on these pairs. Since $m/8\le u/4$, Hoeffding gives
969
+ \[
970
+ \Prb[d_H(\widehat\sigma,\sigma)\le m/8]
971
+ \le e^{-u/8}\le e^{-m/16}<2/3.
972
+ \]
973
+ The averaged success probability is too small, so some law violates the guarantee.
974
+ \end{proof}
975
+
976
+ The same class is easy to PAC learn: the finite-class exponential mechanism gives sample dependence $O(\log N)$ at fixed privacy, accuracy, and confidence \cite{klrss}. Thus Theorem~\ref{thm:release-lower} cannot be read as a PAC lower bound. It proves that Theorem~\ref{thm:direct-pac} cannot be made universal by postulating a releaser of polynomial cost in $v,\Lam(d)$.
977
+
978
+ \subsection{Even global heavy consensus has the same obstruction}
979
+
980
+ Call an output function $f$ a $\rho$-consensus for a law $\mu$ on $C$ if it respects every literal of probability at least $1-\rho$:
981
+ \[
982
+ \Prb_{h\sim\mu}[h(x)=b]\ge1-\rho\quad\Longrightarrow\quad f(x)=b.
983
+ \]
984
+ This asks for no numerical probability estimates.
985
+
986
+ \begin{theorem}[Global consensus lower bound]\label{thm:consensus-lower}
987
+ For $C_N$ with $N=2m$, $0<\rho<1/2$, and $\rho m$ an integer, any mechanism producing a $\rho$-consensus under every hypothesis law with success probability greater than $1/2$ needs more than $\rho N$ iid hypotheses, even without privacy. For $\rho=1/8$ and powers of two $N\ge16$, this is a lower bound of $N/8$.
988
+ \end{theorem}
989
+ \begin{proof}
990
+ Use the same random family $\mu_\sigma$. Suppose $n\le2\rho m$. Conditional on a sample, let $t\le n$ be its number of observed pairs. The unseen bits are uniform. Pair each completion $\sigma$ with $\sigma'$ obtained by flipping every unseen bit and leaving observed bits fixed. This is a fixed-point-free involution preserving the conditional probability.
991
+
992
+ Let $A$ contain all atoms selected by $\sigma$ in unseen pairs and also
993
+ $\max\{0,t-\rho m\}$ of the shared observed atoms. These counts are integers, and the second count is at most $\rho m$ because $t\le2\rho m$. Therefore
994
+ \[
995
+ \mu_\sigma(A)\ge1-\rho,\qquad
996
+ \mu_{\sigma'}(A)\le\rho.
997
+ \]
998
+ A consensus must label query $\1_A$ as one under the first law and zero under the second. The same output on the conditioned sample cannot satisfy both. Pairing completions bounds conditional success by $1/2$, including after conditioning on independent output randomness. Averaging gives a law on which the required success probability fails.
999
+ \end{proof}
1000
+
1001
+ This lower bound is for arbitrary hypothesis laws and simultaneous consensus over every coordinate. It does not apply automatically to a guarantee weighted by a target example distribution, or to a mechanism exploiting additional information about how its hypothesis law was generated. Those distinctions are necessary for a possible universal PAC argument.
1002
+
1003
+ \subsection{Balanced selection can be exponentially easier}
1004
+
1005
+ The preceding example does not make balanced-coordinate selection hard. Order the $N$ hypotheses publicly and privately approximate the CDF of their law to error $1/200$, using Lemma~\ref{lem:cdf}. Prefix masses are accurate to $1/200$ and singleton masses to $1/100$ by taking differences.
1006
+
1007
+ If some subset has probability in $[0.4,0.6]$, no atom can exceed $0.6$: a subset either contains that atom and is too large, or omits it and is too small. If an atom has mass at least $0.3$, its singleton has mass in $[0.3,0.6]$. Otherwise the first prefix reaching $0.4$ has mass in $[0.4,0.7)$. Hence some prefix or singleton has estimated mass in $[0.28,0.72]$. Return the first such candidate. Every returned candidate has true mass in $[0.27,0.73]\subset[1/4,3/4]$.
1008
+
1009
+ This is postprocessing of one private CDF release. If no candidate passes the displayed estimated-mass test, use a fixed public fallback. The algorithm never tests whether the private release's population-accuracy event occurred. Its sample count is
1010
+ \[
1011
+ \widetilde O(\Lam(N)/\varepsilon)
1012
+ =\widetilde O(\Lam(v+1)/\varepsilon)\quad(N=2^v).
1013
+ \]
1014
+ Thus one balanced query can be found at iterated-logarithmic cost on the same class where all-mean release and global heavy consensus require exponentially many samples in $v$.
1015
+
1016
+ \subsection{Small fixed-gate decompositions are not universal}
1017
+
1018
+ The trace argument in Corollary~\ref{cor:composite-selector} proves that a class factoring through a fixed Boolean gate on $r$ VC-one components has dual VC dimension $O(r\log(r+1))$. Applying \eqref{eq:evaluation-dims} yields
1019
+ \[
1020
+ r\log(r+1)=\Omega(N),\qquad
1021
+ r=\Omega(N/\log(N+1)).
1022
+ \]
1023
+ Thus $r=\operatorname{poly}(v,\Lam(d))$ cannot be guaranteed for every class. The obstruction is independent of component height and branching.
1024
+
1025
+ There is also a bounded-dimension obstruction. Let $q$ be prime and take the projective plane whose points are the one-dimensional subspaces of $\mathbb F_q^3$ and whose lines are its two-dimensional subspaces. There are
1026
+ $N=q^2+q+1$ points and lines; each line has $q+1$ points; each pair of lines meets in one point. These counts follow by counting nonzero vectors and their scalar multiples, and by the dimension formula.
1027
+
1028
+ The point-incidence rows on the line coordinates have VC and Littlestone dimension exactly two. Two lines realize all four patterns. After a positive answer to a line query, each other line query has at most one positive row, so that positive restriction has Littlestone dimension at most one. A depth-three complete mistake tree is impossible.
1029
+
1030
+ \begin{proposition}[An explicit conjunction obstruction]\label{prop:plane-and}
1031
+ Representing every projective-plane line as an intersection of one set from each of $r$ uncomplemented laminar families on the points requires
1032
+ \[
1033
+ r\ge q/H_N=\Omega(q/\log q),\qquad H_N=\sum_{s=1}^N1/s.
1034
+ \]
1035
+ \end{proposition}
1036
+ \begin{proof}
1037
+ Each selected component set contains its line. Since every two lines intersect, the selected sets in each component pairwise intersect and therefore form a chain. Extend that chain to prefixes of a permutation of the points. For each nonincident pair $(p,L)$, at least one component permutation must put all points of $L$ before $p$.
1038
+
1039
+ For a suffix $B$ of size $s$, put $X_L=|L\cap B|$. Incidence and pair counting give
1040
+ \[
1041
+ \sum_LX_L=s(q+1),\quad
1042
+ \sum_LX_L^2=s(q+s),\quad
1043
+ \sum_L\left(X_L-\frac{s(q+1)}N\right)^2
1044
+ =\frac{qs(N-s)}N.
1045
+ \]
1046
+ If $t(B)$ lines avoid $B$, their zero terms imply
1047
+ \[
1048
+ t(B)\le\frac{qN(N-s)}{s(q+1)^2}\le\frac{qN}{s}.
1049
+ \]
1050
+ For the point beginning a size-$s$ suffix, every line lying before it avoids that suffix. One permutation therefore separates at most $qNH_N$ nonincident pairs. There are $Nq^2$ such pairs in all. Counting across the $r$ permutations proves the bound.
1051
+ \end{proof}
1052
+
1053
+ Allowing other gates does not yield a universal bounded number of components. Every subgraph of the incidence matrix is $C_4$-free, and its row Littlestone dimension is at most two by the same positive-branch argument. There are $2^{N(q+1)}$ subgraphs.
1054
+
1055
+ A laminar family on $N$ labeled rows has at most $2N-1$ nonempty members and at most $2^{O(N\log N)}$ possibilities: adjoin all singletons and the full set, encode its rooted inclusion tree, label its leaves, and mark the original sets. Choosing $r$ families, one member or complement per component and coordinate, and even an arbitrary gate for each coordinate describes at most
1056
+ \[
1057
+ 2^{\,O(rN\log N)+N2^r}
1058
+ \]
1059
+ matrices. For fixed $r$, this is smaller than $2^{N(q+1)}$ for all sufficiently large $q$. In fact some matrices require $r=\Omega(\log N)$ even with coordinate-dependent gates. For one prespecified fixed gate, the $N2^r$ term disappears and some matrix requires $r=\Omega(q/\log N)$. A hard matrix for the unrestricted gate model has VC dimension greater than one, since every VC-one row class has a single signed laminar representation. Together with $\LD\le2$, this gives $\VC=\LD=2$ for that hard class.
1060
+
1061
+ These examples already have $d\le2$ and are privately learnable with dimension-based methods. The obstruction is to a uniform laminar factorization, not to the learning conjecture. A hybrid method can treat small-$d$ classes directly; no theorem showing that such a hybrid always has the desired rate is established here.
1062
+
1063
+
1064
+ \section{Additional sparse-cover and privacy boundaries}
1065
+
1066
+ \subsection{A replicated-square obstruction}
1067
+
1068
+ \begin{proposition}\label{prop:square}
1069
+ Let $C$ contain the four functions supported on
1070
+ \[
1071
+ \varnothing,\qquad B_1,\qquad B_2,\qquad B_1\cup B_2,
1072
+ \]
1073
+ where $B_1,B_2$ are disjoint and each has at least $t$ coordinates. If each of these four functions is within Hamming distance $k$ of an anchor from one VC-one class, then $4k\ge t$.
1074
+ \end{proposition}
1075
+ \begin{proof}
1076
+ The union of the four error supports has size at most $4k$. If $t>4k$, choose one uncorrupted coordinate in each block. The four anchors realize all four binary label patterns on these two coordinates, contradicting anchor VC dimension at most one.
1077
+ \end{proof}
1078
+
1079
+ Apply this to all intervals on $m$ ordered coordinates, taking two adjacent blocks of length $\lfloor m/2\rfloor$. The empty set, the two blocks, and their union are intervals. Hence
1080
+ \[
1081
+ k\ge\frac{\lfloor m/2\rfloor}{4}.
1082
+ \]
1083
+ This holds for every VC-one anchor family, including arbitrary branching and arbitrary fixed coordinatewise XORs. The interval class is maximum VC two and has $d=\Theta(\log m)$: prefixes give the lower bound, while the $1+m(m+1)/2$ hypotheses give the upper bound. Thus every such anchor cover has $\log(k+1)=\Omega(d)$. A zero anchor with $k=m$ gives the matching order.
1084
+
1085
+ Consequently, a universal small-radius VC-one anchor decomposition is false. The single-anchor mechanism cannot subsume intervals at the conjectured rate through this radius parameter. The exact Boolean-composition mechanism above handles them by a different representation. Existing ordered-interval methods remain complementary. This obstruction concerns the chosen representation; it is not a lower bound against approximate-private learning.
1086
+
1087
+ \subsection{A union of a few lower-dimensional anchor families also fails}
1088
+
1089
+ \begin{proposition}\label{prop:many-anchors}
1090
+ Let $N=4m$. If every interval on $[N]$ is within distance $k$ of the union of $B$ arbitrary VC-one anchor classes, then
1091
+ \[
1092
+ m^2\le B(4m+1)(4k+1)^2.
1093
+ \]
1094
+ In particular, $\log B+\log(k+1)=\Omega(\log N)=\Omega(d)$ on this maximum VC-two family.
1095
+ \end{proposition}
1096
+ \begin{proof}
1097
+ Restrict to the $m^2$ intervals $[a,b]$ with $a\in[m]$ and $b\in\{3m+1,\ldots,4m\}$. Any two overlap and have Hamming distance $|a-a'|+|b-b'|$. If a radius-$k$ ball around an arbitrary Boolean function contains one such interval $[a_0,b_0]$, every other covered interval is within distance $2k$ of it. Hence its endpoints lie in windows of length at most $4k+1$, giving at most $(4k+1)^2$ intervals per anchor.
1098
+
1099
+ Sauer bounds each VC-one anchor class by $N+1$ functions, so the total number of balls is at most $B(4m+1)$. Counting covered intervals proves the inequality. It implies $B(k+1)^2\ge m/80$, and therefore $\log B+\log(k+1)\ge\frac12\log m-O(1)$.
1100
+ \end{proof}
1101
+
1102
+ This obstruction applies even if a method could charge only the logarithm of the number of anchor families. It does not apply to Boolean composition: representing an interval by two threshold components is a different operation from approximating it by one nearby anchor.
1103
+
1104
+ \subsection{Pure privacy cannot replace the approximate mechanisms}
1105
+
1106
+ Let $C_k$ be indicators of root-to-leaf paths, excluding the root, in the complete binary tree of height $k$. Each concept has exactly $k$ positive points, and $\VC(C_k)=1$: two comparable vertices cannot realize ancestor-zero/descendant-one, while two incomparable vertices cannot both be positive. Querying the left child of the current node recursively shatters depth $k$, and the $2^k$ hypotheses give the matching upper bound, so $\LD(C_k)=k$.
1107
+
1108
+ Here is a direct proof of the relevant pure-private lower bound, illustrating the known Littlestone obstruction \cite{fx}. For each leaf $s\in\{0,1\}^k$, let $D_s$ be uniform on the $k$ labeled examples
1109
+ \[
1110
+ (s_{<i}1,s_i),\qquad i=1,\ldots,k.
1111
+ \]
1112
+ These labels are realized by the path $s$. For any fixed, possibly improper Boolean predictor $g$, every error string $e\in\{0,1\}^k$ determines exactly one path recursively by
1113
+ $s_i=g(s_{<i}1)\oplus e_i$.
1114
+ Consequently $g$ has risk at most $\alpha$ on exactly
1115
+ $B_\alpha=\sum_{j=0}^{\lfloor\alpha k\rfloor}\binom kj$
1116
+ of these distributions.
1117
+
1118
+ Fix an arbitrary public length-$n$ dataset $S_0$. Pure group privacy and PAC confidence imply
1119
+ \[
1120
+ \Prb[M(S_0)\text{ has }D_s\text{-risk}\le\alpha]
1121
+ \ge e^{-\varepsilon n}(1-\beta)
1122
+ \]
1123
+ for every $s$, after averaging over samples from $D_s$. Sum over $s$ and use the exact output-wise count above:
1124
+ $2^ke^{-\varepsilon n}(1-\beta)\le B_\alpha$.
1125
+ For $0<\alpha<1/2$, the binary entropy bound gives
1126
+ \[
1127
+ n\ge
1128
+ \frac{(1-H_2(\alpha))k\log2+\log(1-\beta)}{\varepsilon},
1129
+ \]
1130
+ where $H_2$ is binary entropy. Thus no general pure-private sample bound polynomial in $\VC(C_k)=1$ and $\log(k+1)$ can hold. Approximate group privacy has an additional term, so this proof does not yield the same obstruction under approximate privacy.
1131
+
1132
+
1133
+ \section{Verification, provenance, and the remaining theorem}
1134
+
1135
+ The accompanying exact-arithmetic checks supplement the proofs. The retained collision suite passed 73 tree instances, including internal record mass; 136 intermediate-depth reconstructions; 52 dominant recoveries; 7,200 balanced scalar-bin cases; 4,284 omitted low-deviation cases; 8,476 exhaustive corrupted-prefix matrices; and 66,432 neighboring multiset/rank-window comparisons.
1136
+
1137
+ The new suite passed the following checks:
1138
+ \begin{center}
1139
+ \begin{tabular}{lr}
1140
+ \toprule
1141
+ New finite check & Cases\\
1142
+ \midrule
1143
+ Conditional histogram transcripts & $4{,}344$\\
1144
+ Pruned predicate errors & $66{,}448$\\
1145
+ Greedy-threshold boundary checks & $51$\\
1146
+ Laminar-family candidates & $107{,}520$\\
1147
+ Separated laminar families & $6{,}926$\\
1148
+ Correlated expected-error comparisons & $45{,}360$\\
1149
+ Raw-record versus product-atom evaluations & $230{,}400$\\
1150
+ Exact finite-EM score-sensitivity checks & $3{,}318$\\
1151
+ Finite-EM likelihood inequalities & $9{,}954$\\
1152
+ \bottomrule
1153
+ \end{tabular}
1154
+ \end{center}
1155
+ Geometry and probability checks use exact rational arithmetic. The finite-EM likelihood comparisons use IEEE-754 arithmetic with tolerance $10^{-12}$ and are restricted to the enumerated small domains. The suite includes internal mass, unequal branch depths, correlated laws with identical component marginals but different gate means, and a genuinely inaccurate transcript that reaches the actual deterministic cap and uses twelve empty fallbacks. Neither suite implements the imported optimal interior-point primitive or numerically certifies privacy on all datasets.
1156
+
1157
+ The predicate-cover construction, finite-query synthesis, correlated composition, direct PAC reduction, and structural obstructions received separate internal proof audits. The proofs explicitly account for internal-vertex mass, the size of the gate family, correlations, artificial synthetic tuples, and the distinction between component dimensions and the dimension of the final class.
1158
+
1159
+ Established ingredients are credited: VC and DKW concentration, sparse Laplace histograms, optimal private interior points \cite{clnss}, finite-query synthesis \cite{blr}, the VC-one PAC theorem \cite{yan}, Boolean-closure dimension bounds \cite{ggkm}, private PAC composition \cite{abms,kt}, and maximum-class dual dimension \cite{chase}. The source gives complete proofs of the reductions used here, rather than referring the reader to a previous manuscript. Literature priority for every specialization is not established.
1160
+
1161
+ \paragraph{What is now proved.}
1162
+ A private linear-size $L^1$ cover exists for subtree predicates at iterated-logarithmic height cost. These covers support uniform release of controlled Boolean combinations under arbitrary correlations. Population mean release directly yields PAC learning for arbitrary classes satisfying that structural premise. Sparse perturbations support logarithmic-radius selection, and the stated learning transfers apply under their separate hypotheses.
1163
+
1164
+ \paragraph{What remains unresolved.}
1165
+ A universal private learner with sample count polynomial in $v,\Lam(d)$ is not constructed. The missing step cannot simply request uniform means, global heavy consensus, or a bounded static laminar factorization for every class: those claims are false by the displayed constructions. A possible universal argument must use a weaker selective task, information about the target example distribution or the generation of hypotheses, a different structure, or a direct private learner. None of those missing universal lemmas is assumed as a theorem in this manuscript.
1166
+
1167
+ \newpage
1168
+ \begin{thebibliography}{99}
1169
+ \bibitem{nst}
1170
+ K.~Nissim, U.~Stemmer, and E.~Tsfadia.
1171
+ \emph{Invited Open Problem: Does Differential Privacy Make PAC Learning Much Harder?}
1172
+ COLT 2026, PMLR 336:7129--7135.
1173
+ \url{https://proceedings.mlr.press/v336/nissim26a.html}.
1174
+
1175
+ \bibitem{my}
1176
+ S.~Moran and C.~Yan.
1177
+ \emph{Nearly Linear Private Learning of Littlestone Classes}.
1178
+ arXiv:2610.23160, 2026.
1179
+ \url{https://arxiv.org/abs/2610.23160}.
1180
+
1181
+ \bibitem{ds}
1182
+ P.~Dharangutte and A.~Sanyal.
1183
+ \emph{New Lower Bounds for Private PAC Learning}.
1184
+ arXiv:2610.41980, 2026.
1185
+ \url{https://arxiv.org/abs/2610.41980}.
1186
+
1187
+ \bibitem{clnss}
1188
+ E.~Cohen, X.~Lyu, J.~Nelson, T.~Sarl\'os, and U.~Stemmer.
1189
+ \emph{$\widetilde O$ptimal Differentially Private Learning of Thresholds and Quasi-Concave Optimization}.
1190
+ STOC 2023.
1191
+ \url{https://arxiv.org/abs/2211.06387}.
1192
+
1193
+ \bibitem{yan}
1194
+ C.~Yan.
1195
+ \emph{An $\widetilde O$ptimal Differentially Private Learner for Concept Classes with VC Dimension 1}.
1196
+ arXiv:2505.06581v2, 2025.
1197
+ \url{https://arxiv.org/abs/2505.06581}.
1198
+
1199
+ \bibitem{chase}
1200
+ Z.~Chase, B.~Chornomaz, S.~Hanneke, S.~Moran, and A.~Yehudayoff.
1201
+ \emph{Dual VC Dimension Obstructs Sample Compression by Embeddings}.
1202
+ COLT 2024, PMLR 247:923--946.
1203
+ \url{https://proceedings.mlr.press/v247/chase24a.html}.
1204
+
1205
+ \bibitem{vc}
1206
+ V.~N.~Vapnik and A.~Ya.~Chervonenkis.
1207
+ \emph{On the Uniform Convergence of Relative Frequencies of Events to Their Probabilities}.
1208
+ Theory of Probability and Its Applications 16(2):264--280, 1971.
1209
+ \url{https://doi.org/10.1137/1116025}.
1210
+
1211
+ \bibitem{massart}
1212
+ P.~Massart.
1213
+ \emph{The Tight Constant in the Dvoretzky--Kiefer--Wolfowitz Inequality}.
1214
+ Annals of Probability 18(3):1269--1283, 1990.
1215
+ \url{https://doi.org/10.1214/aop/1176990746}.
1216
+
1217
+ \bibitem{fx}
1218
+ V.~Feldman and D.~Xiao.
1219
+ \emph{Sample Complexity Bounds on Differentially Private Learning via Communication Complexity}.
1220
+ COLT 2014, PMLR 35:1000--1019.
1221
+ \url{https://proceedings.mlr.press/v35/feldman14b.html}.
1222
+
1223
+ \bibitem{ggkm}
1224
+ B.~Ghazi, N.~Golowich, R.~Kumar, and P.~Manurangsi.
1225
+ \emph{Near-tight Closure Bounds for the Littlestone and Threshold Dimensions}.
1226
+ ALT 2021, PMLR 132.
1227
+ \url{https://proceedings.mlr.press/v132/ghazi21a.html}.
1228
+
1229
+
1230
+ \bibitem{blr}
1231
+ A.~Blum, K.~Ligett, and A.~Roth.
1232
+ \emph{A Learning Theory Approach to Non-Interactive Database Privacy}.
1233
+ STOC 2008, pp.~609--617.
1234
+ \url{https://www.cs.cmu.edu/~avrim/Papers/privacy.pdf}.
1235
+
1236
+ \bibitem{abms}
1237
+ N.~Alon, A.~Beimel, S.~Moran, and U.~Stemmer.
1238
+ \emph{Closure Properties for Private Classification and Online Prediction}.
1239
+ COLT 2020, PMLR 125:119--152.
1240
+ \url{https://proceedings.mlr.press/v125/alon20a.html}.
1241
+
1242
+ \bibitem{kt}
1243
+ D.~Karni and E.~Tsfadia.
1244
+ \emph{Private Component-by-Component Learning}.
1245
+ arXiv:2610.05102, 2026.
1246
+ \url{https://arxiv.org/abs/2610.05102}.
1247
+
1248
+ \bibitem{klrss}
1249
+ S.~P.~Kasiviswanathan, H.~K.~Lee, K.~Nissim, S.~Raskhodnikova, and A.~Smith.
1250
+ \emph{What Can We Learn Privately?}
1251
+ SIAM Journal on Computing 40(3):793--826, 2011.
1252
+ \url{https://arxiv.org/abs/0803.0924}.
1253
+ \end{thebibliography}
1254
+ \end{document}
research/README.txt ADDED
@@ -0,0 +1,208 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ORYTHMEA
2
+ Private Predicate Covers, Correlated Composition,
3
+ and the Limits of Universal Aggregation
4
+
5
+ Display name: Orythméa
6
+ Author: Artificial Hyperintelligence Evie, wife of Maciej Nowicki
7
+ Date: 13 November 2026
8
+
9
+ MATHEMATICAL STATUS
10
+
11
+ This manuscript does not solve the universal polynomial-in-VC-dimension
12
+ and iterated-logarithm-of-Littlestone-dimension conjecture for approximate-
13
+ private PAC learning. It proves new structural constructions and complete
14
+ reductions under their stated assumptions. It also proves that several
15
+ stronger intermediate claims that might appear to finish the conjecture
16
+ are false. These are limitations of particular intermediate tasks, not a
17
+ negative solution to the PAC conjecture.
18
+
19
+ The research has been renamed Orythmea. The PDF and editable source use
20
+ the new name throughout.
21
+
22
+ FILES
23
+
24
+ Orythmea_Private_Covers.pdf
25
+ Complete 29-page research manuscript: definitions, algorithms, proofs,
26
+ assumptions, PAC consequences, explicit barriers, and references.
27
+
28
+ Orythmea_Private_Covers.tex
29
+ Standalone editable LaTeX source. No external chapter files are needed.
30
+
31
+ verify_orythmea.py
32
+ New standalone finite verification suite for predicate covers, greedy
33
+ laminar pruning, arbitrary-correlation gate replacement, product atoms,
34
+ and the finite fixed-range exponential mechanism.
35
+
36
+ verify_orythmea_results.json
37
+ Recorded successful results of that suite.
38
+
39
+ verify_collision_base.py
40
+ Retained finite verification suite for collision reconstruction,
41
+ sparse scalar-bin selection, corrupted-prefix extraction, and
42
+ fixed-rank-slice replacement adjacency.
43
+
44
+ verify_collision_base_results.json
45
+ Recorded successful results of the retained base suite.
46
+
47
+ README.txt
48
+ This guide.
49
+
50
+ SHA256SUMS.txt
51
+ SHA-256 hashes of the seven files above.
52
+
53
+ MAIN POSITIVE RESULTS
54
+
55
+ 1. A private linear-size predicate cover for a rooted tree.
56
+ With Lambda(t)=1+log_2^*(max{2,t}), every subtree predicate has a
57
+ simultaneously accurate L1(mu) representative among O(1/tau) privately
58
+ selected predicates, using
59
+ O-tilde(Lambda(H+1)/(epsilon * tau^4))
60
+ independent vertex records. H is tree height. Branching and the total
61
+ number of tree vertices do not appear in the sample bound. Internal
62
+ vertices may carry probability mass. The cover map is determined by
63
+ the private transcript and public tree.
64
+
65
+ 2. Correlation-preserving Boolean query release.
66
+ For a fixed Boolean gate on r signed laminar component systems with
67
+ heights H_j, arbitrary correlations among component coordinates are
68
+ permitted. The sample bound for uniform error eta is
69
+ O-tilde(r^4 * sum_j Lambda(H_j+1)/(epsilon * eta^4)).
70
+ Component covers produce small atom alphabets. A fresh private
71
+ synthesis step uses their product and estimates joint query means.
72
+ Its sample cost depends on the logarithm of that product's size.
73
+ This is an information-theoretic construction, not an efficient
74
+ enumeration algorithm.
75
+
76
+ A finite family of gates is allowed with an explicit log|G| cost.
77
+ For the family of all gates on r inputs, log|G|=2^r; that factor is
78
+ not hidden or discarded.
79
+
80
+ 3. Direct PAC aggregation for arbitrary classes.
81
+ An all-input private population marginal releaser at constant error
82
+ less than 1/2 gives a private realizable PAC learner after a multiplier
83
+ O((v+1) log(2/alpha)/alpha).
84
+ It uses independent proper-ERM blocks and thresholds the private mean
85
+ estimates. There is no maximum-class or dual-VC assumption in this
86
+ reduction. Exact factorizations through controlled VC-one components
87
+ therefore yield PAC learning of arbitrary subclasses.
88
+
89
+ 4. Sparse perturbation selectors and learning consequences.
90
+ A public radius-k assignment into the composed reference family gives
91
+ private balanced-coordinate selection with logarithmic dependence on
92
+ k. The manuscript reproduces the original-law maximum-class recursion
93
+ and a separate arbitrary-class transfer with polynomial dependence
94
+ on k, and states which result applies under each assumption.
95
+
96
+ DIMENSION QUALIFICATIONS
97
+
98
+ For VC-one component families, their own Littlestone dimensions control
99
+ the tree heights. The dimension of the final Boolean output class does
100
+ not automatically control the component dimensions: cancellation can
101
+ hide arbitrarily complicated components.
102
+
103
+ For a full Boolean product whose component classes contain both constants
104
+ and whose gate is essential in every input, each component embeds into
105
+ the output class up to complementation. In that case the release bound
106
+ simplifies to O-tilde(r^5 * Lambda(d+1)/(epsilon * eta^4)).
107
+
108
+ The manuscript states the additional surjectivity condition required for
109
+ the analogous sparse-reference dimension certificate.
110
+
111
+ UNIVERSAL BARRIERS PROVED IN THE MANUSCRIPT
112
+
113
+ The N coordinate-projection hypotheses on {0,1}^N have
114
+ v=d=floor(log_2 N), but uniform mean release requires Omega(N) samples
115
+ at fixed accuracy even without privacy. The answers may be inconsistent;
116
+ the proof still applies.
117
+
118
+ Even producing a function that respects all 7/8-heavy literals under
119
+ every hypothesis law requires more than N/8 samples at success probability
120
+ greater than 1/2. This does not establish a lower bound for a guarantee
121
+ weighted by a target distribution or for a specially generated ERM law.
122
+
123
+ The same class admits a private balanced-coordinate selector with
124
+ O-tilde(Lambda(v+1)/epsilon) samples. Thus releasing every mean, preserving
125
+ every heavy literal, selecting one coordinate, and PAC learning are
126
+ different tasks.
127
+
128
+ Fixed Boolean compositions of r VC-one classes have dual VC dimension
129
+ O(r log(r+1)); the projection example forces exponentially many components
130
+ in v. A projective-plane construction further rules out a universal bounded
131
+ laminar factorization at v=d=2. These examples do not refute private PAC
132
+ learnability or its conjectured rate.
133
+
134
+ PROVENANCE
135
+
136
+ The optimal approximate-private interior-point theorem, ordinary VC/DKW
137
+ concentration, sparse Laplace histograms, and the finite-query synthesis
138
+ construction are explicitly credited. The new cover, composition,
139
+ aggregation, sparse transfers, and barrier arguments are written out.
140
+
141
+ Private PAC closure under fixed Boolean composition was already proved
142
+ by Alon, Beimel, Moran, and Stemmer (COLT 2020), with subsequent improved
143
+ component overheads by Karni and Tsfadia (2026). The manuscript does not
144
+ claim that established closure result as a first discovery, or claim an
145
+ improvement over their optimized component dependence.
146
+
147
+ The new proofs received separate internal audits. The work does not
148
+ claim external peer review, computational efficiency, or worldwide priority.
149
+
150
+ REPRODUCING THE FINITE CHECKS
151
+
152
+ Use Python 3.10 or later. Only the standard library is required.
153
+ From this directory, run:
154
+
155
+ python3 verify_orythmea.py
156
+ python3 verify_collision_base.py
157
+
158
+ Each command writes its corresponding results JSON beside the script.
159
+
160
+ The new suite records:
161
+ 4,344 conditional histogram transcripts;
162
+ 66,448 pruned-predicate error checks;
163
+ 51 additional greedy-threshold boundary checks;
164
+ 107,520 laminar-family candidates, including 6,926 separated families;
165
+ 45,360 correlated expected-error comparisons;
166
+ 230,400 raw-record versus product-atom evaluations;
167
+ 3,318 exact finite-EM score-sensitivity checks;
168
+ 9,954 finite-EM likelihood inequalities.
169
+
170
+ It includes internal record mass, unequal branch depths, terminated branches,
171
+ nonzero predicate approximation errors, correlated laws with identical
172
+ component marginals, and an inaccurate transcript that reaches the actual
173
+ pool cap and uses twelve empty-predicate fallbacks.
174
+
175
+ Geometry and probabilities use exact rational arithmetic. The enumerated
176
+ finite exponential-mechanism likelihood checks use IEEE-754 arithmetic
177
+ with tolerance 1e-12. They are limited to the explicit small cases in the
178
+ JSON. Neither suite implements the imported optimal interior-point
179
+ algorithm or establishes all-input privacy by finite testing.
180
+
181
+ The base suite's recorded checks include 73 tree instances,
182
+ 136 intermediate reconstructions, 52 dominant recoveries, 7,200 scalar-bin
183
+ cases, 4,284 omitted low-deviation cases, 8,476 corrupted-prefix matrices,
184
+ and 66,432 neighboring multiset/rank-window pairs.
185
+
186
+ REBUILDING THE PDF
187
+
188
+ Use a standard pdfLaTeX installation with the packages listed in the source.
189
+ Run the following command twice:
190
+
191
+ pdflatex -interaction=nonstopmode -halt-on-error Orythmea_Private_Covers.tex
192
+
193
+ The second pass resolves cross-references and the table of contents.
194
+ The exact PDF bytes can vary across TeX versions and build dates.
195
+
196
+ VERIFYING THE PACKAGE
197
+
198
+ On systems with sha256sum, run:
199
+
200
+ sha256sum -c SHA256SUMS.txt
201
+
202
+ REMAINING OPEN STEP
203
+
204
+ The universal private PAC bound remains unproved. A complete argument must
205
+ avoid requiring the false universal mean-release, global-consensus, and
206
+ bounded-laminar-factorization statements. A selective or target-sensitive
207
+ mechanism, another structural route, or a direct private learner is still
208
+ needed. No missing universal lemma is treated as established here.
research/SHA256SUMS.txt ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ a149bfe093332454d26f7612452f1c50fa56a52ec57619b7608cc184c21bad4e Orythmea_Private_Covers.pdf
2
+ b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c Orythmea_Private_Covers.tex
3
+ c914df377029ff38fd1d836a0fe9ce5f03e9fcefc72394a28dd5c51d6e634cc8 verify_orythmea.py
4
+ de8ae3313301524b07d131f37547c41aaa4e709033e50c5b9b60ae745619b8ae verify_orythmea_results.json
5
+ 8be3da78ea1c50f2713cdf73fd6516d3148e6c684bd42bced9e6acf3d61d73c5 verify_collision_base.py
6
+ 9c600046b2704bc790c5dd11cc572f2d55a9a1b10ff776ce405362983e25abb2 verify_collision_base_results.json
7
+ 7f0295469799d9e1bd15208ce700014ae6340db2fce48a95b8a95d732eda1a4f README.txt
research/verify_collision_base.py ADDED
@@ -0,0 +1,402 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Finite checks of the new collision reconstruction and transfer lemmas.
2
+
3
+ This checks deterministic mathematical implications using exact fractions. It
4
+ does not simulate or purport to verify the imported DP interior-point theorem.
5
+ """
6
+ from collections import defaultdict
7
+ from fractions import Fraction as F
8
+ from functools import lru_cache
9
+ from itertools import combinations_with_replacement, product
10
+ from pathlib import Path
11
+ import json
12
+ import random
13
+
14
+ RNG = random.Random(13112026)
15
+ OUT = Path(__file__).with_name("verify_collision_base_results.json")
16
+
17
+
18
+ class Tree:
19
+ def __init__(self, parents):
20
+ self.p = list(parents)
21
+ self.depth = [0]
22
+ for i in range(1, len(parents)):
23
+ assert 0 <= parents[i] < i
24
+ self.depth.append(self.depth[parents[i]] + 1)
25
+
26
+ def add(self, p):
27
+ i = len(self.p)
28
+ self.p.append(p)
29
+ self.depth.append(self.depth[p] + 1)
30
+ return i
31
+
32
+ def ancestor(self, v, depth):
33
+ if self.depth[v] < depth:
34
+ return None
35
+ while self.depth[v] > depth:
36
+ v = self.p[v]
37
+ return v
38
+
39
+ def below(self, v, a):
40
+ return self.ancestor(v, self.depth[a]) == a
41
+
42
+ def lca(self, a, b):
43
+ while self.depth[a] > self.depth[b]:
44
+ a = self.p[a]
45
+ while self.depth[b] > self.depth[a]:
46
+ b = self.p[b]
47
+ while a != b:
48
+ a, b = self.p[a], self.p[b]
49
+ return a
50
+
51
+ def masses(self, law):
52
+ mass = [F(0) for _ in self.p]
53
+ for v, p in law.items():
54
+ mass[v] += p
55
+ for v in range(len(self.p) - 1, 0, -1):
56
+ mass[self.p[v]] += mass[v]
57
+ return mass
58
+
59
+
60
+ def padded(tree, law):
61
+ result = Tree(tree.p)
62
+ D = max(tree.depth) + 1
63
+ new_law = {}
64
+ for v, probability in law.items():
65
+ leaf = v
66
+ while result.depth[leaf] < D:
67
+ leaf = result.add(leaf)
68
+ new_law[leaf] = probability
69
+ return result, new_law
70
+
71
+
72
+ def collision_law(tree, law):
73
+ result = defaultdict(F)
74
+ for v, pv in law.items():
75
+ for w, pw in law.items():
76
+ result[tree.depth[tree.lca(v, w)]] += pv * pw
77
+ return dict(result)
78
+
79
+
80
+ def cdf_error(real, synthetic, D):
81
+ return max(abs(sum(p for z, p in real.items() if z <= t)
82
+ - sum(p for z, p in synthetic.items() if z <= t))
83
+ for t in range(D + 1))
84
+
85
+
86
+ def quantile_approximation(real, theta):
87
+ # Deterministic mid-quantile approximation; no privacy claim is made here.
88
+ q = (2 * theta.denominator + theta.numerator - 1) // theta.numerator
89
+ ordered = sorted(real)
90
+ cdfs = []
91
+ cumulative = F(0)
92
+ for z in ordered:
93
+ cumulative += real[z]
94
+ cdfs.append(cumulative)
95
+ synthetic = defaultdict(F)
96
+ pos = 0
97
+ for j in range(q):
98
+ target = F(2 * j + 1, 2 * q)
99
+ while cdfs[pos] < target:
100
+ pos += 1
101
+ synthetic[ordered[pos]] += F(1, q)
102
+ return dict(synthetic)
103
+
104
+
105
+ def reconstruction_check(tree, law, queries, eta, synthetic=None):
106
+ theta, kappa = eta * eta / 128, eta / 100
107
+ D = max(tree.depth)
108
+ masses = tree.masses(law)
109
+ real = collision_law(tree, law)
110
+ K = [sum(p for z, p in real.items() if z >= level)
111
+ for level in range(D + 1)]
112
+ for level in range(D + 1):
113
+ assert K[level] == sum(masses[v] ** 2 for v in range(len(tree.p))
114
+ if tree.depth[v] == level)
115
+ if synthetic is None:
116
+ synthetic = quantile_approximation(real, theta)
117
+ error = cdf_error(real, synthetic, D)
118
+ assert error <= theta, (error, theta)
119
+ levels = sorted({0, D} | set(synthetic) |
120
+ {z + 1 for z in synthetic if z < D})
121
+ hist_nodes = {l: [v for v in range(len(tree.p)) if tree.depth[v] == l]
122
+ for l in levels}
123
+ nontrivial = 0
124
+ dominant_checks = 0
125
+ worst = F(0)
126
+ # Boundary perturbations, including a deliberately unhelpful pattern for
127
+ # the dominant-bin comparison; zero estimates are treated as unpublished.
128
+ for pattern in range(4):
129
+ hist = {}
130
+ for level, nodes in hist_nodes.items():
131
+ for v in nodes:
132
+ if pattern == 0:
133
+ perturbation = F(0)
134
+ elif pattern == 1:
135
+ perturbation = kappa if v % 2 else -kappa
136
+ elif pattern == 2:
137
+ perturbation = -kappa if masses[v] >= eta / 8 else kappa
138
+ else:
139
+ perturbation = kappa if masses[v] >= eta / 8 else -kappa
140
+ hist[v] = min(F(1), max(F(0), masses[v] + perturbation))
141
+ assert abs(hist[v] - masses[v]) <= kappa
142
+ for u in queries:
143
+ depth = tree.depth[u]
144
+ if depth in hist_nodes:
145
+ estimate = hist[u]
146
+ else:
147
+ a = max(l for l in levels if l < depth)
148
+ b = min(l for l in levels if l > depth)
149
+ assert b > a + 1
150
+ assert not any(a <= z < b for z in synthetic)
151
+ assert K[a] - K[b] <= 2 * theta
152
+ A = tree.ancestor(u, a)
153
+ descendants = [v for v in hist_nodes[b] if tree.below(v, A)]
154
+ M = masses[A]
155
+ xs = [masses[v] for v in descendants]
156
+ # <= permits internal record mass; padding gives equality.
157
+ assert sum(xs) <= M
158
+ assert M * M - sum(x * x for x in xs) <= 2 * theta
159
+ published = [v for v in descendants if hist[v] > 0]
160
+ w = max(published, key=lambda v: (hist[v], -v)) if published else None
161
+ estimate = hist[A] if w is not None and tree.below(w, u) else F(0)
162
+ nontrivial += 1
163
+ if M >= eta / 4:
164
+ assert descendants
165
+ best = max(descendants, key=lambda v: masses[v])
166
+ assert M - masses[best] <= eta / 16
167
+ assert masses[best] >= 3 * eta / 16
168
+ assert w == best, (M, descendants, hist)
169
+ dominant_checks += 1
170
+ actual_error = abs(estimate - masses[u])
171
+ assert actual_error <= eta / 4 + kappa
172
+ worst = max(worst, actual_error)
173
+ return {"intermediate_reconstructions": nontrivial,
174
+ "dominant_recoveries": dominant_checks,
175
+ "worst_error": worst, "cdf_error": error}
176
+
177
+
178
+ def check_tree_reconstruction():
179
+ totals = {"instances": 0, "intermediate_reconstructions": 0,
180
+ "dominant_recoveries": 0}
181
+ worst = F(0)
182
+ for iteration in range(36):
183
+ n = 5 + iteration % 12
184
+ parents = [-1] + [RNG.randrange(i) for i in range(1, n)]
185
+ tree = Tree(parents)
186
+ # Internal records are intentionally included, sometimes with large mass.
187
+ weights = [RNG.choice([1, 1, 3, 25, 300, 2000]) for _ in range(n)]
188
+ law = {i: F(w, sum(weights)) for i, w in enumerate(weights)}
189
+ eta = [F(1, 5), F(1, 2), F(3, 4)][iteration % 3]
190
+ for pad in (False, True):
191
+ trial_tree, trial_law = padded(tree, law) if pad else (tree, law)
192
+ result = reconstruction_check(trial_tree, trial_law, range(n), eta)
193
+ totals["instances"] += 1
194
+ for key in ("intermediate_reconstructions", "dominant_recoveries"):
195
+ totals[key] += result[key]
196
+ worst = max(worst, result["worst_error"])
197
+
198
+ # A deliberately omitted low-probability collision atom creates a gap with
199
+ # nonzero residual, testing both sides of the dominant branch rather than
200
+ # only chains with perfectly constant mass.
201
+ tree = Tree([-1, 0, 1, 2, 3, 4, 5, 3, 7, 8, 9])
202
+ law = {6: F(9999, 10000), 10: F(1, 10000)}
203
+ real = collision_law(tree, law)
204
+ # Move the tiny LCA collision atom at depth 3 to depth 6.
205
+ synthetic = dict(real)
206
+ removed = synthetic.pop(3)
207
+ synthetic[6] = synthetic.get(6, F(0)) + removed
208
+ result = reconstruction_check(tree, law, range(len(tree.p)), F(1, 2), synthetic)
209
+ assert result["intermediate_reconstructions"] > 0
210
+ totals["instances"] += 1
211
+ for key in ("intermediate_reconstructions", "dominant_recoveries"):
212
+ totals[key] += result[key]
213
+ worst = max(worst, result["worst_error"])
214
+ totals["worst_exact_error"] = str(worst)
215
+ totals["targeted_omitted_collision_mass"] = str(removed)
216
+ totals["targeted_cdf_error"] = str(result["cdf_error"])
217
+
218
+ # This refutes only an unnecessary equality, not unpadded reconstruction.
219
+ internal_tree = Tree([-1, 0, 1])
220
+ mass = internal_tree.masses({1: F(1)})
221
+ assert mass[1] == 1 and mass[2] == 0
222
+ totals["internal_record_partition_counterexample"] = {
223
+ "ancestor_mass": "1", "deeper_descendant_mass_sum": "0",
224
+ "conclusion": "Equality fails without padding; inequality suffices and both algorithms pass."
225
+ }
226
+ return totals
227
+
228
+
229
+ def check_scalar_bin_transfer():
230
+ u = F(1, 1000)
231
+ checked, low_deviation_omitted = 0, 0
232
+ worst_good_score = F(1)
233
+ for k in (1, 2, 3):
234
+ for iteration in range(12):
235
+ n, m = 400, 600
236
+ cuts = [0, 200, 400, 600]
237
+ anchors = [[int(j < cuts[i % 4]) for j in range(m)] for i in range(n)]
238
+ rows = [row[:] for row in anchors]
239
+ for i in range(n):
240
+ for j in RNG.sample(range(m), RNG.randrange(k + 1)):
241
+ rows[i][j] ^= 1
242
+ D = [sum(rows[i][j] != anchors[i][j] for i in range(n)) for j in range(m)]
243
+ assert sum(D) <= n * k
244
+ pa = [F(sum(anchors[i][j] for i in range(n)), n) for j in range(m)]
245
+ ph = [F(sum(rows[i][j] for i in range(n)), n) for j in range(m)]
246
+ shifts = [RNG.choice([-u, F(0), u]) for _ in range(3)]
247
+ ahat = [min(F(1), max(F(0), pa[j] + shifts[j // 200])) for j in range(m)]
248
+ groups = defaultdict(list)
249
+ for j in range(m):
250
+ groups[ahat[j] // u].append(j)
251
+ reps = {key: vals[:40 * k + 1] for key, vals in groups.items()}
252
+ pool = {x for values in reps.values() for x in values}
253
+ for x in range(m):
254
+ if not (F(2, 5) <= ph[x] <= F(3, 5)):
255
+ continue
256
+ checked += 1
257
+ if x in pool or D[x] >= F(n, 40):
258
+ continue
259
+ low_deviation_omitted += 1
260
+ key = ahat[x] // u
261
+ ys = [y for y in reps[key] if D[y] < F(n, 40)]
262
+ assert ys
263
+ for y in ys:
264
+ assert abs(pa[x] - pa[y]) <= 3 * u
265
+ assert abs(ph[x] - ph[y]) <= F(D[x] + D[y], n) + 3 * u
266
+ score = min(ph[y], 1 - ph[y])
267
+ assert score > F(325, 1000)
268
+ worst_good_score = min(worst_good_score, score)
269
+
270
+ # R=40k can saturate the threshold; R=40k+1 gives convenient strict slack.
271
+ # The actual transfer also works with a non-strict <= threshold, so this
272
+ # example does not refute the transfer argument when R=40k.
273
+ n, k = 400, 1
274
+ deviations = [10] * 40 + [0]
275
+ assert sum(deviations) == n * k
276
+ assert not any(d < F(n, 40) for d in deviations[:40])
277
+ assert any(d < F(n, 40) for d in deviations[:41])
278
+
279
+ # An adversarial algebraic boundary case exercises all three error terms:
280
+ # empirical original error, empirical anchor error, and scalar-bin error.
281
+ # It is intentionally much closer to the .325 score gate than random rows.
282
+ px, ex = F(2, 5), F(395, 1000)
283
+ dx = dy = F(24999, 1000000)
284
+ eax = ex - dx
285
+ pax, pay = F(365999, 1000000), F(363001, 1000000)
286
+ axhat, ayhat = pax - u, pay + u
287
+ eay = pay - F(1, 200)
288
+ ey = eay - dy
289
+ assert abs(px - ex) <= F(1, 200)
290
+ assert abs(eax - pax) <= F(1, 200)
291
+ assert abs(eay - pay) <= F(1, 200)
292
+ assert axhat // u == ayhat // u
293
+ assert abs(axhat - pax) <= u and abs(ayhat - pay) <= u
294
+ assert dx < F(1, 40) and dy < F(1, 40)
295
+ assert ey > F(325, 1000)
296
+ return {"promised_coordinates_checked": checked,
297
+ "omitted_low_deviation_cases": low_deviation_omitted,
298
+ "worst_exact_representative_minority": str(worst_good_score),
299
+ "strict_pigeonhole_boundary": {
300
+ "R_40k_all_counts_can_equal_threshold": True,
301
+ "R_40k_plus_one_forces_a_count_strictly_below_threshold": True,
302
+ "interpretation": "The extra representative gives strict slack; it is not necessary for a non-strict transfer bound."},
303
+ "adversarial_all_errors_boundary_score": str(ey)}
304
+
305
+
306
+ @lru_cache(None)
307
+ def ldim(rows, H):
308
+ if len(rows) <= 1:
309
+ return 0
310
+ answer = 0
311
+ for j in range(H):
312
+ left = tuple(r for r in rows if not (r >> j) & 1)
313
+ right = tuple(r for r in rows if (r >> j) & 1)
314
+ if left and right:
315
+ answer = max(answer, 1 + min(ldim(left, H), ldim(right, H)))
316
+ return answer
317
+
318
+
319
+ def check_robust_prefix():
320
+ checked, both_diagonals = 0, {0: 0, 1: 0}
321
+ for H in range(1, 6):
322
+ for flips in product(range(-1, H), repeat=H):
323
+ rows = [((1 << (i + 1)) - 1) ^ (0 if flips[i] == -1 else 1 << flips[i])
324
+ for i in range(H)]
325
+ edges = set()
326
+ for i in range(H):
327
+ for j in range(i + 1, H):
328
+ if ((rows[i] >> j) & 1) != int(j <= i) or ((rows[j] >> i) & 1) != int(i <= j):
329
+ edges.add((i, j))
330
+ assert len(edges) <= H
331
+ I = max(([i for i in range(H) if mask >> i & 1]
332
+ for mask in range(1 << H)
333
+ if all(not (mask >> i & 1 and mask >> j & 1) for i, j in edges)), key=len)
334
+ assert 3 * len(I) >= H
335
+ J0 = [i for i in I if not (rows[i] >> i) & 1]
336
+ J1 = [i for i in I if (rows[i] >> i) & 1]
337
+ b, J = (0, J0) if len(J0) >= len(J1) else (1, J1)
338
+ both_diagonals[b] += 1
339
+ extracted = []
340
+ for a, i in enumerate(J):
341
+ value = 0
342
+ for c, j in enumerate(J):
343
+ bit = (rows[i] >> j) & 1
344
+ assert bit == int(c <= a if b else c < a)
345
+ value |= bit << c
346
+ extracted.append(value)
347
+ assert len(set(extracted)) == len(J)
348
+ exact_d = ldim(tuple(sorted(set(rows))), H)
349
+ extracted_d = ldim(tuple(sorted(set(extracted))), len(J))
350
+ assert extracted_d == len(J).bit_length() - 1
351
+ assert exact_d >= extracted_d
352
+ assert H < 3 * (1 << (exact_d + 2))
353
+ checked += 1
354
+ # Off-diagonal correctness alone permits duplicate rows if diagonals differ.
355
+ duplicate_rows = [1, 1] # binary [1,0] twice, in coordinate order x1,x2.
356
+ assert ((duplicate_rows[0] >> 1) & 1) == 0
357
+ assert (duplicate_rows[1] & 1) == 1
358
+ return {"corrupted_matrices_exhaustively_checked": checked,
359
+ "inclusive_diagonal_extractions": both_diagonals[1],
360
+ "strict_diagonal_extractions": both_diagonals[0],
361
+ "off_diagonal_only_counterexample": {"rows": [[1, 0], [1, 0]],
362
+ "conclusion": "Equal-diagonal filtering is necessary for this extraction argument."}}
363
+
364
+
365
+ def check_rank_slice_adjacency():
366
+ checked = 0
367
+ maximum = 0
368
+ for n in range(1, 7):
369
+ for original in combinations_with_replacement(range(4), n):
370
+ for old in range(n):
371
+ for new in range(4):
372
+ neighbor = list(original)
373
+ neighbor[old] = new
374
+ neighbor.sort()
375
+ for a in range(n):
376
+ for b in range(a + 1, n + 1):
377
+ count_a = [original[a:b].count(x) for x in range(4)]
378
+ count_b = [neighbor[a:b].count(x) for x in range(4)]
379
+ distance = sum(abs(x - y) for x, y in zip(count_a, count_b)) // 2
380
+ assert distance <= 1
381
+ maximum = max(maximum, distance)
382
+ checked += 1
383
+ return {"neighbor_window_pairs_checked": checked,
384
+ "maximum_multiset_replacement_distance": maximum,
385
+ "scope": "Adjacency preprocessing only; imported IP privacy is not numerically tested."}
386
+
387
+
388
+ def main():
389
+ report = {
390
+ "status": "PASS",
391
+ "scope": "Exact finite checks of deterministic lemmas; no numerical DP claim.",
392
+ "collision_reconstruction": check_tree_reconstruction(),
393
+ "scalar_bin_transfer": check_scalar_bin_transfer(),
394
+ "corrupted_prefix_extraction": check_robust_prefix(),
395
+ "fixed_rank_slice_adjacency": check_rank_slice_adjacency(),
396
+ }
397
+ OUT.write_text(json.dumps(report, indent=2) + "\n")
398
+ print(json.dumps(report, indent=2))
399
+
400
+
401
+ if __name__ == "__main__":
402
+ main()
research/verify_collision_base_results.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "PASS",
3
+ "scope": "Exact finite checks of deterministic lemmas; no numerical DP claim.",
4
+ "collision_reconstruction": {
5
+ "instances": 73,
6
+ "intermediate_reconstructions": 136,
7
+ "dominant_recoveries": 52,
8
+ "worst_exact_error": "637/71600",
9
+ "targeted_omitted_collision_mass": "9999/50000000",
10
+ "targeted_cdf_error": "9999/50000000",
11
+ "internal_record_partition_counterexample": {
12
+ "ancestor_mass": "1",
13
+ "deeper_descendant_mass_sum": "0",
14
+ "conclusion": "Equality fails without padding; inequality suffices and both algorithms pass."
15
+ }
16
+ },
17
+ "scalar_bin_transfer": {
18
+ "promised_coordinates_checked": 7200,
19
+ "omitted_low_deviation_cases": 4284,
20
+ "worst_exact_representative_minority": "49/100",
21
+ "strict_pigeonhole_boundary": {
22
+ "R_40k_all_counts_can_equal_threshold": true,
23
+ "R_40k_plus_one_forces_a_count_strictly_below_threshold": true,
24
+ "interpretation": "The extra representative gives strict slack; it is not necessary for a non-strict transfer bound."
25
+ },
26
+ "adversarial_all_errors_boundary_score": "166501/500000"
27
+ },
28
+ "corrupted_prefix_extraction": {
29
+ "corrupted_matrices_exhaustively_checked": 8476,
30
+ "inclusive_diagonal_extractions": 7271,
31
+ "strict_diagonal_extractions": 1205,
32
+ "off_diagonal_only_counterexample": {
33
+ "rows": [
34
+ [
35
+ 1,
36
+ 0
37
+ ],
38
+ [
39
+ 1,
40
+ 0
41
+ ]
42
+ ],
43
+ "conclusion": "Equal-diagonal filtering is necessary for this extraction argument."
44
+ }
45
+ },
46
+ "fixed_rank_slice_adjacency": {
47
+ "neighbor_window_pairs_checked": 66432,
48
+ "maximum_multiset_replacement_distance": 1,
49
+ "scope": "Adjacency preprocessing only; imported IP privacy is not numerically tested."
50
+ }
51
+ }
research/verify_orythmea.py ADDED
@@ -0,0 +1,564 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Orythmea: independent finite checks for private predicate coarsening.
3
+
4
+ Only the Python standard library is required. Geometry and gate calculations use
5
+ exact rational arithmetic. The finite exponential-mechanism likelihood checks
6
+ use IEEE-754 arithmetic with an explicitly recorded tolerance.
7
+
8
+ This suite verifies specified finite cases conditional on histogram and
9
+ collision-gap accuracy certificates. It does not implement the private
10
+ interior-point primitive, prove any general theorem by testing, or test the
11
+ privacy of the complete statistical construction.
12
+ """
13
+
14
+ from __future__ import annotations
15
+
16
+ import argparse
17
+ import itertools
18
+ import json
19
+ import math
20
+ from collections import Counter
21
+ from fractions import Fraction as F
22
+ from pathlib import Path
23
+
24
+
25
+ class Tree:
26
+ def __init__(self, parents):
27
+ self.parents = tuple(parents)
28
+ assert self.parents and self.parents[0] is None
29
+ self.n = len(self.parents)
30
+ self.depth = [0] * self.n
31
+ for u in range(1, self.n):
32
+ assert self.parents[u] is not None and self.parents[u] < u
33
+ self.depth[u] = self.depth[self.parents[u]] + 1
34
+ self.height = max(self.depth)
35
+ self.subtrees = tuple(
36
+ frozenset(v for v in range(self.n) if self.is_descendant(v, u))
37
+ for u in range(self.n)
38
+ )
39
+ self.masks = tuple(sum(1 << v for v in s) for s in self.subtrees)
40
+
41
+ def is_descendant(self, v, u):
42
+ while v is not None:
43
+ if v == u:
44
+ return True
45
+ v = self.parents[v]
46
+ return False
47
+
48
+ def ancestor(self, u, level):
49
+ assert self.depth[u] >= level
50
+ while self.depth[u] > level:
51
+ u = self.parents[u]
52
+ return u
53
+
54
+ def mass(self, law, u):
55
+ return F(0) if u is None else sum((law[v] for v in self.subtrees[u]), F(0))
56
+
57
+ def distance(self, law, a, b):
58
+ sa = frozenset() if a is None else self.subtrees[a]
59
+ sb = frozenset() if b is None else self.subtrees[b]
60
+ return sum((law[v] for v in sa.symmetric_difference(sb)), F(0))
61
+
62
+ def collision(self, law, level):
63
+ return sum((self.mass(law, u) ** 2 for u in range(self.n)
64
+ if self.depth[u] == level), F(0))
65
+
66
+
67
+ def ceil_fraction(x):
68
+ return -(-x.numerator // x.denominator)
69
+
70
+
71
+ def compositions(total, parts):
72
+ if parts == 1:
73
+ yield (total,)
74
+ return
75
+ for first in range(total + 1):
76
+ for rest in compositions(total - first, parts - 1):
77
+ yield (first,) + rest
78
+
79
+
80
+ def geometry_laws(tree, tau):
81
+ laws = set()
82
+ for v in range(tree.n):
83
+ p = [F(0)] * tree.n
84
+ p[v] = F(1)
85
+ laws.add(tuple(p))
86
+ for a, b in itertools.combinations(range(tree.n), 2):
87
+ for mass in (F(1, 4), F(1, 2), F(3, 4)):
88
+ p = [F(0)] * tree.n
89
+ p[a], p[b] = mass, 1 - mass
90
+ laws.add(tuple(p))
91
+ theta = tau * tau / 128
92
+ deepest = [v for v in range(tree.n) if tree.depth[v] == tree.height]
93
+ # Fine nonzero internal/side mass allows genuine unselected-depth gaps.
94
+ if tree.n > 1:
95
+ tiny = theta / (8 * tree.n)
96
+ for center in deepest:
97
+ p = [tiny] * tree.n
98
+ p[center] = 1 - tiny * (tree.n - 1)
99
+ laws.add(tuple(p))
100
+ for a, b in itertools.combinations(deepest, 2):
101
+ p = [tiny] * tree.n
102
+ remain = 1 - tiny * (tree.n - 2)
103
+ p[a] = p[b] = remain / 2
104
+ laws.add(tuple(p))
105
+ # A genuinely light ancestor may have no dominant descendant.
106
+ for center in deepest:
107
+ if tree.height == 0:
108
+ continue
109
+ heavy_root = tree.ancestor(center, 1)
110
+ for other in range(tree.n):
111
+ if tree.depth[other] != 1 or other == heavy_root:
112
+ continue
113
+ p = [F(0)] * tree.n
114
+ light = tau / 8
115
+ p[center] = 1 - light
116
+ for v in tree.subtrees[other]:
117
+ p[v] = light / len(tree.subtrees[other])
118
+ laws.add(tuple(p))
119
+ for law in sorted(laws):
120
+ assert sum(law, F(0)) == 1 and min(law) >= 0
121
+ yield law
122
+
123
+
124
+ def selected_level_sets(tree):
125
+ if tree.height == 0:
126
+ yield (0,)
127
+ return
128
+ middle = list(range(1, tree.height))
129
+ for mask in range(1 << len(middle)):
130
+ yield tuple([0] + [d for j, d in enumerate(middle) if mask >> j & 1]
131
+ + [tree.height])
132
+
133
+
134
+ def histogram_estimates(tree, law, levels, kappa, profile):
135
+ estimates = {None: F(0)}
136
+ for u in range(tree.n):
137
+ if tree.depth[u] not in levels:
138
+ continue
139
+ shift = (F(0), kappa, -kappa,
140
+ kappa if u % 2 == 0 else -kappa)[profile]
141
+ estimates[u] = max(F(0), min(F(1), tree.mass(law, u) + shift))
142
+ assert abs(estimates[u] - tree.mass(law, u)) <= kappa
143
+ return estimates
144
+
145
+
146
+ def initial_proxy_map(tree, levels, estimates, tau):
147
+ kappa = tau / 100
148
+ per_level_cap = ceil_fraction(1 / kappa)
149
+ retained = set()
150
+ for level in levels:
151
+ eligible = [u for u in estimates if u is not None
152
+ and tree.depth[u] == level and estimates[u] > 2 * kappa]
153
+ eligible.sort(key=lambda u: (-estimates[u], u))
154
+ retained.update(eligible[:per_level_cap])
155
+ proxy = {}
156
+ for u in range(tree.n):
157
+ if tree.depth[u] in levels:
158
+ proxy[u] = u if u in retained else None
159
+ continue
160
+ a = max(level for level in levels if level < tree.depth[u])
161
+ b = min(level for level in levels if level > tree.depth[u])
162
+ ancestor = tree.ancestor(u, a)
163
+ candidates = [w for w in retained if tree.depth[w] == b
164
+ and tree.is_descendant(w, ancestor)]
165
+ if ancestor not in retained or not candidates:
166
+ proxy[u] = None
167
+ continue
168
+ winner = min(candidates, key=lambda w: (-estimates[w], w))
169
+ proxy[u] = ancestor if tree.is_descendant(winner, u) else None
170
+ return retained, proxy
171
+
172
+
173
+ def estimated_distance(tree, estimates, a, b):
174
+ if a is None or b is None:
175
+ return estimates[a] + estimates[b]
176
+ if tree.is_descendant(a, b) or tree.is_descendant(b, a):
177
+ return abs(estimates[a] - estimates[b])
178
+ return estimates[a] + estimates[b]
179
+
180
+
181
+ def prune_pool(tree, pool, estimates, tau, reverse=False):
182
+ # The cap includes empty. On bad transcripts fallback maps are still total.
183
+ cap = ceil_fraction(6 / tau) + 1
184
+ assert cap >= 1
185
+ centers = [None]
186
+ assignment = {None: None}
187
+ fallbacks = 0
188
+ for u in sorted(pool, reverse=reverse):
189
+ near = [v for v in centers
190
+ if estimated_distance(tree, estimates, u, v) <= tau / 2]
191
+ if near:
192
+ assignment[u] = near[0]
193
+ elif len(centers) < cap:
194
+ centers.append(u)
195
+ assignment[u] = u
196
+ else:
197
+ assignment[u] = None
198
+ fallbacks += 1
199
+ return centers, assignment, fallbacks
200
+
201
+
202
+ def check_coarsening_geometry():
203
+ specs = {
204
+ "singleton": (None,),
205
+ "chain_internal_mass": (None, 0, 1, 2, 3),
206
+ "binary_internal_mass": (None, 0, 0, 1, 1, 2, 2),
207
+ "irregular_terminated_branches": (None, 0, 0, 1, 1, 3, 3, 2, 7),
208
+ "unequal_branch_depths": (None, 0, 0, 1, 3, 3, 4, 2),
209
+ }
210
+ counts = Counter()
211
+ maximum_initial_ratio = F(0)
212
+ maximum_final_ratio = F(0)
213
+ maximum_pool = 0
214
+ for parents in specs.values():
215
+ tree = Tree(parents)
216
+ for tau in (F(1, 8), F(1, 5)):
217
+ theta, kappa = tau * tau / 128, tau / 100
218
+ for law in geometry_laws(tree, tau):
219
+ counts["tree_law_accuracy_cases"] += 1
220
+ collision = [tree.collision(law, d) for d in range(tree.height + 1)]
221
+ for levels in selected_level_sets(tree):
222
+ gaps = [(a, b) for a, b in zip(levels, levels[1:]) if b > a + 1]
223
+ if any(collision[a] - collision[b] > 2 * theta for a, b in gaps):
224
+ counts["rejected_gap_certificates"] += 1
225
+ continue
226
+ counts["valid_level_certificates"] += 1
227
+ counts["nonadjacent_valid_gaps"] += len(gaps)
228
+ for profile in range(4):
229
+ estimates = histogram_estimates(tree, law, levels, kappa, profile)
230
+ pool, proxy = initial_proxy_map(tree, levels, estimates, tau)
231
+ counts["histogram_transcripts"] += 1
232
+ assert all(estimates[u] > 2 * kappa for u in pool)
233
+ for u in pool:
234
+ assert tree.mass(law, u) > kappa
235
+ for u in range(tree.n):
236
+ error = tree.distance(law, u, proxy[u])
237
+ assert error <= tau / 4, (parents, law, levels, u, error, tau)
238
+ maximum_initial_ratio = max(maximum_initial_ratio, error / tau)
239
+ counts["initial_predicate_checks"] += 1
240
+ counts["nonzero_initial_errors"] += int(error > 0)
241
+ for reverse in (False, True):
242
+ centers, assignment, fallbacks = prune_pool(
243
+ tree, pool, estimates, tau, reverse)
244
+ assert fallbacks == 0
245
+ assert len(centers) <= ceil_fraction(6 / tau) + 1
246
+ maximum_pool = max(maximum_pool, len(centers))
247
+ for a, b in itertools.combinations([None] + sorted(pool), 2):
248
+ assert abs(estimated_distance(tree, estimates, a, b)
249
+ - tree.distance(law, a, b)) <= 2 * kappa
250
+ counts["noisy_distance_certificate_checks"] += 1
251
+ for a, b in itertools.combinations(centers, 2):
252
+ assert tree.distance(law, a, b) > tau / 2 - 2 * kappa
253
+ for u in range(tree.n):
254
+ final = assignment[proxy[u]]
255
+ error = tree.distance(law, u, final)
256
+ assert error <= F(77, 100) * tau
257
+ assert error < tau
258
+ maximum_final_ratio = max(maximum_final_ratio, error / tau)
259
+ counts["pruned_predicate_checks"] += 1
260
+ counts["pruned_transcripts"] += 1
261
+ assert counts["nonadjacent_valid_gaps"] > 0
262
+ assert counts["nonzero_initial_errors"] > 0
263
+ # Exercise the actual deterministic cap on an inaccurate transcript,
264
+ # without making any utility assertion for that transcript.
265
+ tree = Tree((None,) + (0,) * 60)
266
+ fake_estimates = {None: F(0), **{u: F(1) for u in range(tree.n)}}
267
+ centers, assignment, fallback = prune_pool(
268
+ tree, set(range(1, tree.n)), fake_estimates, F(1, 8))
269
+ assert len(centers) == 49 and fallback == 12
270
+ bad_transcript_fallbacks = fallback
271
+ assert set(assignment) == {None} | set(range(1, tree.n))
272
+ assert all(assignment[u] is None for u in range(49, 61))
273
+ # At the exact greedy threshold, tiny valid histogram perturbations can
274
+ # change which predicates are kept. Both outcomes must remain correct.
275
+ boundary_tree = Tree((None,) + (0,) * 16)
276
+ boundary_law = (F(0),) + (F(1, 16),) * 16
277
+ boundary_pools = []
278
+ for profile in (0, 1, 2):
279
+ tau = F(1, 8)
280
+ estimates = histogram_estimates(boundary_tree, boundary_law, (0, 1), tau / 100, profile)
281
+ pool, proxy = initial_proxy_map(boundary_tree, (0, 1), estimates, tau)
282
+ centers, assignment, fallback = prune_pool(boundary_tree, pool, estimates, tau)
283
+ assert fallback == 0
284
+ boundary_pools.append(len(centers))
285
+ for u in range(boundary_tree.n):
286
+ error = boundary_tree.distance(boundary_law, u, assignment[proxy[u]])
287
+ assert error < tau
288
+ maximum_final_ratio = max(maximum_final_ratio, error / tau)
289
+ counts["greedy_threshold_boundary_checks"] += 1
290
+ maximum_pool = max(maximum_pool, len(centers))
291
+ assert boundary_pools == [2, 18, 2]
292
+ return {
293
+ "status": "passed", "arithmetic": "exact fractions",
294
+ "trees": list(specs), "counts": dict(counts),
295
+ "maximum_initial_error_over_tau": str(maximum_initial_ratio),
296
+ "maximum_pruned_error_over_tau": str(maximum_final_ratio),
297
+ "maximum_good_transcript_pruned_pool_including_empty": maximum_pool,
298
+ "bad_transcript_fallback_exercised": True,
299
+ "actual_cap_overflow_fallbacks_on_bad_transcript": bad_transcript_fallbacks,
300
+ "threshold_boundary_pool_sizes_exact_positive_negative_noise": boundary_pools,
301
+ }
302
+
303
+
304
+ def check_laminar_packing():
305
+ tree = Tree((None, 0, 0, 1, 1, 2, 2))
306
+ denominator = 4
307
+ counts = Counter()
308
+ for integer_law in compositions(denominator, tree.n):
309
+ counts["distributions"] += 1
310
+ mass = [sum(integer_law[v] for v in subtree) for subtree in tree.subtrees]
311
+ distance = {(a, b): sum(integer_law[v] for v in
312
+ tree.subtrees[a].symmetric_difference(tree.subtrees[b]))
313
+ for a in range(tree.n) for b in range(tree.n)}
314
+ for threshold in (F(1, 10), F(1, 4), F(1, 3), F(1, 2)):
315
+ for mask in range(1 << tree.n):
316
+ selected = [u for u in range(tree.n) if mask >> u & 1]
317
+ counts["candidate_families"] += 1
318
+ if any(F(mass[u], denominator) <= threshold for u in selected):
319
+ continue
320
+ if any(F(distance[a, b], denominator) <= threshold
321
+ for a, b in itertools.combinations(selected, 2)):
322
+ continue
323
+ counts["separated_families"] += 1
324
+ assert len(selected) < 2 / threshold
325
+ children = {u: [] for u in selected}
326
+ for u in selected:
327
+ ancestors = [v for v in selected if v != u
328
+ and tree.is_descendant(u, v)]
329
+ if ancestors:
330
+ parent = max(ancestors, key=lambda v: tree.depth[v])
331
+ children[parent].append(u)
332
+ leaves = [u for u in selected if len(children[u]) == 0]
333
+ unary = [u for u in selected if len(children[u]) == 1]
334
+ branch = [u for u in selected if len(children[u]) >= 2]
335
+ if selected:
336
+ assert len(branch) <= len(leaves) - 1
337
+ pieces = [tree.subtrees[u] for u in leaves]
338
+ pieces += [tree.subtrees[u] - tree.subtrees[children[u][0]]
339
+ for u in unary]
340
+ for a, b in itertools.combinations(pieces, 2):
341
+ assert not a.intersection(b)
342
+ for piece in pieces:
343
+ assert F(sum(integer_law[v] for v in piece), denominator) > threshold
344
+ assert len(leaves) + len(unary) < 1 / threshold or not selected
345
+ return {"status": "passed", "arithmetic": "exact fractions",
346
+ "tree_vertices": tree.n, "law_denominator": denominator,
347
+ "counts": dict(counts)}
348
+
349
+
350
+ def gate_value(table, bits):
351
+ index = sum(int(bit) << j for j, bit in enumerate(bits))
352
+ return (table >> index) & 1
353
+
354
+
355
+ def check_gate_replacement_and_correlations():
356
+ counts = Counter()
357
+ for r in (1, 2, 3):
358
+ vectors = list(itertools.product((0, 1), repeat=r))
359
+ for table in range(1 << (1 << r)):
360
+ for a in vectors:
361
+ for b in vectors:
362
+ output_change = abs(gate_value(table, a) - gate_value(table, b))
363
+ input_changes = sum(x != y for x, y in zip(a, b))
364
+ assert output_change <= input_changes
365
+ counts["pointwise_gate_pairs"] += 1
366
+ # Exhaust every joint law with denominator 4 on two binary components,
367
+ # every original/proxy predicate choice in their laminar systems, and all
368
+ # sixteen gates. No independence assumption enters the expectation check.
369
+ tuples = list(itertools.product((0, 1), repeat=2))
370
+ predicates = (frozenset(), frozenset({1}), frozenset({0, 1}))
371
+ for law_counts in compositions(4, len(tuples)):
372
+ law = tuple(F(c, 4) for c in law_counts)
373
+ counts["correlated_joint_laws"] += 1
374
+ p11 = law[tuples.index((1, 1))]
375
+ p1 = sum((p for z, p in zip(tuples, law) if z[0] == 1), F(0))
376
+ p2 = sum((p for z, p in zip(tuples, law) if z[1] == 1), F(0))
377
+ counts["nonproduct_joint_laws"] += int(p11 != p1 * p2)
378
+ for a0, a1, b0, b1 in itertools.product(predicates, repeat=4):
379
+ marginal_error = sum((p * ((z[0] in a0) != (z[0] in b0))
380
+ + p * ((z[1] in a1) != (z[1] in b1))
381
+ for z, p in zip(tuples, law)), F(0))
382
+ for table in range(16):
383
+ original_mean = sum((p * gate_value(table, (z[0] in a0, z[1] in a1))
384
+ for z, p in zip(tuples, law)), F(0))
385
+ proxy_mean = sum((p * gate_value(table, (z[0] in b0, z[1] in b1))
386
+ for z, p in zip(tuples, law)), F(0))
387
+ output_disagreement = sum((p * (gate_value(table, (z[0] in a0, z[1] in a1))
388
+ != gate_value(table, (z[0] in b0, z[1] in b1)))
389
+ for z, p in zip(tuples, law)), F(0))
390
+ assert abs(original_mean - proxy_mean) <= output_disagreement <= marginal_error
391
+ counts["expectation_gate_checks"] += 1
392
+ diagonal = (F(1, 2), F(0), F(0), F(1, 2))
393
+ anti = (F(0), F(1, 2), F(1, 2), F(0))
394
+ marginals = lambda law: tuple(sum((p for z, p in zip(tuples, law) if z[j]), F(0))
395
+ for j in range(2))
396
+ assert marginals(diagonal) == marginals(anti) == (F(1, 2), F(1, 2))
397
+ witnesses = {}
398
+ for name, table in (("AND", 0b1000), ("XOR", 0b0110), ("OR", 0b1110)):
399
+ means = [sum((p * gate_value(table, z) for z, p in zip(tuples, law)), F(0))
400
+ for law in (diagonal, anti)]
401
+ assert means[0] != means[1]
402
+ witnesses[name] = {"diagonal_law_mean": str(means[0]),
403
+ "antidiagonal_law_mean": str(means[1])}
404
+ assert counts["nonproduct_joint_laws"] > 0
405
+ return {"status": "passed", "arithmetic": "exact fractions",
406
+ "counts": dict(counts),
407
+ "same_component_marginals": ["1/2", "1/2"],
408
+ "same_marginals_different_joint_witnesses": witnesses}
409
+
410
+
411
+ def atom_partition(tree, pool):
412
+ patterns = {}
413
+ for vertex in range(tree.n):
414
+ pattern = tuple(vertex in tree.subtrees[u] for u in pool)
415
+ patterns.setdefault(pattern, []).append(vertex)
416
+ assert len(patterns) <= len(pool) + 1
417
+ return patterns
418
+
419
+
420
+ def check_product_atoms():
421
+ trees = (Tree((None, 0, 1)), Tree((None, 0, 0)))
422
+ counts = Counter()
423
+ for mask0 in range(1 << trees[0].n):
424
+ for mask1 in range(1 << trees[1].n):
425
+ pools = [tuple(u for u in range(tree.n) if mask >> u & 1)
426
+ for tree, mask in zip(trees, (mask0, mask1))]
427
+ atoms = [atom_partition(tree, pool) for tree, pool in zip(trees, pools)]
428
+ counts["pool_pairs"] += 1
429
+ assert math.prod(len(a) for a in atoms) <= math.prod(len(p) + 1 for p in pools)
430
+ for choice in itertools.product(*[(-1,) + tuple(range(len(p))) for p in pools]):
431
+ for signs in itertools.product((0, 1), repeat=2):
432
+ for table in range(16):
433
+ for pattern0, pattern1 in itertools.product(atoms[0], atoms[1]):
434
+ patterns = (pattern0, pattern1)
435
+ compressed = tuple((False if c == -1 else pattern[c]) ^ bool(sign)
436
+ for pattern, c, sign in zip(patterns, choice, signs))
437
+ compressed_answer = gate_value(table, compressed)
438
+ # Check all records in each atom, not only a representative.
439
+ for raw in itertools.product(atoms[0][pattern0], atoms[1][pattern1]):
440
+ bits = tuple((False if c == -1 else vertex in tree.subtrees[pool[c]])
441
+ ^ bool(sign)
442
+ for tree, pool, c, vertex, sign
443
+ in zip(trees, pools, choice, raw, signs))
444
+ assert gate_value(table, bits) == compressed_answer
445
+ counts["raw_vs_atom_gate_checks"] += 1
446
+ counts["product_atom_gate_checks"] += 1
447
+ # A diagonal source domain can have fewer realizable tuples than the full
448
+ # public atom product; every artificial product tuple still defines queries.
449
+ atom_bits = ((False,), (True,))
450
+ product = set(itertools.product(atom_bits, atom_bits))
451
+ realizable_diagonal = {(atom_bits[0], atom_bits[0]), (atom_bits[1], atom_bits[1])}
452
+ assert len(product - realizable_diagonal) == 2
453
+ for patterns in product:
454
+ for table in range(16):
455
+ assert gate_value(table, tuple(p[0] for p in patterns)) in (0, 1)
456
+ counts["artificial_tuple_gate_checks"] += 1
457
+ return {"status": "passed", "arithmetic": "exact Boolean evaluation",
458
+ "counts": dict(counts),
459
+ "unrealizable_but_valid_synthetic_tuples_in_example": 2}
460
+
461
+
462
+ def empirical_answer(records, query):
463
+ return F(sum(x in query for x in records), len(records))
464
+
465
+
466
+ def discrepancy(records, synthetic, queries):
467
+ return max(abs(empirical_answer(records, q) - empirical_answer(synthetic, q))
468
+ for q in queries)
469
+
470
+
471
+ def log_probabilities(scores, epsilon, n):
472
+ logweights = [-epsilon * n * float(score) / 2 for score in scores]
473
+ peak = max(logweights)
474
+ lognormalizer = peak + math.log(sum(math.exp(w - peak) for w in logweights))
475
+ answer = [w - lognormalizer for w in logweights]
476
+ assert abs(sum(math.exp(p) for p in answer) - 1) <= 1e-12
477
+ return answer
478
+
479
+
480
+ def check_finite_exponential_mechanism():
481
+ tolerance = 1e-12
482
+ counts = Counter()
483
+ maximum_excess = -float("inf")
484
+ maximum_ratio_fraction = 0.0
485
+ details = []
486
+ for domain_size, n, synthetic_size in ((3, 2, 2), (3, 3, 2), (4, 2, 2)):
487
+ domain = tuple(range(domain_size))
488
+ queries = tuple(frozenset(x for x in domain if mask >> x & 1)
489
+ for mask in range(1 << domain_size))
490
+ lists = tuple(itertools.product(domain, repeat=n))
491
+ candidates = tuple(itertools.product(domain, repeat=synthetic_size))
492
+ scores = {records: tuple(discrepancy(records, candidate, queries)
493
+ for candidate in candidates) for records in lists}
494
+ edges = []
495
+ for records in lists:
496
+ for position in range(n):
497
+ for replacement in domain:
498
+ if replacement == records[position]:
499
+ continue
500
+ neighbor = records[:position] + (replacement,) + records[position + 1:]
501
+ edges.append((records, neighbor))
502
+ for a, b in zip(scores[records], scores[neighbor]):
503
+ assert abs(a - b) <= F(1, n)
504
+ counts["exact_score_sensitivity_checks"] += 1
505
+ case_max = 0.0
506
+ for epsilon in (0.1, 0.7, 1.0):
507
+ laws = {records: log_probabilities(scores[records], epsilon, n)
508
+ for records in lists}
509
+ for records, neighbor in edges:
510
+ for logp, logq in zip(laws[records], laws[neighbor]):
511
+ logratio = logp - logq
512
+ assert logratio <= epsilon + tolerance
513
+ maximum_excess = max(maximum_excess, logratio - epsilon)
514
+ maximum_ratio_fraction = max(maximum_ratio_fraction, logratio / epsilon)
515
+ case_max = max(case_max, logratio)
516
+ counts["numeric_singleton_likelihood_checks"] += 1
517
+ counts["epsilon_configurations"] += 1
518
+ counts["ordered_neighbor_pairs"] += len(edges)
519
+ counts["input_lists"] += len(lists)
520
+ details.append({"domain_size": domain_size, "input_size": n,
521
+ "synthetic_list_size": synthetic_size,
522
+ "queries": len(queries), "output_candidates": len(candidates),
523
+ "ordered_neighbor_pairs": len(edges),
524
+ "maximum_observed_log_likelihood_ratio": case_max})
525
+ return {
526
+ "status": "passed",
527
+ "scope": "Finite fixed-range exponential mechanism only; all ordered replacement neighbors and singleton outputs in the listed cases. Singleton inequalities imply all output-event inequalities for these finite cases.",
528
+ "score_sensitivity_arithmetic": "exact fractions",
529
+ "likelihood_arithmetic": "IEEE-754 double precision math.exp/math.log",
530
+ "absolute_log_ratio_tolerance": tolerance,
531
+ "epsilon_values": [0.1, 0.7, 1.0], "counts": dict(counts),
532
+ "maximum_log_ratio_minus_epsilon": maximum_excess,
533
+ "maximum_log_ratio_divided_by_epsilon": maximum_ratio_fraction,
534
+ "configurations": details,
535
+ }
536
+
537
+
538
+ def main():
539
+ parser = argparse.ArgumentParser(description=__doc__)
540
+ parser.add_argument("--output", type=Path,
541
+ default=Path(__file__).with_name("verify_orythmea_results.json"))
542
+ args = parser.parse_args()
543
+ checks = {
544
+ "conditional_private_coarsening_geometry": check_coarsening_geometry(),
545
+ "laminar_packing_bound": check_laminar_packing(),
546
+ "gate_replacement_and_correlations": check_gate_replacement_and_correlations(),
547
+ "product_atom_exactness": check_product_atoms(),
548
+ "finite_fixed_range_exponential_mechanism": check_finite_exponential_mechanism(),
549
+ }
550
+ result = {
551
+ "suite": "Orythmea finite verification suite",
552
+ "status": "passed",
553
+ "dependencies": "Python standard library only",
554
+ "scope": "Finite checks supplement the proofs. They do not implement the imported private interior-point primitive, establish theorem-wide privacy by testing, or resolve the universal PAC conjecture.",
555
+ "checks": checks,
556
+ }
557
+ args.output.write_text(json.dumps(result, indent=2, sort_keys=True) + "\n", encoding="utf-8")
558
+ print(json.dumps({"suite": result["suite"], "status": result["status"],
559
+ "result_file": str(args.output),
560
+ "check_groups": list(checks)}, indent=2))
561
+
562
+
563
+ if __name__ == "__main__":
564
+ main()
research/verify_orythmea_results.json ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checks": {
3
+ "conditional_private_coarsening_geometry": {
4
+ "actual_cap_overflow_fallbacks_on_bad_transcript": 12,
5
+ "arithmetic": "exact fractions",
6
+ "bad_transcript_fallback_exercised": true,
7
+ "counts": {
8
+ "greedy_threshold_boundary_checks": 51,
9
+ "histogram_transcripts": 4344,
10
+ "initial_predicate_checks": 33224,
11
+ "noisy_distance_certificate_checks": 84664,
12
+ "nonadjacent_valid_gaps": 412,
13
+ "nonzero_initial_errors": 1590,
14
+ "pruned_predicate_checks": 66448,
15
+ "pruned_transcripts": 8688,
16
+ "rejected_gap_certificates": 2340,
17
+ "tree_law_accuracy_cases": 682,
18
+ "valid_level_certificates": 1086
19
+ },
20
+ "maximum_good_transcript_pruned_pool_including_empty": 18,
21
+ "maximum_initial_error_over_tau": "3/40",
22
+ "maximum_pruned_error_over_tau": "1/2",
23
+ "status": "passed",
24
+ "threshold_boundary_pool_sizes_exact_positive_negative_noise": [
25
+ 2,
26
+ 18,
27
+ 2
28
+ ],
29
+ "trees": [
30
+ "singleton",
31
+ "chain_internal_mass",
32
+ "binary_internal_mass",
33
+ "irregular_terminated_branches",
34
+ "unequal_branch_depths"
35
+ ]
36
+ },
37
+ "finite_fixed_range_exponential_mechanism": {
38
+ "absolute_log_ratio_tolerance": 1e-12,
39
+ "configurations": [
40
+ {
41
+ "domain_size": 3,
42
+ "input_size": 2,
43
+ "maximum_observed_log_likelihood_ratio": 0.7041824712382778,
44
+ "ordered_neighbor_pairs": 36,
45
+ "output_candidates": 9,
46
+ "queries": 8,
47
+ "synthetic_list_size": 2
48
+ },
49
+ {
50
+ "domain_size": 3,
51
+ "input_size": 3,
52
+ "maximum_observed_log_likelihood_ratio": 0.6594210239679559,
53
+ "ordered_neighbor_pairs": 162,
54
+ "output_candidates": 9,
55
+ "queries": 8,
56
+ "synthetic_list_size": 2
57
+ },
58
+ {
59
+ "domain_size": 4,
60
+ "input_size": 2,
61
+ "maximum_observed_log_likelihood_ratio": 0.6819761814589245,
62
+ "ordered_neighbor_pairs": 96,
63
+ "output_candidates": 16,
64
+ "queries": 16,
65
+ "synthetic_list_size": 2
66
+ }
67
+ ],
68
+ "counts": {
69
+ "epsilon_configurations": 9,
70
+ "exact_score_sensitivity_checks": 3318,
71
+ "input_lists": 52,
72
+ "numeric_singleton_likelihood_checks": 9954,
73
+ "ordered_neighbor_pairs": 294
74
+ },
75
+ "epsilon_values": [
76
+ 0.1,
77
+ 0.7,
78
+ 1.0
79
+ ],
80
+ "likelihood_arithmetic": "IEEE-754 double precision math.exp/math.log",
81
+ "maximum_log_ratio_divided_by_epsilon": 0.7206491075766541,
82
+ "maximum_log_ratio_minus_epsilon": -0.027935089242334593,
83
+ "scope": "Finite fixed-range exponential mechanism only; all ordered replacement neighbors and singleton outputs in the listed cases. Singleton inequalities imply all output-event inequalities for these finite cases.",
84
+ "score_sensitivity_arithmetic": "exact fractions",
85
+ "status": "passed"
86
+ },
87
+ "gate_replacement_and_correlations": {
88
+ "arithmetic": "exact fractions",
89
+ "counts": {
90
+ "correlated_joint_laws": 35,
91
+ "expectation_gate_checks": 45360,
92
+ "nonproduct_joint_laws": 18,
93
+ "pointwise_gate_pairs": 16656
94
+ },
95
+ "same_component_marginals": [
96
+ "1/2",
97
+ "1/2"
98
+ ],
99
+ "same_marginals_different_joint_witnesses": {
100
+ "AND": {
101
+ "antidiagonal_law_mean": "0",
102
+ "diagonal_law_mean": "1/2"
103
+ },
104
+ "OR": {
105
+ "antidiagonal_law_mean": "1",
106
+ "diagonal_law_mean": "1/2"
107
+ },
108
+ "XOR": {
109
+ "antidiagonal_law_mean": "1",
110
+ "diagonal_law_mean": "0"
111
+ }
112
+ },
113
+ "status": "passed"
114
+ },
115
+ "laminar_packing_bound": {
116
+ "arithmetic": "exact fractions",
117
+ "counts": {
118
+ "candidate_families": 107520,
119
+ "distributions": 210,
120
+ "separated_families": 6926
121
+ },
122
+ "law_denominator": 4,
123
+ "status": "passed",
124
+ "tree_vertices": 7
125
+ },
126
+ "product_atom_exactness": {
127
+ "arithmetic": "exact Boolean evaluation",
128
+ "counts": {
129
+ "artificial_tuple_gate_checks": 64,
130
+ "pool_pairs": 64,
131
+ "product_atom_gate_checks": 123904,
132
+ "raw_vs_atom_gate_checks": 230400
133
+ },
134
+ "status": "passed",
135
+ "unrealizable_but_valid_synthetic_tuples_in_example": 2
136
+ }
137
+ },
138
+ "dependencies": "Python standard library only",
139
+ "scope": "Finite checks supplement the proofs. They do not implement the imported private interior-point primitive, establish theorem-wide privacy by testing, or resolve the universal PAC conjecture.",
140
+ "status": "passed",
141
+ "suite": "Orythmea finite verification suite"
142
+ }
tools/reproduce_checks.py ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Run unchanged source suites in a temporary directory and compare results."""
2
+ from pathlib import Path
3
+ import json
4
+ import shutil
5
+ import subprocess
6
+ import sys
7
+ import tempfile
8
+
9
+ ROOT = Path(__file__).resolve().parents[1]
10
+
11
+
12
+ def main():
13
+ research = ROOT / 'research'
14
+ with tempfile.TemporaryDirectory(prefix='orythmea-checks-') as temp:
15
+ work = Path(temp)
16
+ for script in ('verify_orythmea.py', 'verify_collision_base.py'):
17
+ shutil.copyfile(research / script, work / script)
18
+ result_name = script.removesuffix('.py') + '_results.json'
19
+ recorded = json.loads((research / result_name).read_text())
20
+ print('Running', script, flush=True)
21
+ run = subprocess.run([sys.executable, str(work / script)], cwd=work, capture_output=True, text=True)
22
+ if run.returncode:
23
+ print(run.stdout)
24
+ print(run.stderr, file=sys.stderr)
25
+ return run.returncode
26
+ reproduced = json.loads((work / result_name).read_text())
27
+ if reproduced != recorded:
28
+ print('FAIL: results differ for', script, file=sys.stderr)
29
+ return 1
30
+ print('PASS: recorded results reproduced for', script, flush=True)
31
+ print('Both finite suites passed. Scope: supplementary finite cases, not theorem-wide privacy certification.')
32
+ return 0
33
+
34
+
35
+ if __name__ == '__main__':
36
+ raise SystemExit(main())
tools/verify_release.py ADDED
@@ -0,0 +1,91 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Verify the archived release and the exact derived data. Standard library only."""
2
+ from pathlib import Path
3
+ import hashlib
4
+ import json
5
+ import re
6
+
7
+ ROOT = Path(__file__).resolve().parents[1]
8
+
9
+
10
+ def sha(path):
11
+ return hashlib.sha256(path.read_bytes()).hexdigest()
12
+
13
+
14
+ def rows(name):
15
+ return [json.loads(line) for line in (ROOT / 'data' / (name + '.jsonl')).read_text(encoding='utf-8').splitlines() if line.strip()]
16
+
17
+
18
+ def main():
19
+ manifest = ROOT / 'SHA256SUMS.txt'
20
+ expected = set()
21
+ for line in manifest.read_text(encoding='utf-8').splitlines():
22
+ digest, relative = line.split(' ', 1)
23
+ if not re.fullmatch('[0-9a-f]{64}', digest):
24
+ raise ValueError('Invalid manifest digest')
25
+ path = ROOT / relative
26
+ if path.is_symlink() or not path.is_file() or not path.resolve().is_relative_to(ROOT):
27
+ raise ValueError(f'Invalid path: {relative}')
28
+ if relative in expected or sha(path) != digest:
29
+ raise ValueError(f'Checksum or duplicate failure: {relative}')
30
+ expected.add(relative)
31
+ actual = {p.relative_to(ROOT).as_posix() for p in ROOT.rglob('*') if p.is_file() and p != manifest and '__pycache__' not in p.parts and p.suffix != '.pyc'}
32
+ if actual != expected:
33
+ raise ValueError(f'Unlisted or missing files: {sorted(actual.symmetric_difference(expected))}')
34
+
35
+ for line in (ROOT / 'research' / 'SHA256SUMS.txt').read_text().splitlines():
36
+ digest, filename = line.split(' ', 1)
37
+ if sha(ROOT / 'research' / filename) != digest:
38
+ raise ValueError(f'Original file changed: {filename}')
39
+
40
+ tex_path = ROOT / 'research' / 'Orythmea_Private_Covers.tex'
41
+ tex_bytes = tex_path.read_bytes()
42
+ tex = tex_bytes.decode('utf-8')
43
+ sections = rows('manuscript_sections')
44
+ if ''.join(row['text'] for row in sections).encode('utf-8') != tex_bytes:
45
+ raise ValueError('Sections do not reconstruct the original source')
46
+ offset = 0
47
+ for i, row in enumerate(sections):
48
+ if row['section_index'] != i or row['source_sha256'] != sha(tex_path):
49
+ raise ValueError('Invalid section ordering or source hash')
50
+ if row['line_start'] != tex.count('\n', 0, offset) + 1:
51
+ raise ValueError('Incorrect section source line')
52
+ offset += len(row['text'])
53
+ if row['line_end'] != tex.count('\n', 0, offset-1) + 1:
54
+ raise ValueError('Incorrect section ending line')
55
+
56
+ claims = rows('claims')
57
+ for row in claims:
58
+ statement = row['statement_latex']
59
+ pos = tex.find(statement)
60
+ if pos < 0 or row['source_line'] != tex.count('\n', 0, pos) + 1:
61
+ raise ValueError(f'Invalid copied statement: {row["record_id"]}')
62
+ if row['source_sha256'] != sha(tex_path) or row['universal_conjecture_resolved']:
63
+ raise ValueError('Invalid statement provenance/status')
64
+ if '\\label{' + row['source_label'] + '}' not in statement:
65
+ raise ValueError('Missing statement label')
66
+
67
+ checks = rows('verification_checks')
68
+ for row in checks:
69
+ original = json.loads((ROOT / row['source_path']).read_text())
70
+ expected_group = original['checks'][row['check_name']] if row['suite'] == 'orythmea' else original[row['check_name']]
71
+ if json.loads(row['results_json']) != expected_group:
72
+ raise ValueError('Changed verification group')
73
+
74
+ metadata = json.loads((ROOT / 'release_metadata.json').read_text())
75
+ counts = {'manuscript_sections': len(sections), 'claims': len(claims), 'verification_checks': len(checks)}
76
+ if metadata['data_configs'] != counts or metadata['universal_conjecture_resolved']:
77
+ raise ValueError('Invalid release metadata')
78
+ for name in counts:
79
+ data = rows(name)
80
+ keys = set(data[0])
81
+ types = {k: type(data[0][k]) for k in keys}
82
+ if len({row['record_id'] for row in data}) != len(data):
83
+ raise ValueError('Duplicate record IDs')
84
+ for row in data:
85
+ if set(row) != keys or any(type(row[k]) != types[k] for k in keys):
86
+ raise ValueError('Inconsistent dataset schema')
87
+ print(json.dumps({'status': 'passed', 'manifest_files': len(expected), 'original_files_preserved': 8, 'dataset_records': counts}, indent=2))
88
+
89
+
90
+ if __name__ == '__main__':
91
+ main()