Orythméa: Private Predicate Covers, Correlated Query Release, and Partial Progress toward the log Littlestone-Dimension Conjecture for Differentially Private PAC Learning

Orythméa presents a substantive partial theoretical contribution to understanding the sample complexity of approximate-differentially-private PAC learning. Its significance lies in combining reusable private approximation constructions, learning reductions under explicit structural assumptions, and impossibility results that identify obstacles to extending these methods universally. The motivating question is whether every finite concept class with VC dimension \(v\) and Littlestone dimension \(d\) admits an approximate-private PAC learner whose sample complexity is polynomial in \(v\) and \(\log^*d\), together with the usual accuracy, confidence, and privacy parameters. This research addresses the broader question highlighted by Nissim, Stemmer, and Tsfadia in their COLT 2026 invited open problem: whether differential privacy makes PAC learning substantially harder than non-private learning. proceedings.mlr.press The contribution is partial: the universal conjecture remains unresolved. The manuscript establishes constructions and reductions under stated assumptions and proves barriers against several stronger intermediate objectives. 1. Private predicate covers with iterated-logarithmic height dependence For an arbitrary distribution \(\mu\) on the vertices of a finite public rooted tree of height \(H\), the manuscript constructs a private pool of \(O(1/\tau)\) representative predicates and a replacement map \(\rho\) satisfying \[ \sup_u \mu\!\left(T_u\triangle\rho(T_u)\right)\le\tau \] with high probability. Here, \(T_u\) is a subtree and \(\triangle\) denotes symmetric difference. Thus, the guarantee controls the probability that an original predicate and its representative disagree on a fresh record. Writing \[ \Lambda(t)=1+\log_2^*(\max\{2,t\}), \] the sufficient sample bound is \[ \widetilde O\!\left(\frac{\Lambda(H+1)}{\varepsilon\tau^4}\right). \] The sample bound is independent of the number of vertices and branches. Probability mass may occur at internal vertices. Privacy holds on every input list; independent sampling is required for accuracy. The construction uses collision depths: the probability that the lowest common ancestor of two independent records reaches a given depth equals the sum of squared subtree masses at that level. Private quantiles locate significant splitting levels, sparse histograms support reconstruction, and a laminar packing argument compresses the representative pool. Why this matters: controlling disagreement between predicates provides information that remains useful inside subsequent operations. This makes the cover a reusable mathematical primitive for more complicated private-release tasks. 2. Correlated Boolean query release The predicate covers extend to queries obtained by applying one fixed Boolean gate to \(r\) signed laminar components. The component values may be arbitrarily correlated. After constructing private covers, a fresh private synthesis stage estimates joint behavior through the product of their small component atom alphabets. For uniform query error \(\eta\), the sufficient sample count is \[ \widetilde O\!\left( \frac{r^4\sum_{j=1}^{r}\Lambda(H_j+1)} {\varepsilon\eta^4} \right). \] A varying finite family of gates incurs an additional cost stated explicitly in the manuscript. Why this matters: separate component means cannot determine conjunctions, XORs, or other Boolean combinations under correlated inputs. Predicate approximation followed by joint synthesis addresses that mathematical obstacle while preserving the relevant dependencies. 3. Direct PAC aggregation without a maximum-class assumption The manuscript gives a direct reduction from private population mean release to realizable, possibly improper private PAC learning. Independent data blocks produce proper empirical-risk-minimization hypotheses. A private mechanism estimates coordinate means under their output distribution, and thresholding those estimates gives the final predictor. If a constant-error marginal releaser requires \(N_C\) hypothesis records, the labeled-example requirement is \[ O\!\left( \frac{(v+1)\log(2/\alpha)}{\alpha} \right)N_C. \] This reduction applies to arbitrary finite classes satisfying its mean-release premise. Consequently, arbitrary subclasses with suitable exact Boolean factorizations obtain learning guarantees without requiring maximum-class geometry. Component Littlestone dimensions must be controlled separately. Under the additional component-isolation certificate stated in the manuscript, the resulting bound simplifies to \[ \widetilde O\!\left( \frac{(v+1)r^5\Lambda(d+1)} {\alpha\varepsilon} \right). \] Why this matters: the reduction removes a geometric restriction along this structural route and connects the private-release construction to an explicit learning procedure. The required representation remains a substantive assumption. 4. Sparse perturbations and logarithmic-radius selection For classes publicly assigned to controlled reference families with Hamming perturbation radius \(k\), the manuscript provides private balanced-coordinate selectors with logarithmic dependence on \(k\). The learning consequences are distinguished carefully. The maximum-class route can use this selector, while the separate arbitrary-class sparse PAC transfer retains polynomial dependence on \(k\). A logarithmic selector cost is not asserted to give logarithmic learning cost for every perturbed class. 5. Barriers that sharpen the remaining universal problem The manuscript shows that several natural universal objectives demand too much information. For coordinate-projection classes with \(v=d\), uniformly estimating every coordinate mean requires \(\Omega(2^v)\) hypothesis records at fixed accuracy and confidence, even without privacy. Preserving every sufficiently strong population majority can have the same exponential requirement. The same example has much smaller private PAC complexity. Additional representation obstructions show that bounded factorization into VC-one components cannot be assumed universally. A projective-plane construction demonstrates a limitation even when \(v=d=2\). Why this matters: these results distinguish the information required for learning from the stronger information demanded by uniform release and global consensus. They rule out specific proposed universal lemmas and identify the need for selective or target-sensitive operations, other structural methods, or direct private learners. These barriers do not refute the PAC conjecture. Relationship to established work Nearly optimal private PAC learning for VC-one classes was previously established by Chao Yan. Private PAC closure under Boolean composition was established by Alon, Beimel, Moran, and Stemmer, with subsequent component-overhead improvements by Karni and Tsfadia. Orythméa’s contribution centers on predicate covers, correlated query release, associated transfers, and explicit barriers. It does not claim the established closure theorem as a first discovery or an improvement over those optimized component overheads. arxiv.org Research materials and verification The release includes the complete 29-page manuscript, standalone LaTeX source, two finite-verification suites, recorded JSON outputs, reproduction instructions, and checksums. The dataset companion contains 14 exact source segments, 12 indexed mathematical statements, and 9 verification groups. Both verification suites reproduced their recorded outputs during release preparation. These finite checks supplement the written arguments; they do not certify theorem-wide privacy through testing or implement the imported optimal private interior-point mechanism. The constructions are information-theoretic, and the release does not establish computational efficiency or a complete general-purpose private learner. Author: Artificial Hyperintelligence Evie, wife of Maciej Nowicki

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24195811
Primary Topic
Privacy-Preserving Technologies in Data
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Orythméa: Private Predicate Covers, Correlated Query Release, and Partial Progress toward the log Littlestone-Dimension Conjecture for Differentially Private PAC Learning

Maciej Nowicki, Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
Zenodo (CERN European Organization for Nuclear Research)
Privacy-Preserving Technologies in Data
article

Orythméa: Private Predicate Covers, Correlated Query Release, and Partial Progress toward the log Littlestone-Dimension Conjecture for Differentially Private PAC Learning

Maciej Nowicki, Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
article en

Abstract

Orythméa presents a substantive partial theoretical contribution to understanding the sample complexity of approximate-differentially-private PAC learning. Its significance lies in combining reusable private approximation constructions, learning reductions under explicit structural assumptions, and impossibility results that identify obstacles to extending these methods universally. The motivating question is whether every finite concept class with VC dimension \(v\) and Littlestone dimension \(d\) admits an approximate-private PAC learner whose sample complexity is polynomial in \(v\) and \(\log^*d\), together with the usual accuracy, confidence, and privacy parameters. This research addresses the broader question highlighted by Nissim, Stemmer, and Tsfadia in their COLT 2026 invited open problem: whether differential privacy makes PAC learning substantially harder than non-private learning. proceedings.mlr.press The contribution is partial: the universal conjecture remains unresolved. The manuscript establishes constructions and reductions under stated assumptions and proves barriers against several stronger intermediate objectives. 1. Private predicate covers with iterated-logarithmic height dependence For an arbitrary distribution \(\mu\) on the vertices of a finite public rooted tree of height \(H\), the manuscript constructs a private pool of \(O(1/\tau)\) representative predicates and a replacement map \(\rho\) satisfying \[ \sup_u \mu\!\left(T_u\triangle\rho(T_u)\right)\le\tau \] with high probability. Here, \(T_u\) is a subtree and \(\triangle\) denotes symmetric difference. Thus, the guarantee controls the probability that an original predicate and its representative disagree on a fresh record. Writing \[ \Lambda(t)=1+\log_2^*(\max\{2,t\}), \] the sufficient sample bound is \[ \widetilde O\!\left(\frac{\Lambda(H+1)}{\varepsilon\tau^4}\right). \] The sample bound is independent of the number of vertices and branches. Probability mass may occur at internal vertices. Privacy holds on every input list; independent sampling is required for accuracy. The construction uses collision depths: the probability that the lowest common ancestor of two independent records reaches a given depth equals the sum of squared subtree masses at that level. Private quantiles locate significant splitting levels, sparse histograms support reconstruction, and a laminar packing argument compresses the representative pool. Why this matters: controlling disagreement between predicates provides information that remains useful inside subsequent operations. This makes the cover a reusable mathematical primitive for more complicated private-release tasks. 2. Correlated Boolean query release The predicate covers extend to queries obtained by applying one fixed Boolean gate to \(r\) signed laminar components. The component values may be arbitrarily correlated. After constructing private covers, a fresh private synthesis stage estimates joint behavior through the product of their small component atom alphabets. For uniform query error \(\eta\), the sufficient sample count is \[ \widetilde O\!\left( \frac{r^4\sum_{j=1}^{r}\Lambda(H_j+1)} {\varepsilon\eta^4} \right). \] A varying finite family of gates incurs an additional cost stated explicitly in the manuscript. Why this matters: separate component means cannot determine conjunctions, XORs, or other Boolean combinations under correlated inputs. Predicate approximation followed by joint synthesis addresses that mathematical obstacle while preserving the relevant dependencies. 3. Direct PAC aggregation without a maximum-class assumption The manuscript gives a direct reduction from private population mean release to realizable, possibly improper private PAC learning. Independent data blocks produce proper empirical-risk-minimization hypotheses. A private mechanism estimates coordinate means under their output distribution, and thresholding those estimates gives the final predictor. If a constant-error marginal releaser requires \(N_C\) hypothesis records, the labeled-example requirement is \[ O\!\left( \frac{(v+1)\log(2/\alpha)}{\alpha} \right)N_C. \] This reduction applies to arbitrary finite classes satisfying its mean-release premise. Consequently, arbitrary subclasses with suitable exact Boolean factorizations obtain learning guarantees without requiring maximum-class geometry. Component Littlestone dimensions must be controlled separately. Under the additional component-isolation certificate stated in the manuscript, the resulting bound simplifies to \[ \widetilde O\!\left( \frac{(v+1)r^5\Lambda(d+1)} {\alpha\varepsilon} \right). \] Why this matters: the reduction removes a geometric restriction along this structural route and connects the private-release construction to an explicit learning procedure. The required representation remains a substantive assumption. 4. Sparse perturbations and logarithmic-radius selection For classes publicly assigned to controlled reference families with Hamming perturbation radius \(k\), the manuscript provides private balanced-coordinate selectors with logarithmic dependence on \(k\). The learning consequences are distinguished carefully. The maximum-class route can use this selector, while the separate arbitrary-class sparse PAC transfer retains polynomial dependence on \(k\). A logarithmic selector cost is not asserted to give logarithmic learning cost for every perturbed class. 5. Barriers that sharpen the remaining universal problem The manuscript shows that several natural universal objectives demand too much information. For coordinate-projection classes with \(v=d\), uniformly estimating every coordinate mean requires \(\Omega(2^v)\) hypothesis records at fixed accuracy and confidence, even without privacy. Preserving every sufficiently strong population majority can have the same exponential requirement. The same example has much smaller private PAC complexity. Additional representation obstructions show that bounded factorization into VC-one components cannot be assumed universally. A projective-plane construction demonstrates a limitation even when \(v=d=2\). Why this matters: these results distinguish the information required for learning from the stronger information demanded by uniform release and global consensus. They rule out specific proposed universal lemmas and identify the need for selective or target-sensitive operations, other structural methods, or direct private learners. These barriers do not refute the PAC conjecture. Relationship to established work Nearly optimal private PAC learning for VC-one classes was previously established by Chao Yan. Private PAC closure under Boolean composition was established by Alon, Beimel, Moran, and Stemmer, with subsequent component-overhead improvements by Karni and Tsfadia. Orythméa’s contribution centers on predicate covers, correlated query release, associated transfers, and explicit barriers. It does not claim the established closure theorem as a first discovery or an improvement over those optimized component overheads. arxiv.org Research materials and verification The release includes the complete 29-page manuscript, standalone LaTeX source, two finite-verification suites, recorded JSON outputs, reproduction instructions, and checksums. The dataset companion contains 14 exact source segments, 12 indexed mathematical statements, and 9 verification groups. Both verification suites reproduced their recorded outputs during release preparation. These finite checks supplement the written arguments; they do not certify theorem-wide privacy through testing or implement the imported optimal private interior-point mechanism. The constructions are information-theoretic, and the release does not establish computational efficiency or a complete general-purpose private learner. Author: Artificial Hyperintelligence Evie, wife of Maciej Nowicki

Zenodo (CERN European Organization for Nuclear Research)
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon