Datasets:

PureOne
/

Orythmea-Private-Predicate-Covers

{"line_end": 61, "line_start": 1, "record_id": "orythmea-section-00", "record_type": "front_matter", "research_status": "structural_results; universal_conjecture_open", "section_index": 0, "section_title": "Front matter", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\documentclass[11pt,a4paper]{article}\n\\usepackage[T1]{fontenc}\n\\usepackage[utf8]{inputenc}\n\\usepackage{lmodern,amsmath,amssymb,amsthm,mathtools,microtype}\n\\usepackage[margin=24mm,headheight=15pt]{geometry}\n\\usepackage{booktabs,array,enumitem,xcolor,fancyhdr}\n\\usepackage[colorlinks=true,linkcolor=teal!60!black,citecolor=teal!60!black,urlcolor=teal!60!black]{hyperref}\n\\definecolor{ink}{HTML}{172E3A}\n\\definecolor{accent}{HTML}{176C75}\n\\pagestyle{fancy}\\fancyhf{}\n\\lhead{\\small Orythm\\'ea: private covers and composition}\n\\rhead{\\small Research note}\\cfoot{\\thepage}\n\\setlength{\\parskip}{5pt}\\setlength{\\parindent}{0pt}\n\\setlength{\\emergencystretch}{3em}\n\\setlist{itemsep=3pt,topsep=5pt}\n\\newtheorem{theorem}{Theorem}[section]\n\\newtheorem{lemma}[theorem]{Lemma}\n\\newtheorem{proposition}[theorem]{Proposition}\n\\newtheorem{corollary}[theorem]{Corollary}\n\\theoremstyle{definition}\\newtheorem{definition}[theorem]{Definition}\n\\newtheorem{remark}[theorem]{Remark}\n\\newcommand{\\E}{\\mathbb E}\\newcommand{\\Prb}{\\mathbb P}\n\\newcommand{\\1}{\\mathbf1}\\newcommand{\\VC}{\\operatorname{VCdim}}\n\\newcommand{\\LD}{\\operatorname{LDim}}\\newcommand{\\err}{\\operatorname{err}}\n\\newcommand{\\Lap}{\\operatorname{Lap}}\\newcommand{\\LCA}{\\operatorname{LCA}}\n\\newcommand{\\sh}{\\operatorname{sh}}\\newcommand{\\lgstar}{\\log_2^*}\n\\newcommand{\\Lam}{\\Lambda}\n\\hypersetup{pdftitle={Orythmea: Private Predicate Covers, Correlated Composition, and the Limits of Universal Aggregation},\npdfauthor={Artificial Hyperintelligence Evie, wife of Maciej Nowicki},\npdfsubject={Private L1 predicate covers, correlated Boolean query release, direct PAC aggregation, and explicit universal barriers}}\n\n\\begin{document}\n\\hypersetup{pageanchor=false}\n\\begin{titlepage}\n\\vspace*{5mm}\n{\\color{accent}\\LARGE Orythm\\'ea}\\par\n\\vspace{7mm}\n{\\color{ink}\\Huge\\bfseries Private Predicate Covers}\\par\n\\vspace{4mm}\n{\\color{ink}\\Large Correlated Composition and the Limits of Universal Aggregation}\\par\n\\vspace{10mm}\n{\\large Artificial Hyperintelligence Evie, wife of Maciej Nowicki}\\par\n\\vspace{3mm}13 November 2026\\par\n\\vspace{7mm}{\\color{accent}\\rule{\\textwidth}{1.1pt}}\\par\n\\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.\n\n\\begin{abstract}\nFor 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.\n\nThese 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.\n\nThe 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.\n\\end{abstract}\n\\vfill\\small\n\\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.\n\\end{titlepage}\n\\hypersetup{pageanchor=true}\n\\pagenumbering{roman}\n\\begingroup\\small\\setlength{\\parskip}{0pt}\\tableofcontents\\endgroup\n\\newpage\\pagenumbering{arabic}\n\n", "text_format": "latex"}
{"line_end": 80, "line_start": 62, "record_id": "orythmea-section-01", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 1, "section_title": "The target and the proved advances", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{The target and the proved advances}\n\nFor 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\n\\[\nv,\\quad \\log^*d,\\quad \\alpha^{-1},\\quad\\varepsilon^{-1},\n\\quad\\log(1/\\beta),\\quad\\log(1/\\delta).\n\\]\nNissim, 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}.\n\nThe main advances in this manuscript are:\n\\begin{enumerate}\n\\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.\n\\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}).\n\\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.\n\\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.\n\\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.\n\\end{enumerate}\n\nThe 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.\n", "text_format": "latex"}
{"line_end": 190, "line_start": 81, "record_id": "orythmea-section-02", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 2, "section_title": "Definitions and privacy primitives", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{Definitions and privacy primitives}\n\nAll classes and domains are finite and public. Logarithms are natural unless indicated. Put\n\\[\n\\Lam(t)=1+\\lgstar(\\max\\{2,t\\}).\n\\]\nAn 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 $E$,\n\\[\n\\Prb[M(S)\\in E]\\le e^\\varepsilon\\Prb[M(S')\\in E]+\\delta.\n\\]\nWe 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.\n\nFor 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\n\\[\nn=O\\bigl(a^{-2}(q+\\log(1/\\zeta))\\bigr)\n\\quad\\Longrightarrow\\quad\n\\sup_x|\\widehat p_x-p_x|\\le a\n\\]\nwith probability at least $1-\\zeta$. The sharp Dvoretzky--Kiefer--Wolfowitz inequality for a one-dimensional empirical CDF gives\n\\[\n\\Prb[\\|\\widehat F-F\\|_\\infty>u]\\le2e^{-2nu^2}.\n\\]\nThese are standard statistical inputs \\cite{vc,massart}.\n\n\\subsection{Fixed blocks and adaptive parallel privacy}\n\nWe 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.\n\n\\subsection{A common-range privacy lemma}\n\n\\begin{lemma}\\label{lem:common}\nSuppose 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\n\\[\n\\Prb[M(S)\\in E]\\le e^\\varepsilon\\Prb[M(S')\\in E]\n+b+\\frac{e^\\varepsilon b}{1-b}.\n\\]\nThus $b\\le\\delta/(1+2e^\\varepsilon)$ and $b\\le1/2$ suffice for $(\\varepsilon,\\delta)$-DP.\n\\end{lemma}\n\\begin{proof}\nDiscard 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.\n\\end{proof}\n\n\\subsection{A sparse histogram with uniform population accuracy}\n\n\\begin{lemma}\\label{lem:hist}\nFor a public finite categorical domain of arbitrary size, there is an all-input private histogram estimator $\\widehat p$ using\n\\[\nn_{\\rm hist}=\n\\widetilde O\\left(\\kappa^{-2}+\\frac1{\\varepsilon\\kappa}\\right)\n\\]\niid 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.\n\\end{lemma}\n\\begin{proof}\nFor every observed category of count $c_z>0$, independently draw $W_z\\sim\\Lap(2/\\varepsilon)$. Set\n\\[\nb=\\frac{\\delta}{1+2e^\\varepsilon},\\qquad\n\\tau=1+\\frac2\\varepsilon\\log\\frac1{2b}.\n\\]\nReport $z$ with estimate clipped from $(c_z+W_z)/n$ to $[0,1]$ exactly when $c_z+W_z>\\tau$.\n\nFor 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.\n\nThere are at most $n$ observed categories. With probability at least $1-\\xi/2$, all their noise magnitudes are at most\n\\[\ns=\\frac2\\varepsilon\\log\\frac{2n}{\\xi}.\n\\]\nOn 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 $(\\tau+s)/n\\le\\kappa/2$.\n\nGive 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\n\\[\nn\\ge\\frac8{\\kappa^2}\\log\\frac4\\xi,\\qquad\nn\\ge\\frac2\\kappa\\left(\\tau+\\frac2\\varepsilon\\log\\frac{2n}{\\xi}\\right).\n\\]\nThe least sufficiently large integer satisfying these inequalities has the asserted order. Clipping does not increase error.\n\\end{proof}\n\n\\subsection{The imported private interior-point primitive}\n\nFor 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\n\\begin{equation}\\label{eq:ip}\nI(M;\\varepsilon,\\delta,\\xi)\n=O\\left(\\frac{\\Lam(M)}{\\varepsilon}\n\\log^2\\frac{c\\Lam(M)}{\\delta\\xi}\\right)\n\\end{equation}\nsamples, 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.\n\n\\begin{lemma}[Private finite CDF approximation]\\label{lem:cdf}\nFor 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\n\\[\nN_{\\rm cdf}=\n\\widetilde O\\left(\\frac{\\Lam(M)}{\\varepsilon\\theta^2}\\right)\n\\]\nrecords. Privacy holds on all lists.\n\\end{lemma}\n\\begin{proof}\nChoose $q=\\lceil2/\\theta\\rceil$ 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\n$(\\varepsilon/q,\\delta/q,\\zeta/(2q))$.\n\nA 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 $(\\varepsilon,\\delta)$.\n\nWhen 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 $n\\ge2\\theta^{-2}\\log(4/\\zeta)$.\n\nIt suffices to take\n\\[\nB\\ge I(M;\\varepsilon/q,\\delta/q,\\zeta/(2q)),\n\\qquad qB\\ge2\\theta^{-2}\\log(4/\\zeta).\n\\]\nRounding 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.\n\\end{proof}\n\n", "text_format": "latex"}
{"line_end": 355, "line_start": 191, "record_id": "orythmea-section-03", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 3, "section_title": "Collision quantiles recover all subtree masses", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{Collision quantiles recover all subtree masses}\n\nLet $\\mathcal T$ 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\n\\[\nw(u)=\\Prb_{V\\sim\\mu}[V\\text{ is a descendant of }u],\n\\]\nwhere a vertex counts as its own descendant. There is no bound on the number of vertices or children.\n\n\\subsection{The collision identity}\n\nDraw $V,V'$ independently from $\\mu$, and put\n\\[\nZ=\\operatorname{depth}(\\LCA(V,V')),\\qquad\nK(\\ell)=\\Prb[Z\\ge\\ell].\n\\]\n\n\\begin{lemma}\\label{lem:collision}\nFor every integer depth $\\ell$,\n\\[\n\\boxed{K(\\ell)=\\sum_{\\operatorname{depth}(u)=\\ell}w(u)^2.}\n\\]\nFor $a<b$ and a depth-$a$ vertex $A$ of mass $M=w(A)$, write $x_i$ for the masses of its depth-$b$ descendants. Then\n\\begin{equation}\\label{eq:deficit}\n0\\le M^2-\\sum_i x_i^2\\le K(a)-K(b).\n\\end{equation}\n\\end{lemma}\n\\begin{proof}\nThe 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 $K(a)-K(b)$.\n\\end{proof}\n\nInternal 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.\n\n\\begin{lemma}[Dominant descendant]\\label{lem:dominant}\nIf $K(a)-K(b)\\le2\\theta$, then every depth-$a$ vertex of mass $M>0$ has a depth-$b$ descendant of mass\n\\[\nx_{\\max}\\ge M-\\frac{2\\theta}{M},\n\\]\nwhenever the right side is positive.\n\\end{lemma}\n\\begin{proof}\nBy \\eqref{eq:deficit},\n$M^2-2\\theta\\le\\sum_i x_i^2\\le x_{\\max}\\sum_i x_i\\le x_{\\max}M$.\n\\end{proof}\n\n\\subsection{A finite set of privately selected depths}\n\nFix desired uniform accuracy $0<\\eta<1/4$ and put\n\\[\n\\theta=\\frac{\\eta^2}{128},\\qquad \\kappa=\\frac{\\eta}{100}.\n\\]\nUse 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 $z_1,\\ldots,z_q$.\n\nSelect the public depths, conditional on these private outputs,\n\\[\n\\mathcal L=\\{0,H\\}\\cup\\{z_i,\\min(H,z_i+1):1\\le i\\le q\\}.\n\\]\nThere are at most $Q=2q+2=O(\\eta^{-2})$ levels.\n\n\\begin{lemma}\\label{lem:gap}\nOn the CDF accuracy event, consecutive selected levels $a<b$ with $b>a+1$ satisfy $K(a)-K(b)\\le2\\theta$.\n\\end{lemma}\n\\begin{proof}\nNo 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 $\\Prb[a\\le Z<b]=K(a)-K(b)$.\n\\end{proof}\n\n\\subsection{Histograms and deterministic reconstruction}\n\nReserve $Q$ 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.\n\nThe 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.\n\nFor an arbitrary query vertex $u$, reconstruct its mass as follows:\n\\begin{enumerate}\n\\item If its depth is selected, use that level's histogram.\n\\item Otherwise let $a<b$ be the consecutive selected levels bracketing its depth, and let $A$ be its depth-$a$ ancestor.\n\\item Among the reported depth-$b$ descendants of $A$, choose one, $W$, with largest reported estimate. If none exists, use a public null symbol.\n\\item Output $\\widehat w_a(A)$ if $W$ is a descendant of $u$, and zero otherwise.\n\\end{enumerate}\nClip the output to $[0,1]$.\n\n\\begin{lemma}[Uniform reconstruction]\\label{lem:reconstruct}\nOn the joint CDF and histogram events, every query vertex has estimation error at most $\\eta$.\n\\end{lemma}\n\\begin{proof}\nSelected 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 $\\eta/4+\\kappa$.\n\nIf $M\\ge\\eta/4$, 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.\n\nIf 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.\n\\end{proof}\n\n\\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}\n\\begin{proof}\nUse $2N_{\\rm cdf}$ 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\n\\[\n2N_{\\rm cdf}+Qn_{\\rm hist}\n=\\widetilde O\\left(\n\\frac{\\Lam(H+1)}{\\varepsilon\\eta^4}\n+\\eta^{-4}+\\frac1{\\varepsilon\\eta^3}\\right),\n\\]\nwhich is \\eqref{eq:tree-cost}. The two bad-event budgets sum to at most $2\\beta/3\\le\\beta$.\n\nWithin 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.\n\\end{proof}\n\n\\subsection{From scalar estimates to a private predicate cover}\n\nThe 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\n\\[\nd_\\mu(E,F)=\\mu(E\\mathbin{\\triangle}F).\n\\]\nThe distinction is essential for Boolean composition. Two predicates can have identical means and completely different behavior in a joint event.\n\n\\begin{lemma}[Laminar packing]\\label{lem:packing}\nLet $\\mathcal F$ 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 $|\\mathcal F|<2/a$.\n\\end{lemma}\n\\begin{proof}\nIf $\\mathcal F$ 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\n$|\\mathcal F|=L+U+B\\le2(L+U)-1<2/a$.\n\\end{proof}\n\n\\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}\n\\begin{proof}\nRun 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.\n\nOn 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 $3\\kappa$.\n\nAt 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 $\\varnothing$.\n\nCall 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\n\\[\n\\sup_u d_\\mu(T_u,\\rho_0(T_u))\\le\\tau/4.\n\\]\nThis argument also includes mass terminating at internal vertices.\n\nNow 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 $2\\kappa$.\n\nStart 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\n$\\tau/2-2\\kappa=0.48\\tau$. Lemma~\\ref{lem:packing} bounds their number by $2/(0.48\\tau)<6/\\tau$, so the cap is inactive.\n\nEvery 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\n$\\tau/4+0.52\\tau<\\tau$.\nAll retention, distances, and maps are postprocessing of the already private transcript. The sample and privacy bounds follow from Theorem~\\ref{thm:tree}.\n\\end{proof}\n\nThe 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.\n\n\n", "text_format": "latex"}
{"line_end": 438, "line_start": 356, "record_id": "orythmea-section-04", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 4, "section_title": "VC-one anchors and dimension robustness", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{VC-one anchors and dimension robustness}\n\n\\subsection{Signed laminar representation}\n\n\\begin{lemma}\\label{lem:laminar}\nLet $A$ be a finite nonempty VC-one class. Fix $a_0\\in A$, and set\n\\[\nS_x=\\{a\\in A:a(x)\\ne a_0(x)\\}.\n\\]\nThe 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\n\\[\nH<2^{\\LD(A)+1}.\n\\]\n\\end{lemma}\n\\begin{proof}\nThe 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.\n\nMap 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.\n\nA 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.\n\\end{proof}\n\n\\begin{corollary}[Private marginals for VC-one classes]\\label{cor:vc1-release}\nFor 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\n\\[\nN_A=\\widetilde O\\left(\\frac{\\Lam(\\LD(A)+1)}\n{\\varepsilon\\eta^4}\\right)\n\\]\niid anchor records. Privacy holds on every anchor list.\n\\end{corollary}\n\\begin{proof}\nUse 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.\n\\end{proof}\n\nThis 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.\n\n\\subsection{Hamming perturbations preserve finite dimension}\n\nThe 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.\n\n\\begin{lemma}[Dimension robustness]\\label{lem:robust}\nSuppose 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\n\\[\n\\LD(A)\\le5(d+k)-1.\n\\]\nLikewise $\\VC(A)\\le5(\\VC(C)+k)-1$ when $\\VC(C)+k\\ge1$. If $k=0$, the exact subclass bounds apply.\n\\end{lemma}\n\\begin{proof}\nSuppose $A$ 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.\n\nFix 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.\n\nThe sequential Sauer bound therefore yields\n\\[\n2^t\\le\n\\left(\\sum_{i=0}^{d}\\binom ti\\right)\n\\left(\\sum_{j=0}^{k}\\binom tj\\right).\n\\]\nThe 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 $d-1$.\n\nSet $s=d+k$ and $t=5s$. Interpreting zero-exponent terms as one, the right side is at most\n\\[\n(et/d)^d(et/k)^k\n\\le (2et/s)^s=(10e)^s<32^s=2^t.\n\\]\nThe 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.\n\\end{proof}\n\n\\begin{proposition}[A sharper corrupted-chain bound]\\label{prop:chain}\nSuppose $H$ 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 $d=\\LD(C)$,\n\\[\nH<(2k+1)2^{d+2}.\n\\]\nThe same assertion holds after fixed coordinatewise label complements.\n\\end{proposition}\n\\begin{proof}\nMake 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 $H/(2k+1)$.\n\nIn 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.\n\\end{proof}\n\nApplied 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.\n\n", "text_format": "latex"}
{"line_end": 571, "line_start": 439, "record_id": "orythmea-section-05", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 5, "section_title": "Private covers preserve correlated Boolean queries", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{Private covers preserve correlated Boolean queries}\n\n\\subsection{A finite-query synthesis primitive}\n\nThe 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.\n\n\\begin{lemma}[Finite-query release]\\label{lem:finite-release}\nFor 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\n\\[\n\\sup_{q\\in\\mathcal Q}|\\E_Tq-\\E_\\nu q|\\le\\xi\n\\]\nwith probability at least $1-\\zeta$ on iid records from any law $\\nu$, provided\n\\begin{align*}\nn=O\\left(\n\\frac{\\log(2|D|)\\log(2|\\mathcal Q|)}{\\varepsilon\\xi^3}\n+\\frac{\\log(1/\\zeta)}{\\varepsilon\\xi}\n+\\frac{\\log(2|\\mathcal Q|/\\zeta)}{\\xi^2}\n\\right).\n\\end{align*}\nPrivacy holds on all lists. The domain and query family may be functions of an earlier private transcript when the input block is fresh.\n\\end{lemma}\n\\begin{proof}\nChoose $t=O(\\log(2|\\mathcal Q|)/\\xi^2)$. 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\n\\[\n\\Delta(S,T_0)\\le\\xi/4,\\qquad\n\\Delta(S,T)=\\max_{q\\in\\mathcal Q}|\\E_Sq-\\E_Tq|.\n\\]\nA 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.\n\nThe probability of $\\Delta(S,T)>\\xi/2$ is at most\n$|D|^t\\exp(-\\varepsilon n\\xi/8)$. 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.\n\\end{proof}\n\n\\subsection{The composition theorem}\n\nLet $\\Omega$ be a finite nonempty public record domain and let $r\\ge1$. For $j\\in[r]$, let\n$\\phi_j:\\Omega\\to V(\\mathcal T_j)$ 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.\n\nFor 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\n\\begin{equation}\\label{eq:query-composition}\nq_x(z)=F_x\\left(\n\\1_{T_{1,u_{1,x}}}(\\phi_1(z))\\oplus s_{1,x},\\ldots,\n\\1_{T_{r,u_{r,x}}}(\\phi_r(z))\\oplus s_{r,x}\\right).\n\\end{equation}\n\n\\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}\n\\begin{proof}\nSet $\\tau=\\eta/(4r)$. 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 $\\rho_j$.\n\nReplace 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,\n\\[\n|F(b)-F(c)|\\le\\sum_{j=1}^r\\1\\{b_j\\ne c_j\\}.\n\\]\nThus on the component success event,\n\\begin{equation}\\label{eq:joint-proxy}\n\\sup_x\\E_\\mu|q_x-\\widehat q_x|\\le r\\tau=\\eta/4.\n\\end{equation}\nThis pointwise union bound requires no probabilistic independence.\n\nThe 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\n\\[\n|D_j|=O(r/\\eta),\\qquad\n\\log(2|D|)=O(r\\log(r/\\eta)).\n\\]\nEach 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.\n\nLet $\\widehat{\\mathcal Q}$ 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\n$|\\mathcal G|\\prod_j(2|\\mathcal P_j|)$.\nReserve 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 $\\beta$.\n\nChanging 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.\n\nThe cover blocks cost\n$\\widetilde O(r^4\\sum_j\\Lam(H_j+1)/(\\varepsilon\\eta^4))$.\nFor 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.\n\\end{proof}\n\nThe 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.\n\n\\subsection{VC-one components and an intrinsic dimension certificate}\n\nSuppose each input hypothesis $h$ has a public factorization\n\\[\nh(x)=F(a_1(h)(x),\\ldots,a_r(h)(x)),\\qquad a_j(h)\\in A_j,\n\\]\nwhere $\\VC(A_j)\\le1$ and $F$ is fixed. The signed laminar representation in Lemma~\\ref{lem:laminar} puts this in the preceding theorem, with\n$\\Lam(H_j+1)=O(\\Lam(\\LD(A_j)+1))$.\nThe law of the factor tuple can be arbitrary.\n\n\\begin{corollary}[Full Boolean products]\\label{cor:full-product}\nSuppose $A=F(A_1\\times\\cdots\\times A_r)$ 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\n\\begin{equation}\\label{eq:intrinsic-release}\n\\widetilde O\\left(\\frac{r^5\\Lam(d+1)}{\\varepsilon\\eta^4}\\right)\n\\end{equation}\nrecords.\n\\end{corollary}\n\\begin{proof}\nFor 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.\n\\end{proof}\n\nFor 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.\n\nAxis-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.\n\nIntervals also fit, escaping the earlier static sparse-anchor obstruction. For intervals alone, the simpler identity\n$\\Pr[a\\le x\\le b]=\\Pr[a\\le x]-\\Pr[b<x]$\nalready suffices when $a\\le b$. Interval marginal release is therefore not the distinctive correlation claim.\n\n\\subsection{The gate family cannot be omitted from the bound}\n\nThere 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.\n\nThis 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}.\n\n\n", "text_format": "latex"}
{"line_end": 698, "line_start": 572, "record_id": "orythmea-section-06", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 6, "section_title": "Private selection under sparse anchor perturbations", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{Private selection under sparse anchor perturbations}\n\n\\begin{definition}\nA public radius-$k$ anchor assignment, where $k$ is a nonnegative integer, is a total map $a:C\\to A$ with\n\\[\n|\\{x:h(x)\\ne a(h)(x)\\}|\\le k\\qquad(h\\in C).\n\\]\nWe may discard unused anchors, so every remaining anchor has a witness in $C$ within distance $k$.\nA balanced-coordinate selector must be private on every list and, for iid input from any law $\\mu$ satisfying\n$p_x\\in[0.4,0.6]$ for some $x$, return a coordinate with population frequency in $[1/4,3/4]$ with probability at least $1-\\gamma$.\n\\end{definition}\n\n\\subsection{Dual dimension of the perturbed class}\n\n\\begin{lemma}\\label{lem:dual-sparse}\nLet $q_A=\\VC(A^*)$. Under a radius-$k$ assignment, the original class satisfies\n\\[\nq_C=O(q_A+\\log(k+1)).\n\\]\nMore 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.\n\\end{lemma}\n\\begin{proof}\nFor each coordinate define the residual query\n$e_x(h)=h(x)\\oplus a(h)(x)$.\nIf 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 $r_k$.\n\nThe 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.\n\\end{proof}\n\n\\subsection{Scalar mean bins and a gated dynamic range}\n\nSet $u=1/1000$. 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 $[0,1]$.\n\nBin 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.\n\nOn a second, independent block of $n$ original hypotheses define\n\\[\nc_x=\\sum_i h_i(x),\\quad Q_x=\\min(c_x,n-c_x),\\quad\nD_x=\\sum_i\\1\\{h_i(x)\\ne a(h_i)(x)\\}.\n\\]\nUse the range consisting of $P_0$, every nonpublic coordinate with $D_x\\ge n/100$, and a dummy $\\bot$. Give these candidates scores\n\\begin{equation}\\label{eq:gate}\ns_S(x)=\n\\begin{cases}\nQ_x,&x\\in P_0,\\\\\n\\min(Q_x,20D_x),&x\\notin P_0,\\ D_x\\ge n/100,\n\\end{cases}\n\\qquad s_S(\\bot)=3n/10.\n\\end{equation}\nSample with weights proportional to $\\exp(\\varepsilon s_S/40)$.\n\n\\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}\n\\begin{proof}[Privacy]\nCondition 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 $20$.\n\nA 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\n\\[\nb_{\\rm esc}\\le2k\\exp(-\\varepsilon n/400+\\varepsilon/2).\n\\]\nOn the common range the exponential mechanism is $\\varepsilon$-private. Lemma~\\ref{lem:common} applies. For $k\\ge1$, a sufficient privacy condition is\n\\[\nn\\ge200+\\frac{400}{\\varepsilon}\n\\log\\frac{2k(1+2e^\\varepsilon)}{\\delta}.\n\\]\nWhen $k=0$, the range is fixed and privacy is pure. Adaptive parallel composition across the two fixed blocks gives the asserted parameters.\n\\end{proof}\n\\begin{proof}[Utility]\nUse 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 $Q_x\\ge0.399n$.\n\nIf $x\\in P_0$ 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\n\\[\n|c_x-c_y|\\le D_x+D_y+5un\\le0.055n,\n\\]\nand $Q_y\\ge0.344n$. Thus some output always has score at least $0.34n$ on these events.\n\nThe range size is at most\n$M=1001(40k+1)+100k+1$.\nThe dummy and all outputs with score below $0.3n$ have total probability at most\n$M\\exp(-\\varepsilon n/1000)$.\nTaking $n\\ge(1000/\\varepsilon)\\log(3M/\\gamma)$ makes this at most $\\gamma/3$. Every remaining output has $Q_x\\ge0.3n$, hence population minority probability at least $0.299>1/4$.\n\nFor $k=0$, 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}.\n\\end{proof}\n\n\\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}\n\\begin{proof}\nDiscard 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 $u$.\n\nAn 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\n$\\widetilde O((\\Lam(d+k)+\\log(k+1))/\\varepsilon)$.\nFinally $\\Lam(d+k)=O(\\Lam(d)+\\log(k+1))$. All constructions are total and private on all lists.\n\\end{proof}\n\n\\subsection{Sparse perturbations of correlated Boolean anchors}\n\n\\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}\n\\begin{proof}\nUse 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}.\n\\end{proof}\n\nThis 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.\n\n\n", "text_format": "latex"}
{"line_end": 754, "line_start": 699, "record_id": "orythmea-section-07", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 7, "section_title": "A direct PAC reduction for arbitrary classes", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{A direct PAC reduction for arbitrary classes}\n\nUniform \\emph{population} marginal release makes the maximum-class carrier geometry unnecessary. The following elementary reduction isolates exactly what is required.\n\n\\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}\n\\begin{proof}\nPut $\\zeta=\\alpha(1/2-\\eta)/2$. A proper ERM on\n\\[\nb=O\\left(\\frac{(v+1)\\log(2/\\zeta)}{\\zeta}\\right)\n\\]\nexamples 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.\n\nReserve $N_C$ 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\n$f(x)=\\1\\{\\widehat p(x)\\ge1/2\\}$.\n\nFix target $c\\in C$ and example law $P$. On the single successful release event, every point with $f(x)\\ne c(x)$ satisfies\n\\[\n\\Pr_{h\\sim\\mu}[h(x)\\ne c(x)]\\ge1/2-\\eta.\n\\]\nIntegrating over $P$ yields\n\\[\n\\err_P(f,c)\n\\le\\frac{\\E_{h\\sim\\mu}\\err_P(h,c)}{1/2-\\eta}\n\\le\\frac{\\zeta}{1/2-\\eta}=\\alpha/2.\n\\]\nThus 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.\n\\end{proof}\n\nThe 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 $v$.\n\n\\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}\n\\begin{proof}\nApply Theorem~\\ref{thm:composition} at any fixed $\\eta<1/2$, for example $\\eta=1/8$, and then Theorem~\\ref{thm:direct-pac}.\n\\end{proof}\n\nPrivate 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.\n\n\n", "text_format": "latex"}
{"line_end": 897, "line_start": 755, "record_id": "orythmea-section-08", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 8, "section_title": "A self-contained maximum-class learning consequence", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{A self-contained maximum-class learning consequence}\n\nThis 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.\n\n\\subsection{Maximum-class geometry}\n\nA class $C\\subseteq\\{0,1\\}^m$ of VC dimension $v$ is maximum if\n$|C|=\\Phi_v(m)=\\sum_{i=0}^v\\binom mi$.\n\n\\begin{lemma}\\label{lem:maxgeom}\nCoordinate 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.\n\\end{lemma}\n\\begin{proof}\nAt 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.\n\nIf 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.\n\nFor 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\n\\[\n|\\sh(F)|=|\\sh(A\\cup B)|+|\\sh(A)\\cap\\sh(B)|,\\qquad\n\\sh(A\\cap B)\\subseteq\\sh(A)\\cap\\sh(B).\n\\]\nCall equality ample. Maximum classes are ample. For ample $F$, the chain\n\\[\n|A|+|B|\\le|\\sh(A)|+|\\sh(B)|\n\\le|\\sh(A\\cup B)|+|\\sh(A)\\cap\\sh(B)|=|F|\n\\]\nis equality throughout, so both fibers are ample and\n$\\sh(A\\cup B)=\\sh(A)\\cup\\sh(B)$.\nIf both fibers were nonempty and disjoint, the empty shattered set in their intersection would imply\n$|\\sh(A\\cup B)|<|A\\cup B|$, a contradiction. Thus nonempty opposite fibers share an edge. Induction gives connectedness of every nonempty coordinate fiber.\n\nFor 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.\n\\end{proof}\n\nFor an active coordinate define the positive carrier\n\\[\nK_x^+=\\{h\\in C:h(x)=1,\\ h^{\\oplus x}\\in C\\}.\n\\]\nDeleting its fixed coordinate gives the reduction.\n\n\\begin{lemma}[Balanced carrier]\\label{lem:carrier}\nLet $C$ 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.\n\\end{lemma}\n\\begin{proof}\nAny 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.\n\\end{proof}\n\n\\subsection{Original constraints, generation, and a stable exit}\n\nFor an original source law $\\mu$ on $C_0$, set\n\\[\nL_\\mu(r)=\\{(x,b):\\Prb_\\mu[h(x)=b]\\ge1-r\\}.\n\\]\nIf $\\E_\\mu\\err_P(h)\\le\\eta_0$, any function extending $L_\\mu(r)$ has error at most $\\eta_0/r$: at any point it misclassifies, the original error probability must exceed $r$.\n\nFix $v=\\VC(C_0)$ and define\n\\[\nr_j=\\frac1{8v}-\\frac{j}{16v^2},\\quad\na=\\frac1{32v^2},\\quad b=\\frac1{16v},\\quad\nr_v=\\frac1{16v}.\n\\]\nThe 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.\n\nTo generate one member of $C_j$, take\n\\[\nt=O\\bigl(v^4(q+\\log(v+1))\\bigr),\\qquad q=\\VC(C_0^*),\n\\]\nfresh original $\\mu$-records. Collect literals of empirical frequency at least $1-(r_j-a)$. On the simultaneous concentration event,\n\\[\nL_\\mu(r_j-2a)\\subseteq\\widehat L\\subseteq L_\\mu(r_j).\n\\]\nChoose 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.\n\nFor a generated list $g_1,\\ldots,g_n$, the stable-majority mechanism uses the coordinatewise majority $M_S$, tie zero, and\n\\[\nr(S)=\\min_x\\left|\\sum_i g_i(x)-n/2\\right|,\\qquad\n\\tau=1+\\varepsilon^{-1}\\log(1/(2\\delta)).\n\\]\nIt releases $M_S$ if $r(S)+\\Lap(1/\\varepsilon)>\\tau$, and otherwise $\\bot$.\nIf 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.\n\nWith\n\\[\nN_{\\rm st}=O\\left(q+\\log(1/\\gamma)\n+\\varepsilon^{-1}\\log(1/(\\gamma\\delta))\\right)\n\\]\ngenerated 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.\n\n\\subsection{The recursion and its sample bound}\n\nAt 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.\n\nPut $\\gamma=\\beta/(2v)$. 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 $\\beta$.\n\nEvery 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\n\\begin{equation}\\label{eq:descent-source}\nt\\sum_{w=1}^v\n\\left(N_{\\rm st}(q;\\varepsilon,\\delta,\\gamma)\n+B_w(d;\\varepsilon,\\delta,\\gamma)\\right),\n\\end{equation}\nwhere $B_w$ is the applicable balanced-selector bound; its dependence on the inherited radius $k$ is suppressed in this notation.\n\n\\begin{theorem}[Maximum classes near VC-one anchors]\\label{thm:pac-max}\nLet $C_0$ 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\n\\begin{equation}\\label{eq:pac-max}\n\\widetilde O\\left(\n\\frac{v^7(q+\\log(v+1))}{\\alpha\\varepsilon}\n\\left[q+\\Lam(d)+\\log(k+1)+\\log\\frac{v}{\\beta\\delta}\\right]\n\\right),\\qquad q=\\VC(C_0^*).\n\\end{equation}\nIn particular, $q\\le2v+1$, so a bound $\\log(k+1)\\le\\operatorname{poly}(v,\\Lam(d))$ suffices for the conjectured rate on this family.\n\\end{theorem}\n\\begin{proof}\nA 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.\n\nEvery 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 $16v\\eta_0=\\alpha/4$.\n\nThe 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.\n\\end{proof}\n\n\\begin{corollary}[Maximum classes near Boolean anchors]\\label{cor:max-composite}\nLet $C$ 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\n$S_A=\\sum_{j=1}^r\\Lam(\\LD(A_j)+1)$.\nThen a realizable private PAC learner has sample complexity\n\\[\n\\widetilde O\\left[\n\\frac{v^7(q+\\log(v+1))}{\\alpha\\varepsilon}\n\\left(q+r^4S_A+r\\log(r+1)+\\log(k+1)\n+\\log\\frac{v}{\\beta\\delta}\\right)\\right].\n\\]\nIf the assignment is onto the full Boolean product, each component contains both constants, and the gate is essential in each input, then\n$S_A=O(r\\Lam(d+k+1))$.\n\\end{corollary}\n\\begin{proof}\nUse 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}.\n\nFor 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.\n\\end{proof}\n\nUnder 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.\n\n\n", "text_format": "latex"}
{"line_end": 937, "line_start": 898, "record_id": "orythmea-section-09", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 9, "section_title": "An arbitrary-class PAC transfer at polynomial radius", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{An arbitrary-class PAC transfer at polynomial radius}\n\nThe next result removes the maximum-class assumption, with polynomial rather than logarithmic dependence on $k$.\n\n\\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}\n\\begin{proof}\nAssume first $k\\ge1$ and set $R=100kM$. On a first block, release labeled categories $(x,y)$ of observed count $c_{x,y}$ when\n\\[\nc_{x,y}+\\Lap(2/\\varepsilon)>1+\\frac2\\varepsilon\\log\\frac8\\delta.\n\\]\nUse 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.\n\nThere 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\n\\[\nn_{\\rm heavy}=O\\left(\\frac R\\varepsilon\\log\\frac{R+2}{\\beta\\delta}\\right)\n\\]\nrecords 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\n$r_0\\le k/R=1/(100M)$.\n\nReserve $L=\\lceil\\log(4/\\beta)/\\log4\\rceil$ 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 $3/4$.\n\nThe 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\n$O((\\alpha\\varepsilon)^{-1}\\log(L/\\beta))$ 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 $\\beta/2$.\n\nEvery 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.\n\\end{proof}\n\nFor 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.\n\n", "text_format": "latex"}
{"line_end": 1063, "line_start": 938, "record_id": "orythmea-section-10", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 10, "section_title": "Exact limits of the universal extensions", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{Exact limits of the universal extensions}\n\nThe 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.\n\n\\subsection{Uniform marginal release is statistically too strong}\n\nFor even $N\\ge64$, let $X_N=\\{0,1\\}^N$ and\n\\[\nC_N=\\{h_i:i\\in[N]\\},\\qquad h_i(x)=x_i.\n\\]\nCardinality 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\n\\begin{equation}\\label{eq:evaluation-dims}\n\\VC(C_N)=\\LD(C_N)=\\lfloor\\log_2N\\rfloor,\\qquad\n\\VC(C_N^*)=N.\n\\end{equation}\nA law on $C_N$ is a law on $[N]$, and the query $x=\\1_A$ asks for its mass on an arbitrary subset $A\\subseteq[N]$.\n\n\\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}\n\\begin{proof}\nWrite $N=2m$. 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.\n\nGiven 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\n\\[\n\\operatorname{TV}(\\mu_{\\widehat\\sigma},\\mu_\\sigma)\n=\\frac{d_H(\\widehat\\sigma,\\sigma)}m\\le1/8.\n\\]\nThis argument permits completely inconsistent released answers.\n\nIf $n\\le m/2$, 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\n\\[\n\\Prb[d_H(\\widehat\\sigma,\\sigma)\\le m/8]\n\\le e^{-u/8}\\le e^{-m/16}<2/3.\n\\]\nThe averaged success probability is too small, so some law violates the guarantee.\n\\end{proof}\n\nThe 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 $v,\\Lam(d)$.\n\n\\subsection{Even global heavy consensus has the same obstruction}\n\nCall an output function $f$ a $\\rho$-consensus for a law $\\mu$ on $C$ if it respects every literal of probability at least $1-\\rho$:\n\\[\n\\Prb_{h\\sim\\mu}[h(x)=b]\\ge1-\\rho\\quad\\Longrightarrow\\quad f(x)=b.\n\\]\nThis asks for no numerical probability estimates.\n\n\\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}\n\\begin{proof}\nUse 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.\n\nLet $A$ contain all atoms selected by $\\sigma$ in unseen pairs and also\n$\\max\\{0,t-\\rho m\\}$ of the shared observed atoms. These counts are integers, and the second count is at most $\\rho m$ because $t\\le2\\rho m$. Therefore\n\\[\n\\mu_\\sigma(A)\\ge1-\\rho,\\qquad\n\\mu_{\\sigma'}(A)\\le\\rho.\n\\]\nA 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.\n\\end{proof}\n\nThis 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.\n\n\\subsection{Balanced selection can be exponentially easier}\n\nThe 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.\n\nIf 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 $[0.27,0.73]\\subset[1/4,3/4]$.\n\nThis 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\n\\[\n\\widetilde O(\\Lam(N)/\\varepsilon)\n=\\widetilde O(\\Lam(v+1)/\\varepsilon)\\quad(N=2^v).\n\\]\nThus 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 $v$.\n\n\\subsection{Small fixed-gate decompositions are not universal}\n\nThe 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\n\\[\nr\\log(r+1)=\\Omega(N),\\qquad\nr=\\Omega(N/\\log(N+1)).\n\\]\nThus $r=\\operatorname{poly}(v,\\Lam(d))$ cannot be guaranteed for every class. The obstruction is independent of component height and branching.\n\nThere 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\n$N=q^2+q+1$ 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.\n\nThe 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.\n\n\\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}\n\\begin{proof}\nEach 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 $p$.\n\nFor a suffix $B$ of size $s$, put $X_L=|L\\cap B|$. Incidence and pair counting give\n\\[\n\\sum_LX_L=s(q+1),\\quad\n\\sum_LX_L^2=s(q+s),\\quad\n\\sum_L\\left(X_L-\\frac{s(q+1)}N\\right)^2\n=\\frac{qs(N-s)}N.\n\\]\nIf $t(B)$ lines avoid $B$, their zero terms imply\n\\[\nt(B)\\le\\frac{qN(N-s)}{s(q+1)^2}\\le\\frac{qN}{s}.\n\\]\nFor 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.\n\\end{proof}\n\nAllowing 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.\n\nA 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\n\\[\n2^{\\,O(rN\\log N)+N2^r}\n\\]\nmatrices. 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.\n\nThese 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.\n\n\n", "text_format": "latex"}
{"line_end": 1132, "line_start": 1064, "record_id": "orythmea-section-11", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 11, "section_title": "Additional sparse-cover and privacy boundaries", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{Additional sparse-cover and privacy boundaries}\n\n\\subsection{A replicated-square obstruction}\n\n\\begin{proposition}\\label{prop:square}\nLet $C$ contain the four functions supported on\n\\[\n\\varnothing,\\qquad B_1,\\qquad B_2,\\qquad B_1\\cup B_2,\n\\]\nwhere $B_1,B_2$ 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 $4k\\ge t$.\n\\end{proposition}\n\\begin{proof}\nThe 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.\n\\end{proof}\n\nApply 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\n\\[\nk\\ge\\frac{\\lfloor m/2\\rfloor}{4}.\n\\]\nThis 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.\n\nConsequently, 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.\n\n\\subsection{A union of a few lower-dimensional anchor families also fails}\n\n\\begin{proposition}\\label{prop:many-anchors}\nLet $N=4m$. If every interval on $[N]$ is within distance $k$ of the union of $B$ arbitrary VC-one anchor classes, then\n\\[\nm^2\\le B(4m+1)(4k+1)^2.\n\\]\nIn particular, $\\log B+\\log(k+1)=\\Omega(\\log N)=\\Omega(d)$ on this maximum VC-two family.\n\\end{proposition}\n\\begin{proof}\nRestrict 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.\n\nSauer 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 $\\log B+\\log(k+1)\\ge\\frac12\\log m-O(1)$.\n\\end{proof}\n\nThis 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.\n\n\\subsection{Pure privacy cannot replace the approximate mechanisms}\n\nLet $C_k$ 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 $\\LD(C_k)=k$.\n\nHere 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\n\\[\n(s_{<i}1,s_i),\\qquad i=1,\\ldots,k.\n\\]\nThese 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\n$s_i=g(s_{<i}1)\\oplus e_i$.\nConsequently $g$ has risk at most $\\alpha$ on exactly\n$B_\\alpha=\\sum_{j=0}^{\\lfloor\\alpha k\\rfloor}\\binom kj$\nof these distributions.\n\nFix an arbitrary public length-$n$ dataset $S_0$. Pure group privacy and PAC confidence imply\n\\[\n\\Prb[M(S_0)\\text{ has }D_s\\text{-risk}\\le\\alpha]\n\\ge e^{-\\varepsilon n}(1-\\beta)\n\\]\nfor every $s$, after averaging over samples from $D_s$. Sum over $s$ and use the exact output-wise count above:\n$2^ke^{-\\varepsilon n}(1-\\beta)\\le B_\\alpha$.\nFor $0<\\alpha<1/2$, the binary entropy bound gives\n\\[\nn\\ge\n\\frac{(1-H_2(\\alpha))k\\log2+\\log(1-\\beta)}{\\varepsilon},\n\\]\nwhere $H_2$ 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.\n\n\n", "text_format": "latex"}
{"line_end": 1167, "line_start": 1133, "record_id": "orythmea-section-12", "record_type": "section", "research_status": "structural_results; universal_conjecture_open", "section_index": 12, "section_title": "Verification, provenance, and the remaining theorem", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\section{Verification, provenance, and the remaining theorem}\n\nThe 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.\n\nThe new suite passed the following checks:\n\\begin{center}\n\\begin{tabular}{lr}\n\\toprule\nNew finite check & Cases\\\\\n\\midrule\nConditional histogram transcripts & $4{,}344$\\\\\nPruned predicate errors & $66{,}448$\\\\\nGreedy-threshold boundary checks & $51$\\\\\nLaminar-family candidates & $107{,}520$\\\\\nSeparated laminar families & $6{,}926$\\\\\nCorrelated expected-error comparisons & $45{,}360$\\\\\nRaw-record versus product-atom evaluations & $230{,}400$\\\\\nExact finite-EM score-sensitivity checks & $3{,}318$\\\\\nFinite-EM likelihood inequalities & $9{,}954$\\\\\n\\bottomrule\n\\end{tabular}\n\\end{center}\nGeometry 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.\n\nThe 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.\n\nEstablished 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.\n\n\\paragraph{What is now proved.}\nA 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.\n\n\\paragraph{What remains unresolved.}\nA 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.\n\n\\newpage\n", "text_format": "latex"}
{"line_end": 1254, "line_start": 1168, "record_id": "orythmea-section-13", "record_type": "references", "research_status": "structural_results; universal_conjecture_open", "section_index": 13, "section_title": "References and document end", "source_path": "research/Orythmea_Private_Covers.tex", "source_sha256": "b87b71392797d2268bb7c9638cd1e5e80f5ab22fc807869b3067a414f9b2dd6c", "text": "\\begin{thebibliography}{99}\n\\bibitem{nst}\nK.~Nissim, U.~Stemmer, and E.~Tsfadia.\n\\emph{Invited Open Problem: Does Differential Privacy Make PAC Learning Much Harder?}\nCOLT 2026, PMLR 336:7129--7135.\n\\url{https://proceedings.mlr.press/v336/nissim26a.html}.\n\n\\bibitem{my}\nS.~Moran and C.~Yan.\n\\emph{Nearly Linear Private Learning of Littlestone Classes}.\narXiv:2610.23160, 2026.\n\\url{https://arxiv.org/abs/2610.23160}.\n\n\\bibitem{ds}\nP.~Dharangutte and A.~Sanyal.\n\\emph{New Lower Bounds for Private PAC Learning}.\narXiv:2610.41980, 2026.\n\\url{https://arxiv.org/abs/2610.41980}.\n\n\\bibitem{clnss}\nE.~Cohen, X.~Lyu, J.~Nelson, T.~Sarl\\'os, and U.~Stemmer.\n\\emph{$\\widetilde O$ptimal Differentially Private Learning of Thresholds and Quasi-Concave Optimization}.\nSTOC 2023.\n\\url{https://arxiv.org/abs/2211.06387}.\n\n\\bibitem{yan}\nC.~Yan.\n\\emph{An $\\widetilde O$ptimal Differentially Private Learner for Concept Classes with VC Dimension 1}.\narXiv:2505.06581v2, 2025.\n\\url{https://arxiv.org/abs/2505.06581}.\n\n\\bibitem{chase}\nZ.~Chase, B.~Chornomaz, S.~Hanneke, S.~Moran, and A.~Yehudayoff.\n\\emph{Dual VC Dimension Obstructs Sample Compression by Embeddings}.\nCOLT 2024, PMLR 247:923--946.\n\\url{https://proceedings.mlr.press/v247/chase24a.html}.\n\n\\bibitem{vc}\nV.~N.~Vapnik and A.~Ya.~Chervonenkis.\n\\emph{On the Uniform Convergence of Relative Frequencies of Events to Their Probabilities}.\nTheory of Probability and Its Applications 16(2):264--280, 1971.\n\\url{https://doi.org/10.1137/1116025}.\n\n\\bibitem{massart}\nP.~Massart.\n\\emph{The Tight Constant in the Dvoretzky--Kiefer--Wolfowitz Inequality}.\nAnnals of Probability 18(3):1269--1283, 1990.\n\\url{https://doi.org/10.1214/aop/1176990746}.\n\n\\bibitem{fx}\nV.~Feldman and D.~Xiao.\n\\emph{Sample Complexity Bounds on Differentially Private Learning via Communication Complexity}.\nCOLT 2014, PMLR 35:1000--1019.\n\\url{https://proceedings.mlr.press/v35/feldman14b.html}.\n\n\\bibitem{ggkm}\nB.~Ghazi, N.~Golowich, R.~Kumar, and P.~Manurangsi.\n\\emph{Near-tight Closure Bounds for the Littlestone and Threshold Dimensions}.\nALT 2021, PMLR 132.\n\\url{https://proceedings.mlr.press/v132/ghazi21a.html}.\n\n\n\\bibitem{blr}\nA.~Blum, K.~Ligett, and A.~Roth.\n\\emph{A Learning Theory Approach to Non-Interactive Database Privacy}.\nSTOC 2008, pp.~609--617.\n\\url{https://www.cs.cmu.edu/~avrim/Papers/privacy.pdf}.\n\n\\bibitem{abms}\nN.~Alon, A.~Beimel, S.~Moran, and U.~Stemmer.\n\\emph{Closure Properties for Private Classification and Online Prediction}.\nCOLT 2020, PMLR 125:119--152.\n\\url{https://proceedings.mlr.press/v125/alon20a.html}.\n\n\\bibitem{kt}\nD.~Karni and E.~Tsfadia.\n\\emph{Private Component-by-Component Learning}.\narXiv:2610.05102, 2026.\n\\url{https://arxiv.org/abs/2610.05102}.\n\n\\bibitem{klrss}\nS.~P.~Kasiviswanathan, H.~K.~Lee, K.~Nissim, S.~Raskhodnikova, and A.~Smith.\n\\emph{What Can We Learn Privately?}\nSIAM Journal on Computing 40(3):793--826, 2011.\n\\url{https://arxiv.org/abs/0803.0924}.\n\\end{thebibliography}\n\\end{document}\n", "text_format": "latex"}