\documentclass[11pt]{amsart} \usepackage[letterpaper,margin=1in,headheight=14pt]{geometry} \usepackage{amsmath,amssymb,mathtools,array} \usepackage[hidelinks]{hyperref} \emergencystretch=2em \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \numberwithin{equation}{section} \newcommand{\N}{\mathbb N} \newcommand{\Z}{\mathbb Z} \newcommand{\E}{\mathbb E} \newcommand{\one}{\mathbf{1}} \newcommand{\floor}[1]{\left\lfloor #1\right\rfloor} \newcommand{\norm}[1]{\lVert #1\rVert} \newcommand{\inner}[2]{\langle #1,#2\rangle} \newcommand{\psum}{\mathcal P} \newcommand{\indep}{\mathcal D} \title{Completeness of two floor-doubling sequences} \author{Jensen Kohlmeyer} \author{Liam Kruer} \date{September 11, 2026} \subjclass[2020]{Primary 11B13; Secondary 11B05, 37A05} \keywords{Erd\H{o}s problem 354, complete sequences, subset sums, binary carries, symbolic dynamics, rigidity, Lean formalization} \begin{document} \begin{abstract} We prove that, whenever $\alpha,\beta>0$ and $\alpha/\beta$ is irrational, every sufficiently large integer is a sum of distinct indexed terms of the two sequences $(\floor{2^n\alpha})_{n\ge0}$ and $(\floor{2^n\beta})_{n\ge0}$. This answers the doubling case of Erd\H{o}s Problem 354 affirmatively. The proof associates a symbolic tower system to each sequence and turns infinitely many missing sums into a nonproduct invariant coupling. Binary carries give two complementary estimates: many marked transitions force decay of correlations, whereas a single common marked transition controls one-step displacement by displacement at a prescribed large time. The latter estimate forces a forward relation between the positions of binary ones in the two parameters. Applying that relation in both directions bounds all zero runs, contradicting the correlation estimate. A complete Lean 4 formalization of the exact conjecture accompanies the paper; its final theorem has no admitted proof and uses only the standard axioms \texttt{propext}, \texttt{Classical.choice}, and \texttt{Quot.sound}. \end{abstract} \maketitle \section{Introduction} A sequence of nonnegative integers is called complete if its finite subsequence sums contain every sufficiently large integer. The number of summands is unrestricted, but each index may be used at most once. The question considered here appears in the study of complete sequences initiated by Graham~\cite{Graham} and is recorded by Erd\H{o}s and Graham~\cite[p.~58]{ErdosGraham}. It is Problem 354 in Bloom's catalogue~\cite{Bloom}. Earlier investigations include Hegyv\'ari~\cite{Hegyvari1}; his later discussion~\cite{Hegyvari2} also describes the connection with binary bracket sequences. \begin{theorem}\label{thm:main} Let $\alpha,\beta$ be positive real numbers such that $\alpha/\beta$ is irrational. There is an integer $N_0$ such that every integer $N\ge N_0$ has a representation \begin{equation} N=\sum_{s\in S}\floor{2^s\alpha}+\sum_{t\in T}\floor{2^t\beta}, \qquad S,T\subset\N_0\text{ finite}. \end{equation} \end{theorem} Here and throughout, $\N_0=\{0,1,2,\ldots\}$. The sets $S$ and $T$ are chosen independently. Thus two equal numerical values occurring at different indices remain two available terms. This is precisely the multiset convention in the formalized problem~\cite{Formal}. The theorem concerns the multiplier $2$. The separate question about a multiplier in $(1,2)$ is outside its scope. The binary recurrence is the starting point. Put \[ h_\gamma(n)=\floor{2^n\gamma},\qquad d_\gamma(n)=h_\gamma(n+1)-2h_\gamma(n)\in\{0,1\}. \] The digit $d_\gamma(n)$ records whether one spacer is inserted after doubling a word of length $h_\gamma(n)$. The resulting tower systems retain both the subset-sum information and the binary carries generated by large translations. The main estimate is a comparison of two displacement energies. If a large translation has a suitable marked carry at stage $m$, then every stage-$m$ observable $f$, bounded by $F$, satisfies \begin{equation} \norm{U^{-1}f-f}_2^2\le 16\norm{U^tf-f}_2^2+36F^2\varepsilon_m+676F^2(\ell+1)w_m. \end{equation} Here $U$ is the Koopman operator of the tower, $w_m=(2^m\gamma)^{-1}$, $\varepsilon_m=1-h_\gamma(m)w_m$, and $2^\ell$ bounds the relevant return query. The constants are absolute. A pair of three-bit carry paths produces the factor 16, and elementary interval counting controls both return regions simultaneously. This estimate is proved in Section~\ref{sec:common}. The final argument has a simple form. Failure of disjointness forces every sufficiently late binary one of $\alpha$ to be followed, at a fixed positive offset and within a bounded distance, by a one of $\beta$. Reversing the parameters gives a bounded forward return to the ones of $\alpha$. This bounds the zero runs of $\alpha$. But bounded zero runs already imply disjointness, by a second use of the carry paths. The contradiction proves completeness. All arguments needed for Theorem~\ref{thm:main}, including the construction of the symbolic systems and the estimates above, are implemented in the accompanying Lean development. Section~\ref{sec:formal} gives the exact target, verification information, and a guide to the proof files. \section{Binary heights and the symbolic tower model} We first work with a parameter $\gamma\ge1$, so that every height is positive. We suppress $\gamma$ from the notation when it is fixed. \begin{lemma}\label{lem:heights} For $n\ge0$, \begin{align} h_{n+1}&=2h_n+d_n,\quad d_n\in\{0,1\},\label{eq:heights}\\ h_n-\sum_{j0$. \section{Missing sums, couplings, and rigidity} A binary-name limit for $\gamma$ is a subsequential empirical limit of the shifts of $v_\gamma$. Given binary-name limits $\nu_\alpha$ and $\nu_\beta$, an anti-joining is a probability measure $\eta$ on $\{0,1\}^{\Z}\times\{0,1\}^{\Z}$ with those marginals, invariant under $\sigma\times\sigma^{-1}$. We write $\indep(\alpha,\beta)$ if every such anti-joining is the product of its marginals. Swapping coordinates and using inverse invariance shows that $\indep(\alpha,\beta)$ is symmetric. \begin{lemma}\label{lem:missing} If $\alpha,\beta\ge1$ and $\indep(\alpha,\beta)$ holds, then $\psum_\alpha+\psum_\beta$ contains all sufficiently large integers. \end{lemma} \begin{proof} If this failed, choose missing integers $N_r\to\infty$ and consider \[ \eta_r=\frac{1}{N_r+1}\sum_{t=0}^{N_r}\delta_{(\sigma^tv_\alpha,\sigma^{N_r-t}v_\beta)}. \] Compactness gives a convergent subsequence. The two endpoint terms are the only discrepancy between $\eta_r$ and its image under $\sigma\times\sigma^{-1}$, so the limit $\eta$ is an anti-joining. Its marginals are binary-name limits. Every $\eta_r$ gives mass zero to the clopen cylinder $\{(x,y):x(0)=y(0)=1\}$, since a point in that cylinder would represent $N_r$. Its limiting mass is also zero. The product of the marginals gives it mass $(1/\alpha)(1/\beta)>0$, a contradiction. \end{proof} We will rule out anti-joinings through their Hilbert-space operators. An anti-joining defines a contraction $J_0:L^2(\nu_\beta)\to L^2(\nu_\alpha)$ by \[ \inner{g}{J_0f}=\int g(x)f(y)\,d\eta(x,y). \] It preserves constants and integrals and satisfies $U_\alpha^kJ_0=J_0U_\beta^{-k}$. If it annihilates $L^2_0(\nu_\beta)$, then $\eta$ is product, first on rectangles and hence on the whole product sigma-algebra. These operators can be tested on the labeled systems of Section 2. Indeed, the factor maps induce isometric embeddings $I_\gamma:L^2(\nu_\gamma)\to L^2(\mu_\gamma)$. The operator $J=I_\alpha J_0I_\beta^*$ again preserves constants and integrals and intertwines $U_\alpha^k$ with $U_\beta^{-k}$. Moreover $I_\alpha^*JI_\beta=J_0$. It therefore suffices to prove that every such operator between tower-name limits annihilates the space of mean-zero functions. \begin{lemma}[Zero blocks]\label{lem:zero} Let $\mu$ be a tower-name limit for $\gamma\ge1$. If $d_\gamma(n+j)=0$ for $j+10$, eventually \begin{equation} \norm{U^{\pm h_\gamma(n_r)}f-f}_2^2\le\norm f_2^2+\epsilon.\label{eq:singlezero} \end{equation} For $f\ne0$, this implies eventually $\inner{f}{U^{\pm h_\gamma(n_r)}f}\ge\norm f_2^2/8$. \end{enumerate} \end{lemma} \begin{proof} Among the $2^L$ stage-$n$ copies in stage $n+L$, every copy before the last is followed immediately by the next one. Indeed $G_n(r)=rh_n$ for $r<2^L$ under the stated zero assumptions; the unrestricted last digit can only place a spacer after the last copy. Translation by $h_n$ preserves $f$ on these first $2^L-1$ copies. The integral of $f^2$ over their complement is at most $2^{-L}\norm f_2^2+F^2\varepsilon_n$. Since $\norm{U^{h_n}f}_2=\norm f_2$, the two functions have equal square integrals on the complement of their equality set. The inequality $|a-b|^2\le2a^2+2b^2$ now proves \eqref{eq:zero}. Refine a fixed earlier observable to stage $n_r$ and pass to the limit. Density gives (i) for every $L^2$ function. Taking $L=2$ proves (ii), again by density. The negative-time displacement has the same norm as the positive-time displacement. Finally use $\norm{U^tf-f}_2^2=2\norm f_2^2-2\inner{f}{U^tf}$ and choose $\epsilon>0$ sufficiently small. \end{proof} \begin{lemma}[Two disjointness criteria]\label{lem:criteria} For $\alpha,\beta\ge1$, either of the following conditions implies $\indep(\alpha,\beta)$. \begin{enumerate} \renewcommand{\labelenumi}{(\roman{enumi})} \item For every tower-name limit of $\alpha$, $\inner{g}{U_\alpha^{h_\beta(n)}f}\to0$ whenever $f$ has mean zero and $g\in L^2$. \item There are $n_r,L_r\to\infty$ with $d_\beta(n_r+j)=0$ for $j+10$, a bound of the form in Lemma~\ref{lem:amplify} exists with $s=1/2$: choose a stage beyond which the norm is at most $\epsilon$ and absorb the finitely many preceding values into $C$. It follows that arbitrarily many such blocks force the carry average to tend to zero. The conclusion passes to $P_{m,q}$ by first including all selected blocks in a finite word and then letting its remaining length tend to infinity. \begin{definition} The digits of $\alpha$ have bounded zero runs if there is an integer $H>0$ such that every interval $[a,a+H)\cap\N_0$ contains a position $j$ with $d_\alpha(j)=1$. \end{definition} Under this hypothesis, each transition $q_k\ne q_{k+1}$ starts a delayed pair of length at most $H+2$. Transitions separated by at least $H+2$ therefore yield disjoint marked blocks, with the uniform parameter $\delta=2^{-H-1}$. \begin{lemma}[Queries at the other height]\label{lem:query} Let $\alpha,\beta\ge1$, $n\to\infty$, and put \[ m=\floor{n/2},\quad L=n-m,\quad t=h_\beta(n),\quad G_{\alpha,m}(q)\le t0$, $\Delta\ge0$, and $0\le rh$, the first region is empty. The second image is the interval of length $h$ starting at $r-h-\Delta-c_1$. Its initial point lies between $-\Delta-c_1$ and $0$. The same cancellation estimate proves \eqref{eq:interval}. \end{proof} Let $a:\Z\to\mathbb R$ be zero outside $[0,h)$ and satisfy $|a|\le F$. For two costs $C_0,C_1$ on uniformly random $K$-bit words, put \begin{align} E_K&=\E\sum_{i=0}^{h-1}\bigl(a(i)-a(\Phi_{C_0,C_1}(i))\bigr)^2,\label{eq:ek}\\ D(a)&=\sum_{i=0}^{h-1}(a(i)-a(i-1))^2,\qquad \E(C_0+\Delta+C_1)\le B.\label{eq:da} \end{align} \begin{lemma}[A paired word detects one-step energy]\label{lem:pair} Suppose $K=k+3+s$. For each fixed prefix of length $k$ and suffix of length $s$, replacing a middle word $(b,0,e)$ by $(b,1,e)$ increases both $C_0$ and $C_1$ by exactly one. Then \begin{equation} D(a)\le16E_K+64F^2B.\label{eq:pair} \end{equation} One also has, without the paired-word assumption, \begin{equation} \E\sum_{i=0}^{h-1}a(\Phi_{C_0,C_1}(i))^2\le\sum_{i=0}^{h-1}a(i)^2+2F^2B.\label{eq:pairuncond} \end{equation} \end{lemma} \begin{proof} The coordinate identity $\Phi_{c_0+1,c_1+1}(i)=\Phi_{c_0,c_1}(i)-1$ holds in both regions, since the region is selected by $i+r0$ such that $d_c(b)\ne d_c(b+1)$. Suppose $a_r\to\infty$, $d_\alpha(a_r)=1$, and $n_r=a_r+b+1$. In every tower-name limit of $\alpha$, \[ U^{h_\beta(n_r)}f\longrightarrow f\quad\Longrightarrow\quad U^{-1}f=f. \] In particular a mean-zero vector with this rigidity property is zero. \end{proposition} \begin{proof} Set $m_r=\floor{n_r/2}$, $L_r=n_r-m_r$, and let $q_r$ be the return query for $t_r=h_\beta(n_r)$ at stage $m_r$. By Lemmas~\ref{lem:query} and~\ref{lem:stable}, eventually the bits of both $q_r$ and $q_r+1$ at positions $k_r=L_r-b-2$, $k_r+1$, $k_r+2$ are respectively $d_c(b+1)$, $d_c(b)$, $d_c(b-1)$. The first two differ, the three bits agree for the two queries, and \[ m_r+k_r+1=n_r-b-1=a_r. \] The common-mark assumptions therefore hold. For a fixed bounded tower observable $g$, refine it to stage $m_r$. Theorem~\ref{thm:common}, with $\ell=n_r+C$, gives \begin{equation} \norm{U^{-1}g-g}_2^2\le16\norm{U^{t_r}g-g}_2^2+R_r(g),\qquad R_r(g)\longrightarrow0,\label{eq:rigidity} \end{equation} by \eqref{eq:queryerrors}. The bound on the refined observable is independent of $r$. Now let $U^{t_r}f\to f$ and choose such a $g$ with $\norm{f-g}_2<\eta$. Isometry gives $\limsup_r\norm{U^{t_r}g-g}_2\le2\eta$. Equation \eqref{eq:rigidity} implies $\norm{U^{-1}g-g}_2\le8\eta$. Another triangle inequality yields $\norm{U^{-1}f-f}_2\le10\eta$. Density lets $\eta\downarrow0$. Ergodicity then gives the mean-zero conclusion. \end{proof} \section{Forward transport and the proof of completeness} Put $A_\gamma=\{n\ge0:d_\gamma(n)=1\}$. For sets $A,B\subset\N_0$ and an integer $s>0$, write $A\rightsquigarrow_s B$ if there are integers $M>0$ and $N$ such that \begin{equation} a\in A,\ a\ge N\quad\Longrightarrow\quad B\cap[a+s,a+s+M)\ne\varnothing.\label{eq:transport} \end{equation} The constants are uniform over all sufficiently late $a\in A$. \begin{proposition}[Non-disjointness forces transport]\label{prop:transport} If $\alpha,\beta\ge1$, $\alpha/\beta$ is irrational, and $\indep(\alpha,\beta)$ fails, then $A_\alpha\rightsquigarrow_s A_\beta$ for some $s>0$. \end{proposition} \begin{proof} Choose $b>0$ with $d_{\beta/\alpha}(b)\ne d_{\beta/\alpha}(b+1)$, possible by Lemma~\ref{lem:heights}. We claim $s=b+1$ works. If \eqref{eq:transport} fails for this $s$, then for each $r$ there is $a_r\ge r$ with $d_\alpha(a_r)=1$ and \[ d_\beta(a_r+b+1+j)=0\qquad(0\le j0$, then $A$ meets every interval of some fixed positive integer length in $\N_0$. \end{lemma} \begin{proof} If $M$ and $M'$ are the two interval lengths, composing the relations shows that every sufficiently late $a\in A$ has another point of $A$ in $[a+s+t,a+s+t+M+M')$. Start from one such $a_0\in A$ and iterate. The resulting sequence increases strictly and has successive gaps bounded by $s+t+M+M'$. It tends to infinity and therefore meets every sufficiently late interval of that length. Increasing the length to include the finite initial segment proves the result. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:main}] First suppose $\alpha,\beta\ge1$. If $\indep(\alpha,\beta)$ failed, at least one of $\alpha,\beta$ would be irrational, since their ratio is irrational. Exchange the parameters if necessary so that $\alpha$ is irrational. Then $A_\alpha$ is unbounded. Proposition~\ref{prop:transport} yields $A_\alpha\rightsquigarrow_s A_\beta$ for some $s>0$. Symmetry of $\indep$ and irrationality of $\beta/\alpha$ give $A_\beta\rightsquigarrow_t A_\alpha$ for some $t>0$. Lemma~\ref{lem:gaps} says that the digits of $\alpha$ have bounded zero runs. Proposition~\ref{prop:bounded} now gives $\indep(\alpha,\beta)$, a contradiction. Thus $\indep(\alpha,\beta)$ always holds, and Lemma~\ref{lem:missing} proves completeness in the normalized case. For arbitrary positive $\alpha,\beta$, choose $k\ge0$ such that $\alpha'=2^k\alpha$ and $\beta'=2^k\beta$ are at least $1$. Their ratio is unchanged. The normalized result gives the representation in Theorem~\ref{thm:main} for $\alpha',\beta'$. Since $h_{\alpha'}(n)=h_\alpha(n+k)$ and likewise for $\beta$, these representations use only terms of the original two sequences. This proves the theorem. \end{proof} \section{Lean formalization and reproducibility}\label{sec:formal} The accompanying source proves the exact declaration \texttt{Erdos354.erdos\_354.parts.i} from the pinned Formal Conjectures repository~\cite{Formal}. The challenge supplies its statement through the macro \texttt{fcTypeOfName\%}. The final declaration is named \texttt{Bounty.target}. Its expanded proposition is \[ \mathrm{True}\longleftrightarrow \bigl[\forall\alpha,\beta>0,\ \alpha/\beta\notin\mathbb Q\Longrightarrow\operatorname{Complete}(\alpha,\beta)\bigr], \] where $\operatorname{Complete}(\alpha,\beta)$ denotes the conclusion of Theorem~\ref{thm:main}. The library's completeness predicate uses finite sets of indices. It is applied to the interleaving of the $\alpha$ and $\beta$ floor sequences with multiplier $2$, exactly as in the theorem. The proof is written in Lean 4~\cite{Lean} using Mathlib~\cite{Mathlib}. Its single-file submission contains 601 theorem declarations, assembled from 130 local source files. The development constructs the compact symbolic spaces, empirical limits, factor and coupling operators, finite carry averages, and their limits. It proves the tower approximation, ergodicity, zero-block rigidity, carry amplification, and displacement estimates before assembling the final digit argument. \begin{table}[tbp] \caption{Guide to the main blocks of the formal proof.}\label{tab:lean} \small \centering \begin{tabular}{@{}>{\raggedright\arraybackslash}p{.37\linewidth}>{\raggedright\arraybackslash}p{.59\linewidth}@{}} \hline Mathematical component & Lean source blocks\\\hline Heights, digits, finite sums & \texttt{BinaryFloors}, \texttt{FiniteSums}, \texttt{FullReturnPositions}\\ Symbolic model and ergodicity & \texttt{TowerLabels}, \texttt{TowerMeasures}, \texttt{TowerGeneration}, \texttt{TowerErgodicity}\\ Missing-sum obstruction and operators & \texttt{SubsetSumCoding}, \texttt{TowerJoiningOperators}, \texttt{IntertwinerCriteria}\\ Zero blocks and correlations & \texttt{TowerRigidityIntegrals}, \texttt{TowerPartialRigidity}, \texttt{InfiniteAdaptiveCorrelation}\\ Bounded zero runs & \texttt{CarryAmplification}, \texttt{BoundedCarryDecay}, \texttt{BoundedZeroDisjointness}\\ The common-mark energy estimate & \texttt{SplitWindowEnergy}, \texttt{AdaptiveArrayCorrelations}, \texttt{MarkedTowerEnergy}\\ Rigidity exclusion and transport & \texttt{CrossHeightRigidity}, \texttt{TransportDisjointness}, \texttt{DigitTransport}\\ Assembly of the exact target & \texttt{ConcreteReduction}, \texttt{FullTarget}\\\hline \end{tabular} \end{table} In particular, the formal proof does not assume a disjointness theorem for a preexisting family of transformations. The systems used by the estimates are the actual empirical limits of the labeled subset-sum names. The main proof blocks are listed in Table~\ref{tab:lean}. The same names appear as source-block comments in the standalone file. The last step of the proof, after all auxiliary declarations have been supplied, is: \begin{verbatim} theorem target : fcTypeOfName% "Erdos354.erdos_354.parts.i" := by exact Erdos354Formal.full_target_of_symbolic_digit_criteria (fun _ _ ha hb hc hz => Erdos354Formal.symbolicallyDisjoint_of_boundedZeroRuns ha hb hc hz) (fun _ _ ha hb hc _ hn => Erdos354Formal.forwardTransport_of_not_symbolicallyDisjoint ha hb hc hn) \end{verbatim} The two arguments are the fully proved concrete digit criteria of Propositions~\ref{prop:bounded} and~\ref{prop:transport}. They are not additional hypotheses of the final theorem. \subsection{Verification record} The proof package records successful local checks with Lean 4.33.1 and the following repository revisions: \begin{center} \begin{tabular}{@{}ll@{}} \hline Repository & Commit\\\hline Formal Conjectures & \texttt{8432eac998110a563e03df65a28c117e97c8c142}\\ Mathlib & \texttt{0df444a360eaa60ab8c11dca51a86af692955474}\\ Challenge validator & \texttt{922a912f3777ac11451e2d145356e78de4dbedb2}\\\hline \end{tabular} \end{center} The standalone file compiles with \texttt{autoImplicit=false} and \texttt{warningAsError=true}. The exact challenge header, submission body, and footer also compile together. The pinned validator's static submission checks pass, and its target inspector reports that the intended and proved propositions coincide. The transitive axiom audit of the completed declaration reports: \begin{verbatim} 'Bounty.target' depends on axioms: [propext, Classical.choice, Quot.sound] \end{verbatim} In particular, \texttt{sorryAx} is absent. The upstream conjecture file contains an admitted declaration. It is imported to recover the target type, and that admitted proof is not a dependency of \texttt{Bounty.target}. No extra axioms are introduced by the submitted proof. The canonical target hash recorded by the inspector is \begin{verbatim} de4262d428cc1a0deb7df106c6951e4d b5d808b9a6d6a4e21c29f05775c50488 \end{verbatim} where the two printed lines are concatenated. The complete compiler and inspector output, source hashes, and pinned task files are included in the original companion package. \subsection{Files and verification command} In the original package, \texttt{Submit.lean} contains the submission body for the website. The site supplies its imports and outer \texttt{Bounty} namespace. \texttt{Challenge.lean} contains the complete standalone source for local use, including those commands. \texttt{Audit.lean}, \texttt{verify.py}, and \texttt{verification.json} provide the audit and its recorded results. With the pinned repositories and their built dependencies available, the original verification is reproduced by: \begin{verbatim} python verify.py --lean PATH_TO_LEAN \ --formal-conjectures PATH_TO_FORMAL_CONJECTURES \ --mathlib PATH_TO_MATHLIB --validator PATH_TO_VALIDATOR \end{verbatim} Here the backslashes denote shell line continuation. The command may also be entered on one line. The record establishes local kernel verification and compliance with the pinned static submission checks. A website submission or an official server verdict is not part of that record. A later public version in the submitted prize package includes a \href{https://github.com/JENW1N/complete-math-solutions/blob/af9feb6c310f5cea035e4ec34400c43fec01817e/proofs/erdos354/KohlmeyerKruerPartI.lean}{direct base-two Lean endpoint} and a \href{https://github.com/JENW1N/complete-math-solutions/blob/af9feb6c310f5cea035e4ec34400c43fec01817e/proofs/erdos354/verification.json}{separate build and axiom record}. See the \href{https://github.com/JENW1N/complete-math-solutions/tree/af9feb6c310f5cea035e4ec34400c43fec01817e/proofs/erdos354}{pinned reproduction guide} for that Mathlib-only package and its attribution for the separate variable-base results. \subsection*{Acknowledgments} We acknowledge substantial assistance from OpenAI's Codex AI assistant in the mathematical exploration, development and revision of the argument, construction of the Lean formalization, and preparation of this manuscript. Formal proof checking was performed by the Lean kernel. We also acknowledge the Lean, Mathlib, and Formal Conjectures contributors whose software and formal statements support the accompanying verification. \begin{thebibliography}{8} \bibitem{Bloom} T. F. Bloom, \emph{Erd\H{o}s Problem \#354}, \url{https://www.erdosproblems.com/354}. \bibitem{ErdosGraham} P. Erd\H{o}s and R. L. Graham, \emph{Old and new problems and results in combinatorial number theory}, Monographies de L'Enseignement Math\'ematique 28, Universit\'e de Gen\`eve, Geneva, 1980. \bibitem{Formal} The Formal Conjectures Authors, \emph{Formal Conjectures}, file \texttt{FormalConjectures/ErdosProblems/354.lean}, \url{https://github.com/google-deepmind/formal-conjectures}. The exact revision used here is recorded in Section~\ref{sec:formal}. \bibitem{Graham} R. L. Graham, \emph{On sums of integers taken from a fixed sequence}, Proceedings of the Washington State University Conference on Number Theory (1971), 22--40. \url{https://mathweb.ucsd.edu/~ronspubs/71_08_integer_sums.pdf}. \bibitem{Hegyvari1} N. Hegyv\'ari, \emph{Some remarks on a problem of Erd\H{o}s and Graham}, Acta Math. Hungar. 53 (1989), 149--154. \bibitem{Hegyvari2} N. Hegyv\'ari, \emph{On Boolean functions defined on bracket sequences}, Discrete Math. Lett. 12 (2023), 45--49. \url{https://doi.org/10.47443/dml.2022.209}. \bibitem{Lean} L. de Moura and S. Ullrich, \emph{The Lean 4 theorem prover and programming language (system description)}, Automated Deduction---CADE 28 (2021). \url{https://lean-lang.org/papers/lean4.pdf}. \bibitem{Mathlib} The mathlib Community, \emph{The Lean mathematical library}, Proceedings of the 9th ACM SIGPLAN International Conference on Certified Programming and Proofs (2020). \url{https://arxiv.org/abs/1910.09336}. \end{thebibliography} \end{document}