{"id": "section-1", "record_type": "manuscript_section", "title": "Aggregation laws", "content": "\\section{Aggregation laws}\nCoefficient compatibility is governed by coupled quadratic relations. A substitution adds coefficients and may change their global compatibility, even when every original relation holds strictly. We establish exact low-degree preservation laws, a sharp cardinality boundary, and constructive weak repair in every degree.\n\n\\begin{theorem}[Principal results]\\label{thm:main}\nLet $q>0$.\n\\begin{enumerate}\n\\item For any symmetric nonnegative matrix $C$ satisfying all $\\T_q$ quartet inequalities and every nonnegative rectangular matrix $A$, the matrix $A^{\\mathsf T}CA$ satisfies all quartet inequalities at the exact optimal universal parameter $\\max\\{q,1\\}$. This controls the full normalized quadratic polynomial under arbitrary nonnegative linear substitutions.\n\\item One merge of rank-$d$ matroid data on $n$ labels preserves strong representability at the same parameter precisely in the regimes in Table~\\ref{tab:threshold}. The same classification holds for universal strong-to-weak preservation.\n\\item Every weak $\\T_q$ matroid input, including zero coefficients, has full normalized weak $\\T_{2\\max(q,1)}$ output after one merge. The output parameter is exactly optimal for $00$\\\\\n$d\\ge2$ & $d\\le n\\le d+1$ & Every $q>0$\\\\\n$d=2$ & $n\\ge4$ & Exactly $q\\ge1$\\\\\n$d\\ge3$ & $n=d+2$ & Exactly $q\\ge1$\\\\\n$d\\ge3$ & $n\\ge d+3$ & No $q>0$\\\\\\bottomrule\n\\end{tabular}\n\\caption{Complete classification for one full normalized merge.}\\label{tab:threshold}\n\\end{table}\n\nA full polynomial and its squarefree projection are distinguished throughout. Normalization is explicit, including repeated monomials. The sparse rank-three pattern behind one obstruction family is credited to Proposition 8.5 of \\cite{BHKL}; every construction used here is proved directly. Historical priority of the new theorem package is not asserted.", "source_section": "1", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"1\"}"} {"id": "section-2", "record_type": "manuscript_section", "title": "Conventions, normalization, and elementary tools", "content": "\\section{Conventions, normalization, and elementary tools}\\label{sec:definitions}\n\\subsection{Strong arrays and weak relations}\nFor $q>0$, a finite list $t_1,\\ldots,t_m\\ge0$ is null over $\\T_q$ when\n\\begin{equation}\\label{eq:null}\n t_k^{1/q}\\le\\sum_{\\ell\\ne k}t_\\ell^{1/q}\\qquad\\text{for every }k.\n\\end{equation}\nA matroid basis family is a nonempty family of equally sized sets satisfying basis exchange: for bases $B,D$ and $i\\in B\\setminus D$, some $j\\in D\\setminus B$ makes $B-i+j$ a basis. For a rank-$d$ coefficient function $c$ on the $d$-subsets of a ground set $E$, strong representability requires the support to be a matroid and requires \\eqref{eq:null} for\n\\begin{equation}\\label{eq:strong}\n \\bigl(c_{I\\cup\\{j\\}}c_{J\\setminus\\{j\\}}:j\\in J\\setminus I\\bigr),\n \\qquad |I|=d-1,\\quad |J|=d+1.\n\\end{equation}\nThe weak relations are the three-term subfamily: for $|K|=d-2$ and four distinct elements $i,j,k,\\ell\\notin K$, the products are\n\\begin{equation}\\label{eq:weak}\n c_{Kij}c_{Kk\\ell},\\qquad c_{Kik}c_{Kj\\ell},\\qquad c_{Ki\\ell}c_{Kjk}.\n\\end{equation}\nConcatenation denotes adjoining distinct indices. All inequalities in this paper are unsigned: the additive inverse of a positive element in this hyperfield equals itself.\n\nIf $m=|J\\setminus I|=2$, the two terms in \\eqref{eq:strong} are identical. ``Strict strong'' always means strictness for $m\\ge3$; it never means making a tautological equality strict.\n\n\\begin{lemma}[Monotonicity and rescaling]\\label{lem:basic}\nIf a list is null at $q$, it is null at every $p\\ge q$. Multiplying coefficients by $\\lambda\\prod_{i\\in S}d_i$, for $\\lambda,d_i>0$, preserves all strong and weak relations and their strictness.\n\\end{lemma}\n\\begin{proof}\nFor monotonicity put $\\theta=q/p\\le1$ and use $(\\sum a_i)^\\theta\\le\\sum a_i^\\theta$ for nonnegative $a_i$. In a fixed strong relation, every product acquires the common factor $\\lambda^2\\prod_{i\\in I}d_i\\prod_{j\\in J}d_j$. The same argument covers weak relations.\n\\end{proof}\n\n\\subsection{What the normalized output means}\nFor a multi-affine input $F=\\sum_{|S|=d}c_Sx^S$, set $G=D_{u,v}F$. Its coefficient function in exponential form is\n\\begin{equation}\\label{eq:rho}\n \\rho_\\beta=\\beta!\\,[y^\\beta]G,\n \\qquad G=\\sum_\\beta\\rho_\\beta\\frac{y^\\beta}{\\beta!}.\n\\end{equation}\nFor squarefree $\\beta$ the factorial equals one. If $A$ avoids $u,v$ and $z$ denotes their merge, then\n\\begin{equation}\\label{eq:output}\n \\rho_A=c_A,\\qquad\n \\rho_{Az}=c_{Au}+c_{Av},\\qquad\n \\rho_{A+2z}=2c_{Auv},\n\\end{equation}\nwhenever the indices have total degree $d$.\n\nA fixed-degree integer set $J$ is M-convex if whenever $x_i>y_i$ for $x,y\\in J$, there is a $j$ with $x_j0$. Since row $j$ is nonzero, choose $\\ell$ with $C_{j\\ell}>0$. The quartet $(i,k,j,\\ell)$ has exactly one positive product, a contradiction. Thus $\\sim$ is an equivalence relation. If $i\\ne j$ are in one class, choose a positive neighbor of $j$; the quartet $(i,i,j,\\ell)$ forces $C_{ii}=0$. Positive diagonal entries therefore occur only in singleton classes.\n\nThe support consists exactly of degree-two integer vectors having total at most one on every zero-diagonal class, and total at most two on each positive-diagonal singleton. These partition-capacity bases satisfy symmetric exchange. Indeed, for $x_i>y_i$, first choose a deficit coordinate in the same block if one exists. Otherwise $x$ has greater total than $y$ on that block, so choose a deficit coordinate in a block where $x$ has smaller total. The exchange respects both capacities for both vectors. Hence the support is M-convex.\n\\end{proof}\n\nFor a usual quadratic weak representation, the capacity-admissible quartet conditions imply the unrestricted matrix conditions defining $\\mathcal C_q$. If no coordinate is mandatory, an excluded repetition has zero diagonal at a capacity-one coordinate or a zero row at a capacity-zero coordinate; its list is zero, an equal pair, or an equal triple. If a coordinate is mandatory, the degree-two support is a star or one monomial, and every unrestricted quartet is again one of these tautological types. Thus the following corollary covers all quadratic weak representations, rather than just a restricted matrix subclass.\n\n\\begin{corollary}[Exact quadratic substitution calculus]\nWrite $F(x)=\\tfrac12x^{\\mathsf T}Cx$ with normalized quadratic coefficient matrix $C\\in\\mathcal C_q$. For every nonnegative linear substitution $x=Ay$, the full normalized output matrix is exactly $A^{\\mathsf T}CA$ and if nonzero, its weak and strong parameter is universally repaired by $p=\\max\\{q,1\\}$. No parameter loss accumulates under any finite composition once the parameter is at least one.\n\\end{corollary}\n\\begin{proof}\nThe matrix identity follows by substitution. The support lemma supplies M-convexity. All rank-two strong relations have at most three terms, so the quartet conditions give both weak and strong representations. A composition of nonnegative matrices is nonnegative; apply Theorem~\\ref{thm:congruence} to their product.\n\\end{proof}", "source_section": "3", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"3\"}"} {"id": "section-4", "record_type": "manuscript_section", "title": "Uniform obstructions in every rank and size", "content": "\\section{Uniform obstructions in every rank and size}\\label{sec:uniform}\nChoose six distinguished labels $1,\\ldots,6$ and a disjoint set $C$ of size $d-3$. The ground set may contain any additional labels. Put\n\\[\n U=C\\cup135,\\qquad Z_1=C\\cup126,\\qquad Z_2=C\\cup346.\n\\]\n\n\\begin{lemma}[Three exceptional bases]\\label{lem:uniform}\nFor $10$, the coefficient vector $c_S=a_S^q$ is strict strong over $\\T_q$ on $U_{d,n}$. Every genuine root-polygon slack is at least\n\\begin{equation}\\label{eq:slack}\n \\delta=\\min\\{e^2,2-b\\}.\n\\end{equation}\n\\end{lemma}\n\\begin{proof}\nAfter taking roots, each strong term is a product of two $a$'s. In a relation with $m\\ge3$ terms, each exceptional coefficient can occur at most once among the factors: two different first factors are distinct, two different second factors are distinct, and a first factor cannot equal a second factor when $m\\ge3$. Every term is at most two; a term containing a low coefficient is at most $be<1$.\n\nFor $m=3$, factors from different terms intersect in $d-1$ elements. The three exceptional bases intersect pairwise in $d-2$ elements. Consequently any two exceptional factors occurring in the relation must occur in the same term. At least two terms therefore equal one.\n\nFor $m=4$, if fewer than three terms are exceptional there are again two unit terms. If all three exceptional bases occur, the union $I\\cup J$ has size $d+3$ and equals $C\\cup[6]$. Complement each basis inside this union. The dual relation has rank three and its exceptional triples are a peak at $246$ and lows at $125,345$. The permutation $(12)(34)(56)$ takes them to the original peak at $135$ and lows at $126,346$. The rank-three four-term products have at least two unit terms. To see this directly, complementary products at the sparse limit have possible lists\n\\[\n (1,1,1,1),\\ (0,1,1,1),\\ (0,0,1,1),\\ (1,1,1,2),\\ (0,1,1,2).\n\\]\nThe two zeros can occur together only when $I=12$ or $34$, in which case the peak product cannot occur. Restoring positive lows and replacing the peak root by $b$ retains the two unit terms.\n\nFor $m\\ge5$, there are at most three exceptional terms, so at least $m-3\\ge2$ unit terms remain. Thus in every case a peak term has slack at least $2-b$. Without a peak term the largest term is at most one, two unit terms remain, and at least one additional term is positive and at least $e^2$. This proves \\eqref{eq:slack} and strictness.\n\\end{proof}\n\n\\begin{theorem}[Explicit all-rank obstruction]\\label{thm:explicit}\nFor every $q>0$, define\n\\begin{equation}\\label{eq:parameters}\n T=2^q,\\quad D=2^{-q}-4^{-q},\\quad\n s=\\frac{qD}{64(q+2)},\\quad\n L=\\frac{3D^2}{128(q+2)},\\quad b=2-\\frac{L}{16}.\n\\end{equation}\nUse the root array of Lemma~\\ref{lem:uniform} with $e=1/4$ and put\n\\begin{equation}\\label{eq:explicit}\n c_S=a_S^q s^{\\one_{6\\in S}}.\n\\end{equation}\nFor every $d\\ge3,n\\ge d+3$, this is positive strict strong on $U_{d,n}$, whereas merging $5,6$ fails a squarefree weak relation by at least $L/2$ in root lengths.\n\\end{theorem}\n\\begin{proof}\nSince $00.\n\\]\nExtra labels never enter these six coefficients, proving the claim in every $d,n$.\n\\end{proof}\n\n\\begin{corollary}[Explicit robustness independent of dimension]\\label{cor:robust}\nLet $c$ be \\eqref{eq:explicit} and put $\\eta=L/2048$. Every positive array $c'$ satisfying\n\\[\n 1-\\eta\\le(c'_S/c_S)^{1/q}\\le1+\\eta\\qquad\\text{for every }S\n\\]\nis still strict strong and still fails after the same merge. Its normalized input root-polygon slack is at least $3L/256$, and its output defect is at least $3L/32$.\n\\end{corollary}\n\\begin{proof}\nFactor the common diagonal-rescaling multiplier from each input relation. Each root product changes by a factor in $[(1-\\eta)^2,(1+\\eta)^2]$. Its original maximum is at most two and its slack at least $L/16$. For any chosen side the new slack is at least\n\\[\n (1-\\eta)^2\\frac L{16}-8\\eta\n \\ge\\frac L{64}-\\frac L{256}=\\frac{3L}{256}.\n\\]\nThe same relative bounds hold on the roots of merged coefficient products: sums preserve the coefficient bounds before taking roots. The sum of the two smaller original output root lengths is less than $16$, using the bounds in the preceding proof. The new defect is therefore at least\n\\[\n (1-\\eta)^2\\frac L2-4\\eta\\cdot16\n \\ge\\frac L8-\\frac L{32}=\\frac{3L}{32}.\n\\]\n\\end{proof}", "source_section": "4", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"4\"}"} {"id": "section-5", "record_type": "manuscript_section", "title": "A guaranteed repair for the full normalized output", "content": "\\section{A guaranteed repair for the full normalized output}\\label{sec:repair}\nThe negative result concerns keeping $q$ unchanged. The next theorem shows that controlled weakening of the output condition always repairs \\emph{weak} compatibility for positive uniform input.\n\n\\begin{theorem}[One-merge repair]\\label{thm:repair}\nLet $c_S>0$ for every $d$-subset of $[n]$, and suppose $c$ is weak over $\\T_q$. For $r=\\max\\{q,1\\}$, the normalized coefficient function \\eqref{eq:rho} of $D_{u,v}F$ is weak over $\\T_{2r}$. If $d=2$, $\\T_r$ suffices. Every larger output parameter also suffices.\n\\end{theorem}\n\\begin{proof}\nThe support is the capped simplex described in Section~\\ref{sec:definitions}, so it remains to verify all tests \\eqref{eq:generalweak}. Let $s_0=\\alpha_z$ and let $k_0$ be the multiplicity of $z$ in the quartet. Capacities give $s_0+k_0\\le2$, so there are exactly six cases.\n\n\\textbf{Case $(s_0,k_0)=(0,0)$.} The products are an original input relation. They remain null at $2r\\ge q$ by monotonicity.\n\n\\textbf{Case $(0,1)$.} Write $K$ for the squarefree common index. Each output product is a sum of two products, one using $u$, one using $v$. For example,\n\\[\n c_{Kij}(c_{Kku}+c_{Kkv})=\\sum_{h\\in\\{u,v\\}}c_{Kij}c_{Kkh}.\n\\]\nFor each fixed $h$, the three $q$th roots form a triangle by the input relation on $i,j,k,h$. Lemma~\\ref{lem:minkowski} gives the output triangle at $p=r$, and hence at $2r$.\n\n\\textbf{Case $(0,2)$.} Let $K$ be the common index and $i,j$ the two nonmerged quartet labels. Put $a=c_{Kij}$, $w=c_{Kuv}$, and $b_i=c_{Kiu}+c_{Kiv}$, with $b_j$ defined similarly. The input quartet $i,j,u,v$ gives\n\\[\n (aw)^{1/q}\\le(c_{Kiu}c_{Kjv})^{1/q}+(c_{Kiv}c_{Kju})^{1/q}.\n\\]\nUsing $(x+y)^q\\le2^{r-1}(x^q+y^q)$ for $q\\ge1$, and subadditivity for $q\\le1$, yields\n\\[\n 2aw\\le2^r(c_{Kiu}c_{Kjv}+c_{Kiv}c_{Kju})\\le2^r b_i b_j.\n\\]\nThe output products are $2aw,b_i b_j,b_i b_j$. Thus they form a null list already at $p=r$.\n\n\\textbf{Case $(1,0)$: the only parameter-doubling case.}\nWrite $\\alpha=C+z$, $|C|=d-3$, and let the quartet be $i,j,k,\\ell$. Define\n\\[\n a^h_{ij}=c_{Cijh}^{1/q}\\quad(h=u,v),\\qquad w_i=c_{Ciuv}^{1/q}>0.\n\\]\nApply the input weak relation with common index $C\\cup\\{u\\}$ and quartet $i,j,k,v$. One of its triangle inequalities is\n\\[\n a^u_{ij}w_k\\le a^u_{ik}w_j+a^u_{jk}w_i.\n\\]\nThe analogous statement holds for $v$. Hence\n\\[\n d^h_{ij}=\\frac{a^h_{ij}}{w_iw_j},\\qquad d^h_{ii}=0,\n\\]\nare metrics on the unmerged labels under consideration. Positivity is essential here and is part of the theorem's hypotheses. For $q\\ge1$ their pointwise $\\ell_q$ aggregate\n\\[\n D_{ij}=((d^u_{ij})^q+(d^v_{ij})^q)^{1/q}\n\\]\nis a metric by Minkowski. For $q\\le1$, each $(d^h)^q$ is a metric, so $D_{ij}=(d^u_{ij})^q+(d^v_{ij})^q$ is a metric. In both cases the merged coefficients satisfy\n\\begin{equation}\\label{eq:metricfactor}\n b_{ij}:=c_{Ciju}+c_{Cijv}=(w_iw_j)^qD_{ij}^{r}.\n\\end{equation}\nAt parameter $p=2r$, the root product associated with a pairing of the quartet equals\n\\[\n (b_{ij}b_{k\\ell})^{1/(2r)}\n =(w_iw_jw_kw_\\ell)^{q/(2r)}\\sqrt{D_{ij}D_{k\\ell}}.\n\\]\nThe common positive multiplier cancels. Lemma~\\ref{lem:snowflake} supplies precisely all three required inequalities.\n\n\\textbf{Case $(1,1)$.} Write $\\alpha=C+z$ and use quartet $i,j,k,z$. Set $W_i=c_{Ciuv}$ and $B_{jk}=c_{Cjku}+c_{Cjkv}$. The three output products are\n\\[\n 2W_iB_{jk},\\quad2W_jB_{ik},\\quad2W_kB_{ij}.\n\\]\nFor each fixed $h=u,v$, the products $W_i c_{Cjkh}$, $W_j c_{Cikh}$, $W_k c_{Cijh}$ form an input weak relation with common index $C\\cup\\{h\\}$. Aggregate the two lists using Lemma~\\ref{lem:minkowski} and discard the shared factor two. This proves nullity at $p=r$.\n\n\\textbf{Case $(2,0)$.} The common index has the form $C+2z$. Each coefficient is twice an original coefficient with $u,v$ present. All three products are multiplied by four, and the input weak relation with common index $C\\cup\\{u,v\\}$ applies.\n\nThese six cases exhaust all admissible tests. In degree two, $|\\alpha|=0$, so only the first three cases occur and $p=r$ suffices throughout. Monotonicity proves the assertion for larger $p$.\n\\end{proof}\n\n\\begin{theorem}[Arbitrary matroid support]\\label{thm:zeros}\nThe normalized output of one merge of any weak $\\T_q$ matroid coefficient array is weak over $\\T_{2\\max\\{q,1\\}}$. In degree two, $\\T_{\\max\\{q,1\\}}$ suffices, with strong output as well.\n\\end{theorem}\n\\begin{proof}\nThe output support is M-convex by Lemma~\\ref{lem:support} below. The five elementary cases in the preceding proof allow zeros unchanged. Consider its critical case $(1,0)$ and fix its quartet. Retain the notation $a^h_{ij}=c_{Cijh}^{1/q}$ and $w_i=c_{Ciuv}^{1/q}$.\n\nIf every $w_i>0$, the preceding metric argument applies directly with pseudometrics: its divisions are valid, and vanishing distances do not affect Minkowski or Lemma~\\ref{lem:snowflake}. Hence assume that at least one weight vanishes.\n\nSuppose some $w_i$ is positive. The weighted triangle inequalities imply that an edge between two zero-weight vertices has $a^h_{ij}=0$: use a positive third vertex. They also imply that for a zero-weight vertex $i$ and positive-weight vertices $j,k$, $a^h_{ij}/w_j=a^h_{ik}/w_k=:A_{i,h}$, by applying both directions of the same inequality. If at least two weights vanish, the output products are either all zero or $(0,B,B)$, so are null at every parameter.\n\nIf exactly one weight, say $w_i$, vanishes, put $M=A_{i,u}^q+A_{i,v}^q$. For positive vertices $j$ we have $b_{ij}=w_j^q M$. On the other three vertices the aggregate metric $D$ in \\eqref{eq:metricfactor} is defined as before, allowing zero distances. If $M=0$ all output products vanish. Otherwise each output root product at parameter $2r$ is a common positive multiple of one of $\\sqrt{D_{jk}},\\sqrt{D_{j\\ell}},\\sqrt{D_{k\\ell}}$. These form a triangle, since square roots of metric distances satisfy the triangle inequality. Lemma~\\ref{lem:snowflake} also holds for pseudometrics by the same proof.\n\nIt remains to treat $w_i=0$ at all four vertices. The input weak relation with common index $C\\cup\\{i\\}$ and quartet $j,k,u,v$ now has two nonzero candidate products, and forces\n\\[\n c_{Ciju}c_{Cikv}=c_{Cijv}c_{Ciku}.\n\\]\nThus nonzero coefficient two-vectors on adjacent edges are proportional. Suppose two disjoint nonzero edge vectors $v_{ij}=(c_{Ciju},c_{Cijv})$ and $v_{k\\ell}$ were not proportional. Every bridge vector $v_{ik},v_{i\\ell},v_{jk},v_{j\\ell}$ is adjacent to both and must be zero. Choose nonzero coordinates $h,g\\in\\{u,v\\}$ to obtain actual matroid bases $C\\cup\\{i,j,h\\}$ and $C\\cup\\{k,\\ell,g\\}$. Basis exchange removing $i$ from the first basis is impossible: exchanging with $k$ or $\\ell$ gives a zero bridge coefficient, and exchanging with $g\\ne h$ gives $c_{Cjuv}=0$. This contradicts the support matroid's basis-exchange axiom.\n\nAll six nonnegative edge vectors therefore lie on one ray (zero vectors allowed). If every vector is zero there is nothing to prove. Otherwise choose a component $h$ positive on that ray. There is a single scalar $\\lambda>0$ such that $b_{ij}=\\lambda c_{Cijh}$ for every edge. The output relation is a rescaled input weak relation, valid already at $q$ and hence at $2r$. This completes all zero cases without a density assumption. Degree two never uses this critical case, so its earlier bound holds for every rank-two matroid.\n\\end{proof}\n\n\\begin{remark}[Scope of the repair]\nThe output is weak; strong output in degree at least three is not proved. The arbitrary-support extension uses the matroid support axiom in its final zero case; no positivity or strong-input hypothesis is needed. Weighted two-variable merges are covered by absorbing their positive weights into diagonal input rescalings.\n\\end{remark}", "source_section": "5", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"5\"}"} {"id": "section-6", "record_type": "manuscript_section", "title": "Sharp thresholds and asymptotically optimal parameter cost", "content": "\\section{Sharp thresholds and asymptotically optimal parameter cost}\\label{sec:optimal}\n\\subsection{Determinants as certified strong inputs}\n\\begin{lemma}\\label{lem:det}\nFor a real rank-$d$ matrix $A$ and any $q>0$, $c_S=|\\det A_S|^q$ is strong over $\\T_q$ on its column matroid.\n\\end{lemma}\n\\begin{proof}\nThe signed determinant Pl\\\"ucker relation sums to zero. The absolute value of each term is at most the sum of the others by the ordinary triangle inequality. These absolute values are exactly the $q$th roots of the products in \\eqref{eq:strong}.\n\\end{proof}\n\nTake four planar points $u_i\\in\\R^2$, their affine lifts $(u_i,1)$, and two additional columns $e_1,e_2\\in\\R^3$. After merging the last two labels,\n\\begin{equation}\\label{eq:planar}\n b_{ij}=|u_{i,1}-u_{j,1}|^q+|u_{i,2}-u_{j,2}|^q.\n\\end{equation}\nThis directly verifies the required output coefficients; no metric theorem is needed for the lower bounds below.\n\n\\begin{proposition}[Necessary output parameter]\\label{prop:lower}\nAny universal parameter $p$ for one merge in degree at least three, even restricted to positive uniform strong $\\T_q$ inputs, must satisfy\n\\begin{equation}\\label{eq:lower}\n p\\ge\\max\\{2,2(q-1)\\}.\n\\end{equation}\nMoreover, in each fixed degree $d\\ge3$ and size $n\\ge d+3$, the infimal universal parameter is strictly greater than $q$.\n\\end{proposition}\n\\begin{proof}\nFor the planar unit square, \\eqref{eq:planar} gives adjacent coefficients one and diagonal coefficients two, independently of $q$. The three output products are $(1,4,1)$. The output $\\T_p$ relation holds exactly when $4^{1/p}\\le2$, or $p\\ge2$.\n\nFor the coordinate diamond $(1,0),(0,1),(-1,0),(0,-1)$ the three products are $(4,4^q,4)$. When $q>1$, the necessary inequality $4^{q/p}\\le2\\cdot4^{1/p}$ is equivalent to $p\\ge2(q-1)$; for $q\\le1$ this bound is vacuous.\n\nThese matrices may have zero minors. Whenever a proposed bound is violated, the failed output inequality is strict. A sufficiently small generic matrix perturbation makes every maximal minor nonzero and preserves this strict failure, while Lemma~\\ref{lem:det} certifies strong input at the fixed $q$. To realize every $d,n$, first add $d-3$ identity columns and rows, then any extra columns, and perturb generically. The selected quartet minors are nonzero before perturbation, so the same failure persists.\n\nFinally, Theorem~\\ref{thm:explicit} provides, at each fixed $q,d,n$, a witness with a positive output defect at parameter $q$. That witness continues to fail for all output parameters in some interval $[q,q+\\eps]$ by continuity. Thus the infimum of universal working parameters is at least $q+\\eps>q$.\n\\end{proof}\n\n\\subsection{Degree two is completely determined}\n\\begin{theorem}[Exact degree-two answer]\\label{thm:d2}\nFor degree-two matroid input, the infimal universal output parameter for one merge is\n\\[\n P_2(q)=\\max\\{q,1\\}.\n\\]\nThe output is both weak and strong at this parameter. Consequently the same-parameter threshold in degree two is exactly $q_2=1$.\n\\end{theorem}\n\\begin{proof}\nThe upper bound is Theorems~\\ref{thm:repair} and \\ref{thm:zeros}. In rank two every strong relation has at most three summands, so weak and strong coincide. Equivalently, polarize the output by replacing the capacity-two coordinate $z$ with two clones. Its coefficient on the two clones is $\\rho_{2z}$ and its coefficients on a clone and $i$ are $\\rho_{zi}$. Every rank-two strong test is a three-term quartet, including the repeated-$z$ test already checked.\n\nFor $p\\ge q$ necessity use four real columns $(1,0),(0,1),(1,1),(1,-1)$: their Pl\\\"ucker product magnitudes are $(2,1,1)$. After raising minors to $q$, the products are $(2^q,1,1)$, which require $p\\ge q$. Add two very small generic columns and merge those columns; the untouched quartet is unchanged and all input minors can be nonzero.\n\nFor $p\\ge1$ necessity take the five real columns\n\\[\n e_1,\\quad e_2,\\quad e_1+e_2,\\quad e_1,\\quad e_2\n\\]\nand merge the last two. On the unmerged labels $1,2,3$ the coefficients are all one. The merged coefficients are $b_1=b_2=1,b_3=2$, for every input exponent $q>0$. The output quartet products are $(2,1,1)$ and require $p\\ge1$. A small generic perturbation gives positive uniform input while retaining strict failure for any fixed proposed $p<1$. Extra columns can be added if desired. The two lower bounds and the upper bound agree.\n\\end{proof}\nDegree one has only linear coefficients; identification preserves it at every parameter.\n\n\\subsection{Exact low-parameter repair and the factor two}\nFor fixed $d\\ge3,n\\ge d+3$, let $P^{\\rm str}_{d,n}(q)$ be the infimum of parameters $p$ for which every positive uniform strong $\\T_q$ input has normalized weak $\\T_p$ output after one specified merge. Define $P^{\\rm w}_{d,n}(q)$ similarly for weak input. Monotonicity makes the universal working sets upper intervals, and Theorem~\\ref{thm:repair} makes them nonempty.\n\n\\begin{corollary}[Two-sided aggregation law]\\label{cor:cost}\nFor $d\\ge3,n\\ge d+3$,\n\\[\n P^{\\rm str}_{d,n}(q)\\le P^{\\rm w}_{d,n}(q)\\le2\\max\\{q,1\\},\n\\]\n\\[\n P^{\\rm str}_{d,n}(q)\\ge\\max\\{2,2(q-1)\\},\n \\qquad P^{\\rm str}_{d,n}(q)>q.\n\\]\nIn particular,\n\\[\n P^{\\rm str}_{d,n}(q)=P^{\\rm w}_{d,n}(q)=2\\quad(0y_h$, M-convex exchange supplies a deficit block $j$ and valid exchanged counts; choose a clone in the second but not the first in that block. This verifies ordinary basis exchange for the lift. Lemma~\\ref{lem:support} proves the needed M-convexity.\n\n\\begin{theorem}[Complete one-merge preservation classification]\\label{thm:minimal}\nFix $q>0$, rank $d\\ge1$, and $n\\ge\\max(d,2)$, with a specified pair to be merged. Universal same-parameter preservation from strong input to strong full normalized output holds exactly in the regimes of Table~\\ref{tab:threshold}. Replacing ``strong output'' by ``weak output'' gives the identical classification. Every excluded regime admits a uniform positive input counterexample.\n\\end{theorem}\n\\begin{proof}\nDegree one is immediate. For $d\\ge2$ and $n\\le d+1$, the polarized output has corank at most one. A strong relation has\n\\[\n |J\\setminus I|\\le n-d+1\\le2\n\\]\nterms, hence is zero or a tautological equal pair. Thus every parameter works.\n\nSuppose $n=d+2$. The critical metric case $(\\alpha_z,k_z)=(1,0)$ in Theorem~\\ref{thm:repair} requires a common set of $d-3$ labels, four quartet labels, and the two merged labels: $d+3$ distinct source elements. It is impossible here. All five remaining cases work already at $p=\\max\\{q,1\\}$, including their zero cases. For $q\\ge1$ the parameter stays $q$. The polarized output now has corank two, so all its strong relations have at most three terms and weak equals strong.\n\nFor $d=2$ and any $n\\ge4$, Theorem~\\ref{thm:congruence} gives same-parameter strong preservation whenever $q\\ge1$.\n\nFor the converse at $q<1$, choose $M>0$ and the rank-two real columns\n\\[\n i=(1,0),\\quad j=(0,1),\\quad\n u=(1,M^{1/q}),\\quad v=(M^{1/q},-1).\n\\]\nTheir absolute-minor $q$-power coefficients are strong by Lemma~\\ref{lem:det}. Merging $u,v$ gives the normalized output product list\n\\[\n 2(1+M^{2/q})^q,\\qquad (M+1)^2,\\qquad(M+1)^2.\n\\]\nThe ratio of the first product to the other two tends to $2$ as $M\\to\\infty$. Nullity at parameter $q$ would require that ratio to be at most $2^q<2$, so sufficiently large $M$ gives strict failure. All six source minors are nonzero. Adding $d-2$ identity coloops gives rank $d$ on $d+2$ elements; a sufficiently small generic real matrix perturbation makes every maximal minor nonzero while preserving the strict failed relation. For degree two, additional generic columns cover every $n>4$ as well.\n\nFinally, if $d\\ge3,n\\ge d+3$, Theorem~\\ref{thm:explicit} supplies a strict positive uniform counterexample at every $q>0$. This exhausts the table and proves both classifications.\n\\end{proof}\n\n\\begin{example}[Smallest exact failure below one]\nAt $q=1/2$ and $M=8$, the four real columns above are\n$(1,0),(0,1),(1,64),(64,-1)$. The nontrivial input root products are $(4097,4096,1)$, so the input is strong. The merged output root products at $q=1/2$ are\n\\[\n 16388,\\qquad6561,\\qquad6561.\n\\]\nThe excess is the integer $3266>0$. Thus failure already occurs on $U_{2,4}$, the smallest possible nontrivial ground set in degree two.\n\\end{example}", "source_section": "7", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"7\"}"} {"id": "section-8", "record_type": "manuscript_section", "title": "Two peaks: exact weak versus strong compatibility", "content": "\\section{Two peaks: exact weak versus strong compatibility}\\label{sec:peaks}\nFix $d\\ge3,n\\ge d+3$ and $|C|=d-3$ disjoint from $[6]$. Set $U=C\\cup135$ and $V=C\\cup246$. These two exceptional bases meet in $d-3$ elements.\n\n\\begin{theorem}[Exact two-peak classification]\\label{thm:twopeak}\nLet $a,b\\ge1$, define root coefficients $a_U=a$, $a_V=b$, and $a_S=1$ otherwise, and put $c_S=a_S^q$. Then\n\\[\n c\\text{ is weak over }\\T_q\\quad\\Longleftrightarrow\\quad a\\le2,\\ b\\le2,\n\\]\n\\[\n c\\text{ is strong over }\\T_q\\quad\\Longleftrightarrow\\quad\n a\\le2,\\ b\\le2,\\ ab\\le3.\n\\]\nAll genuine weak inequalities are strict when $a,b<2$, and all genuine strong inequalities are strict when additionally $ab<3$.\n\\end{theorem}\n\\begin{proof}\nA weak relation has union size $d+2$, whereas $U\\cup V$ has size $d+3$. Thus it contains at most one peak, and its root products are $(1,1,1)$, $(a,1,1)$, or $(b,1,1)$. Both peak types occur; this proves the weak equivalence and its strictness statement.\n\nFor a four-term strong relation, exceptional factors cannot lie in different terms: factors in different terms intersect in either $d-1$ or $d-2$ elements, whereas $|U\\cap V|=d-3$. If both peaks occur they pair together and give $(ab,1,1,1)$. Such a relation exists, with $I=C\\cup\\{1,3\\}$ and $J=C\\cup\\{2,4,5,6\\}$. It forces $ab\\le3$. Relations with a single peak are harmless under $a,b\\le2$. A relation with at least five terms has at most two exceptional terms, retaining at least three unit terms; its largest term is at most two. Both peaks cannot occupy one product in this case, since the intersection of the two factors of a product would have to have size $d+1-m0$. It forces $P^{\\rm w}_{d,n}(q)>2$ whenever $q>\\log_2 3$. No matching upper curve is asserted here.\n\n\\begin{example}[Three integer levels at parameter two]\nOn every $U_{d,n}$ in this section, set $c_U=23$, $c_V=14$, and every other coefficient to $6$. After removing the common root scale $\\sqrt6$, the peaks are $\\sqrt{23/6}<2$ and $\\sqrt{14/6}<2$, with product $\\sqrt{322}/6<3$. The input is strict strong by Theorem~\\ref{thm:twopeak}. Its merged product list is $(580,144,144)$, which fails $\\T_2$ because\n\\[\n 580-(12+12)^2=4>0.\n\\]\n\\end{example}\n\n\\begin{theorem}[A quantitative strong-input barrier at two]\\label{thm:twentyninths}\nFor every $d\\ge3,n\\ge d+3$,\n\\[\n P^{\\rm str}_{d,n}(2)>\\frac{20}{9}.\n\\]\n\\end{theorem}\n\\begin{proof}\nUse the three exceptional bases $U=C\\cup135$, $Z_1=C\\cup126$, $Z_2=C\\cup346$. Assign\n\\[\n c_U=3999,\\quad c_{Z_1}=c_{Z_2}=1,\\quad\n c_S=5000\\ (6\\in S\\text{ otherwise}),\\quad\n c_S=1000\\ (6\\notin S\\text{ otherwise}).\n\\]\nThis is Lemma~\\ref{lem:uniform} at $q=2$, root peak $\\sqrt{3999/1000}<2$, root low $1/\\sqrt{5000}<1/2$, followed by scaling label six by five and globally by $1000$. Hence all genuine strong input inequalities are strict.\n\nThe merged coefficients are $g_{12}=g_{34}=1001$, $g_{14}=g_{23}=g_{24}=6000$, and $g_{13}=8999$. Divide their three products by $10^6$ to obtain\n\\[\n A=\\frac{26997}{500},\\qquad B=36,\\qquad D=\\frac{1002001}{1000000}.\n\\]\nExact integer-power comparisons give\n\\[\n A^9>\\left(\\frac{6019}{1000}\\right)^{20},\\quad\n 36^9<\\left(\\frac{5016}{1000}\\right)^{20},\\quad\n D^9<\\left(\\frac{1001}{1000}\\right)^{20}.\n\\]\nFor a short manual or executable check, multiply each inequality by its positive denominators. For example, the first is exactly\n$26997^9\\,1000^{20}>6019^{20}\\,500^9$; the other two are equally direct integer comparisons. They imply\n\\[\n A^{9/20}-B^{9/20}-D^{9/20}>\\frac{2}{1000}>0.\n\\]\nThus the fixed strict input fails at output parameter $20/9$. Continuity in the output parameter gives failure on a nonempty interval above $20/9$, proving the strict inequality for the infimal universal repair parameter. All selected relations are unchanged by extra ground labels.\n\\end{proof}", "source_section": "8", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"8\"}"} {"id": "section-9", "record_type": "manuscript_section", "title": "Exact classification of weighted squarefree aggregation", "content": "\\section{Exact classification of weighted squarefree aggregation}\\label{sec:classification}\nFix $d\\ge3$ and $q>0$. Let an operator substitute\n\\[\n x_i\\longmapsto w_i y_{\\phi(i)},\\qquad w_i\\ge0,\n\\]\nand then take the squarefree part. Only the active labels $E_+=\\{i:w_i>0\\}$ matter. Let $m=|\\phi(E_+)|$. Write $A_{\\phi,w}$ for this squarefree operator. A zero output is allowed; otherwise its coefficient array has rank $d$.\n\n\\begin{theorem}[Weighted set-map classification]\\label{thm:classification}\nThe following are equivalent:\n\\begin{enumerate}\n\\item For every strong $\\T_q$ rank-$d$ input, $A_{\\phi,w}$ is zero or strong over $\\T_q$.\n\\item For every strong $\\T_q$ rank-$d$ input, $A_{\\phi,w}$ is zero or weak over $\\T_q$.\n\\item Either $\\phi|_{E_+}$ is injective, or $m\\le d+1$.\n\\end{enumerate}\nWhen condition 3 fails, failure can be witnessed by a positive uniform input on the original ground set, with every genuine input inequality strict.\n\\end{theorem}\n\\begin{proof}\nCondition 1 implies 2. If the active map is injective, the operation is deletion, positive rescaling, and relabeling; strong representability is preserved, unless every rank-$d$ term is deleted. If $m\\le d+1$, any nonzero rank-$d$ array on these labels is strong: for $m=d$ it has at most one basis, and for $m=d+1$ every strong relation has only its two identical terms. Its support is a matroid, since complements of the possible bases are the nonloops of a rank-one matroid. For $m0$. Input positivity and strictness are preserved. In the limit $\\lambda\\to0$, the six output coefficients of the selected weak relation converge to \\eqref{eq:six}; contributions involving any additional element of the same output fibers vanish. Inactive variables contribute nothing. The limiting output relation fails by at least $L/2$, so it still fails for all sufficiently small positive $\\lambda$. This contradicts condition 2 and proves necessity. Every input coefficient remains positive for such $\\lambda$, as asserted.\n\\end{proof}\n\n\\begin{remark}[Why the squarefree qualifier is necessary]\nThe low-output-count exception concerns rank-$d$ arrays on at most $d+1$ labels. The full polynomial may have repeated monomials and additional weak tests. Theorem~\\ref{thm:classification} is not a classification of all full-polynomial linear preservers, nor of arbitrary matrices of linear forms.\n\\end{remark}\n\n\\begin{corollary}[A guaranteed rule for repeated squarefree merging]\nStart with positive weak data on $U_{d,n}$. Perform $k$ binary identifications, taking the squarefree shadow after each, and assume the number of labels remains at least $d$. The resulting array is weak over $\\T_p$ for\n\\[\n p=2^k\\max\\{q,1\\}.\n\\]\nFor $d=2$ the better parameter $\\max\\{q,1\\}$ suffices after any number of such steps.\n\\end{corollary}\n\\begin{proof}\nEach squarefree shadow remains positive on the new uniform support, and its weak tests are among the full output weak tests just proved. Iterate Theorem~\\ref{thm:repair}; in degree two iterate $r\\mapsto\\max\\{r,1\\}$ instead.\n\\end{proof}\nThe exponential bound is a safe composition bound, not an optimal cost for a partition or for multiple merges.", "source_section": "9", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"9\"}"} {"id": "section-10", "record_type": "manuscript_section", "title": "A robust Lorentzian obstruction to strong compatibility", "content": "\\section{A robust Lorentzian obstruction to strong compatibility}\\label{sec:lorentzian}\nThe preceding repair results distinguish weak and strong output. This section gives an independent explicit obstruction: even a Lorentzian cubic can fail strong compatibility in the same parameter that controls its local quadratic geometry.\n\nFor a positive multi-affine cubic on six labels, we use the standard Lorentzian criterion: each quadratic derivative has a Hessian with at most one positive eigenvalue. Uniform support supplies the required M-convexity. This is the usual derivative characterization in \\cite{BH}. The active Hessian omits the zero row and column of the differentiated variable. Signature $(1,4)$ of all six active Hessians implies membership in the relative interior of the Lorentzian $U_{3,6}$ stratum.\n\n\\begin{theorem}[Interior Lorentzian obstruction]\\label{thm:lorentz}\nPartition $[6]$ as $P=\\{1,3\\}$, $Q=\\{2,4,6\\}$, and $\\{s\\}=\\{5\\}$. For $t>1$ and $0<\\delta<1$, put\n\\[\n a=3-\\delta,\\qquad h=4-t^{-2}-\\delta.\n\\]\nAssign the twenty cubic coefficients by the following exhaustive composition classes:\n\\begin{center}\n\\begin{tabular}{lrrrrrr}\n\\toprule\nClass & $PPs$ & $QQQ$ & $PPQ$ & $sQQ$ & $PsQ$ & $PQQ$\\\\\nNumber & 1 & 1 & 3 & 3 & 6 & 6\\\\\nCoefficient & $a$ & $h$ & 1 & 1 & $t$ & $t$\\\\\\bottomrule\n\\end{tabular}\n\\end{center}\nEvery active derivative Hessian has signature $(1,4)$. The strong relation indexed by $I=P$ and $J=Q\\cup\\{s\\}$ has product list $(ah,1,1,1)$. Therefore any parameter $p$ containing all Lorentzian $U_{3,6}$ coefficient arrays among strong $\\T_p$ arrays must satisfy\n\\[\n p\\ge\\log_3 12.\n\\]\n\\end{theorem}\n\\begin{proof}\nAll coefficients are positive. Symmetry reduces the six active Hessians to three types; the quotient matrices below use orthonormal constant vectors on their symmetry blocks.\n\nFor differentiation in $s$, the quotient on $P,Q$ is\n\\[\n \\begin{pmatrix}a&\\sqrt6t\\\\\\sqrt6t&2\\end{pmatrix}.\n\\]\nIts determinant $2a-6t^2$ is negative, giving one positive and one negative eigenvalue. The other three modes have eigenvalues $-a,-1,-1$.\n\nFor differentiation in a member of $P$, the symmetry blocks have sizes $1,1,3$, and the quotient is\n\\[\n \\begin{pmatrix}\n 0&a&\\sqrt3\\\\ a&0&\\sqrt3t\\\\ \\sqrt3&\\sqrt3t&2t\n \\end{pmatrix}.\n\\]\nIts determinant is $2at\\delta>0$ and the sum of its principal two-by-two minors is $-a^2-3-3t^2<0$. Thus its three nonzero eigenvalues have one positive and two negative signs: their product is positive, while the all-positive possibility is excluded by the negative sum of pairwise products. The other two modes are $-t,-t$.\n\nFor differentiation in a member of $Q$, the symmetry blocks have sizes $2,2,1$, and the quotient is\n\\[\n \\begin{pmatrix}\n 1&2t&\\sqrt2t\\\\ 2t&h&\\sqrt2\\\\ \\sqrt2t&\\sqrt2&0\n \\end{pmatrix}.\n\\]\nIts determinant is $2t^2\\delta>0$ and the sum of its principal two-by-two minors is $h-6t^2-2<0$. Again it has one positive and two negative eigenvalues. The other two modes are $-1,-h$. This proves signature $(1,4)$ in every case.\n\nIn the specified strong relation, the $s$ term is $c_{PPs}c_{QQQ}=ah$; each of the other three terms is $c_{PPQ}c_{sQQ}=1$. Strong nullity requires $(ah)^{1/p}\\le3$, so $p\\ge\\log_3(ah)$. Let $\\delta\\downarrow0$ and $t\\to\\infty$ to obtain $p\\ge\\log_3 12$.\n\\end{proof}\n\n\\begin{corollary}[A robust four-level integer certificate]\\label{cor:lorentzbox}\nTake $t=2$, $\\delta=1/4$, and multiply every coefficient by four. The six coefficient classes become\n\\[\n 11,\\quad14,\\quad4,\\quad4,\\quad8,\\quad8.\n\\]\nThis Lorentzian array fails strong $\\T_2$; moreover every coefficient perturbation of absolute size at most $1/1000$ retains both properties within the multi-affine coefficient space.\n\\end{corollary}\n\\begin{proof}\nThe strong product list is $(154,16,16,16)$, and $154>144=(4+4+4)^2$ proves failure. Exact active-Hessian determinants, in variable order, are\n\\[\n 11264,\\ 7168,\\ 11264,\\ 7168,\\ 52096,\\ 7168.\n\\]\nEach maximum absolute row sum is at most $35$. All eigenvalues therefore have absolute value at most $35$, while the determinant product implies every absolute eigenvalue is at least $7168/35^4$. A perturbation of the twenty coefficients by at most $\\eps=1/1000$ changes each active Hessian by operator norm at most $4\\eps$, because each row has only four off-diagonal entries. The exact inequality\n$4/1000<7168/35^4$ prevents any eigenvalue from crossing zero, so the active signatures remain $(1,4)$.\n\nThe worst permitted strong-product comparison still has positive squared excess:\n\\[\n (11-\\eps)(14-\\eps)-9(4+\\eps)^2\n =10-97\\eps-8\\eps^2>0.\n\\]\nAll coefficients remain positive. This gives the stated explicit relative open box.\n\\end{proof}\nThis is a lower-bound theorem for a distinct inclusion problem. It does not establish finiteness or a matching upper bound for the Lorentzian-to-strong parameter.", "source_section": "10", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"10\"}"} {"id": "section-11", "record_type": "manuscript_section", "title": "Concrete certificates and verification", "content": "\\section{Concrete certificates and verification}\\label{sec:verification}\nThe package contains the complete source, an executable verifier, and its JSON results. Universal real-parameter statements are established by the proofs above. Finite computations are certificates for selected constructions and checks against transcription mistakes.\n\n\\subsection{A strict rational witness at $q=1$}\nOn $U_{3,6}$ take root coefficients $a_{135}=199/100$, $a_{126}=a_{346}=1/4$, and all others one; rescale label six by $s=1/4$. Lemma~\\ref{lem:uniform} proves strict strong input. The relevant output coefficients are\n\\[\n g_{12}=g_{34}=17/16,\\quad\n g_{14}=g_{23}=g_{24}=5/4,\\quad\n g_{13}=56/25.\n\\]\nThe failed triangle defect, with no numerical roots required, is\n\\[\n \\frac{56}{25}\\frac54-\\left(\\frac54\\right)^2-\n \\left(\\frac{17}{16}\\right)^2=\\frac{139}{1280}>0.\n\\]\nThis example is convenient for an independent manual check of strictness, merging, and exact rational failure.\n\n\\subsection{A strict integer witness at $q=2$}\nAn explicit integer construction is\n\\[\n c_{135}=4095,\\quad c_{126}=c_{346}=8,\\quad\n c_S=64\\ (6\\in S\\text{ otherwise}),\\quad\n c_S=1024\\ (6\\notin S\\text{ otherwise}).\n\\]\nIts six merged coefficients are $g_{12}=g_{34}=1032$, $g_{14}=g_{23}=g_{24}=1088$, and $g_{13}=4159$. Failure is certified by\n\\[\n 4159\\cdot1088-(1088+1032)^2=30592>0.\n\\]\nThe input is obtained from Lemma~\\ref{lem:uniform} with root peak $\\sqrt{4-1/1024}$ and low $1/\\sqrt8$, followed by label-six scaling $1/16$ and global scaling $1024$. Exact enumeration provides a second certificate.\n\n\\subsection{What is checked}\nThe delivered scripts separate exact certificates from numerical checks.\n\\begin{itemize}\n\\item The core verifier enumerates all strong relations in eight rank/size configurations, certifies the rational and integer examples, and checks all six normalized output cases as formal polynomial identities. It also certifies 567 boundary-case relations exactly and checks 264 numerical instances containing 11,424 normalized weak relations.\n\\item The congruence verifier checks three formal tensor identities, 16 exact rational matrix instances, 128 numerical substitution/composition instances at 90 digits, the minimal corank example, the two-peak classification in selected ranks and sizes, and the integer levels $6,14,23$.\n\\item The strong-parameter verifier certifies every one of 225 indexed input relations for each of two integer witnesses, then checks the output twentieth-power comparisons entirely with integers and fractions.\n\\item The Lorentzian verifier proves twelve derivative quotient identities symbolically and checks 372 principal minors exactly at the rational examples. It also verifies the strict integer example and every stated robustness constant.\n\\end{itemize}\nAll delivered checks pass. Universal real-parameter statements are established by the written proofs, rather than inferred from samples. The scripts are reproducibility certificates, not a proof-assistant formalization.", "source_section": "11", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"11\"}"} {"id": "section-12", "record_type": "manuscript_section", "title": "Interpretation, limits, and the next mathematical target", "content": "\\section{Interpretation, limits, and the next mathematical target}\nThe decisive structural distinction is between homogeneous rescaling and additive aggregation. Rescaling acts by one common multiplier in each relation. Aggregation sums coefficients, and a common-index slice becomes a pointwise aggregate of two metrics after a shared normalization. Passing to a square root restores the Ptolemy property of that aggregate; this is the source of the factor two in the repair.\n\nThe exact quadratic theorem allows arbitrary nonnegative linear maps and their compositions. The cardinality table identifies the complete same-parameter boundary for a full merge. Above that boundary, explicit uniform obstructions and a constructive weak repair coexist. The two-peak family resolves weak versus strong feasibility in a complete parameter region, while the Lorentzian family exhibits a different robust obstruction to strong compatibility.\n\nFor automated algebra, the practical consequence is a mathematically sound composition rule: a system may lower its claimed output certificate from strong $\\T_q$ to weak $\\T_{2\\max(q,1)}$ after one merge of matroid data. If the same parameter or a strong output is required, new relation checks remain necessary. The paper supplies constant-size regression witnesses and a general operator rejection criterion. It does not measure an AI speedup or establish recursive self-improvement of a model.\n\nA focused remaining question is the exact function $P^{\\rm str}_{d,n}(q)$ for $q>1$. The current bounds differ by at most two and have the exact asymptotic factor. A second open extension is a comparable repair theorem when the initial input already has repeated coordinates, rather than being multi-affine. A third is higher-degree classification for arbitrary nonnegative linear substitutions rather than weighted set maps; the quadratic case is completely covered by Theorem~\\ref{thm:congruence}. None is silently included in the proved assertions.\n\n\\appendix", "source_section": "12", "verification_method": "standalone_manuscript", "verification_status": "source_text", "artifact_paths": ["paper/Ilyoxune_Aggregation_Theorem.pdf", "paper/Ilyoxune_Aggregation_Theorem.tex"], "payload_json": "{\"content_format\": \"latex\", \"section_number\": \"12\"}"} {"id": "section-a", "record_type": "manuscript_section", "title": "Why merging matroid support is M-convex", "content": "\\section{Why merging matroid support is M-convex}\n\\begin{lemma}\\label{lem:support}\nThe image of a matroid's basis indicators under one coordinate merge is M-convex.\n\\end{lemma}\n\\begin{proof}\nLet $\\phi:E\\to H$ merge $u,v$ into $z$, leaving other labels distinct. Define the integral monotone submodular function $f(A)=r_M(\\phi^{-1}(A))$. Its basis images are exactly the integer vectors\n\\[\n x\\ge0,\\qquad x(H)=d,\\qquad x(A)\\le f(A)\\quad(A\\subseteq H).\n\\]\nNecessity is immediate. Conversely, every unmerged $x_i\\le1$. Let $S$ be the unmerged labels with $x_i=1$, and put $k=x_z\\le2$. The inequalities imply that $S$ is independent and $r_M(S\\cup\\{u,v\\})\\ge d$. Thus $S$ extends to a basis using exactly $k=d-|S|$ elements of $\\{u,v\\}$, as required.\n\nWe prove symmetric exchange for these vectors. Let $x_i>y_i$. Tight sets for $x$ are closed under unions and intersections by submodularity, as are tight sets for $y$. Let $U$ be the union of all $x$-tight sets avoiding $i$ (including the empty set), and $V$ the intersection of all $y$-tight sets containing $i$ (including $H$). Then $i\\in V\\setminus U$. An exchange $x-e_i+e_j$ preserves all upper inequalities if $j\\notin U$; $y+e_i-e_j$ does so if $j\\in V$. Integer slack makes every inequality that is not tight harmless.\n\nTightness and submodularity give\n\\[\n x(V\\setminus U)\\le f(U\\cup V)-f(U)\n \\le f(V)-f(U\\cap V)\\le y(V\\setminus U).\n\\]\nSince $x_i>y_i$, some $j\\in V\\setminus U$ has $x_j