Nhoryun-Omega-III / data /manuscript_sections.jsonl
PureOne's picture
Upload Nhoryun Omega III research dataset release 3.0.0
868c741 verified
Raw History Blame Contribute Delete
280 kB
{"format": "latex", "id": "00_frontmatter", "sha256": "ea767fb6a4182eaa4d4894ad4ffa1fd9e27d124d67ded243e3a5e2dee0284718", "source_path": "research/chapters/00_frontmatter.tex", "text": "\\chapter*{Abstract}\n\\addcontentsline{toc}{chapter}{Abstract}\n\\markboth{Abstract}{}\n\nConvexity of an objective does not determine the discrete curvature of\nits optimization trajectory. This monograph develops a certificate-based\naccount of that separation and connects it to decisions in recursive\nlearning. For constant-step gradient descent on a convex $L$-smooth\nobjective, let $h=\\eta L$, $D_k=f(x_k)-f(x_{k+1})$, and\n$\\kappa_k=D_k-D_{k+1}$. We establish the exact universal threshold\n\\[\n h_{\\mathrm{iso}}=\\frac{19-\\sqrt{17}}8\n\\]\nfor the compensation law $D_{k+1}+D_{k+2}\\leq2D_k$.\nWe also identify the distinct exact threshold\n$h_*=1.863425364363717\\ldots$ for excluding two adjacent strictly\nnegative second differences. The latter constant is defined by an\nisolated root of an explicit degree-seven polynomial. Both converses\nare realized by globally smooth, strongly convex, real-analytic\nobjectives. The positive results are finite interpolation certificates;\nthe converses use exact polynomial transcripts, a positive-slack\nquadratic repair, constructive interpolation, and quantified smoothing.\n\nFor every $m\\geq1$, an explicit two-dimensional real-analytic objective\nhas $m$ consecutive negative second differences despite global strong\nconvexity and geometric convergence. With $M=m+1$, its optimal\ncondition number is exactly $128M$, its step is $2-1/(128M)$, and\n\\[\n \\kappa_k\\leq-\\frac3{128M^2}\\qquad(0\\leq k<m).\n\\]\nIts analytic smoothing width is of order\n$1/[M\\sqrt{\\log M}]$, obtained from Gaussian chamber tails and\ncancellation of a common quadratic trace. The unsmoothed orbit has an\nexact third-order loss recurrence, and the construction extends to\nfixed positive-definite metrics and controlled block compositions.\nMore generally, every finite strict trajectory condition admits an\nanalytic strongly convex realization at the same step, when the\npositive strong-convexity parameter is not fixed in advance.\n\nUsing established smooth convex interpolation as a foundation, we give\ncomplete finite transcript and prediction descriptions for prescribed\nfirst-order methods. Compact normalized model sets yield attained\nprediction intervals and maximal sound threshold decisions. A residual\ncertificate language admits exact rational verification and is complete\nfor every strict decision margin. Deterministic trajectory tubes cover\ngradient error, variable steps, and objective drift.\n\nThe information limits are equally explicit. An arbitrarily long full\nvalue-and-gradient history can leave the entire residual-gap interval\n$[v^2/(2L),v^2/(2\\mu)]$ unresolved. Even real-analytic objectives can\nmatch every observed derivative through any prescribed finite order\nand produce different future losses. Finally, a recursive acceptance\ncontroller combines simultaneous validation, a finite oracle inequality,\nall-attempt compute accounting, and conditional convergence under fair\nproposals. These are mathematical guarantees for declared models, not\nan empirical claim of neural-training acceleration or an unconditional\nsolution to recursive self-improvement.\n\n\\bigskip\n\\noindent\\textbf{Keywords.} Optimization curves; gradient descent;\ndiscrete concavity; smooth convex interpolation; semidefinite\ncertificates; real-analytic counterexamples; adaptive validation;\nrecursive self-improvement.\n\n\\chapter*{Reading guide and evidence contract}\n\\addcontentsline{toc}{chapter}{Reading guide and evidence contract}\n\\markboth{Reading guide}{}\n\nThis is a self-contained mathematical manuscript. Every research theorem\nis accompanied by its proof. Standard foundations are identified as such\nand proved where they are used to construct or certify the main results.\nThe exposition distinguishes a universal inequality, a finite\ninterpolation witness, an explicit global objective, and a conditional\nlearning-system guarantee. They are not interchangeable forms of evidence.\n\nReaders primarily interested in optimization theory can begin with\nChapters~\\ref{ch:universal}--\\ref{ch:analytic}. Readers interested in\nautomated reasoning about learning curves can begin with\nChapter~\\ref{ch:transcripts}, after the interpolation construction in\nChapter~\\ref{ch:foundations}. The implications and limits for recursive\nlearning are developed together in Chapters~\\ref{ch:information}\nand~\\ref{ch:recursive}. The final chapter describes exactly what the\nexecutable checks establish.\n\n\\begin{center}\n\\small\n\\begin{tabularx}{\\textwidth}{@{}>{\\raggedright\\arraybackslash}p{0.25\\textwidth}>{\\raggedright\\arraybackslash}X@{}}\n\\toprule\nEvidence level & Meaning in this manuscript\\\\\n\\midrule\nWritten proof & A quantified mathematical argument under stated assumptions.\\\\\nExact arithmetic audit & Reproducible checks of algebraic identities,\nfinite rational instances, interval bounds, and implementation behavior.\\\\\nNumerical exploration & Floating-point optimization or plotting used for\nillustration; it is not treated as an exact certificate.\\\\\nIndependent internal check & A second derivation or mathematical review\nwithin this research effort, not external replication.\\\\\nApplication theorem & A guarantee for an explicitly declared model of\ntraining, evaluation, proposals, and resources.\\\\\n\\bottomrule\n\\end{tabularx}\n\\end{center}\n\nThe accompanying programs do not formally verify the whole manuscript in\na proof assistant. They do not numerically integrate the Gaussian\nconvolutions and then infer their universal properties. Analyticity,\nglobal curvature bounds, all-horizon statements, and probabilistic\ncoverage follow from the written proofs. The computations independently\ncheck finite and algebraic components for which exact executable checks\nare informative.\n\nThe classical $7/4$ boundary, smooth convex interpolation, performance\nestimation, and concentration inequalities have established sources.\nTheir use is acknowledged in the text and bibliography. The targeted\nliterature search described in the verification chapter is not a claim\nof exhaustive historical priority. Publication impact and practical\nacceleration are separate from the mathematical assertions proved here.\n"}
{"format": "latex", "id": "01_overview", "sha256": "6491c9a806d382edc1c9c2c6c19e7bd19b7d49980e1c5756fdfd9727e7a69930", "source_path": "research/chapters/01_overview.tex", "text": "\\chapter{Questions, results, and their implications}\n\\label{ch:overview}\n\n\\section{The central question}\n\nAn automated learning system frequently decides whether to continue a\ntraining run, switch an optimizer, enlarge a model, or accept a proposed\nreplacement. A curve that appears to be improving slowly can make\ncontinuation seem unpromising. That inference is justified only if the\navailable assumptions restrict future improvement.\n\nThe simplest mathematical setting already contains a substantial\ndistinction. For exact gradient descent on a convex $L$-smooth objective,\na normalized step $h=\\eta L$ in $(0,2)$ makes the objective decrease.\nIt does not always make the improvements decrease. Barzilai, Shamir and\nZamani proved that discrete convexity of the loss sequence is universally\nguaranteed exactly up to $h=7/4$, and asked whether concavity can persist\nfor several consecutive steps \\cite{barzilai2025}. The question is about\nthe geometry of an orbit, not merely the geometry of the objective.\n\nThis manuscript supplies explicit long runs and a sharper account of\nthe transition into them. There are three distinct local thresholds:\nthe loss-convexity threshold $7/4$; a two-step compensation threshold;\nand a strictly larger adjacent-concavity threshold. Their separation\nshows that a local increase in improvement, failure of a particular\naveraged law, and persistence of increasing improvement are different\nphenomena even in a globally smooth strongly convex setting.\nRelated finite consecutive examples and rational threshold brackets are\nalready available in \\texttt{optcurves} \\cite{optcurves2026}. The\ncontributions here are stated relative to that work in\nChapter~\\ref{ch:verification}, including the exact adjacent boundary\nand the quantitative arbitrary-length analytic construction.\n\n\\section{The exact local phase boundaries}\n\nWith $D_k=f_k-f_{k+1}$ and $\\kappa_k=D_k-D_{k+1}$, define\n\\begin{equation}\n h_c=\\frac74,\\qquad\n h_{\\mathrm{iso}}=\\frac{19-\\sqrt{17}}8.\n \\label{eq:overview-boundaries}\n\\end{equation}\nLet $h_*$ be the unique root in\n$(1.863425364,1.863425365)$ of\n\\begin{align}\n P(h)={}&5000h^7-33602h^6+97161h^5-175306h^4\\notag\\\\\n &+235672h^3-224304h^2+126080h-33536.\n \\label{eq:overview-polynomial}\n\\end{align}\nThe isolating interval is part of the exact definition; the decimal\napproximation is a display aid. The three values satisfy\n\\[\n h_c=1.75<h_{\\mathrm{iso}}=1.8596117967\\ldots\n <h_*=1.8634253643\\ldots<2.\n\\]\n\n\\begin{center}\n\\small\n\\begin{tabularx}{\\textwidth}{@{}p{0.25\\textwidth}X@{}}\n\\toprule\nNormalized step & Exact universal statement and matching possibility\\\\\n\\midrule\n$0<h\\leq h_c$ & Every second difference is nonnegative.\\\\\n$h_c<h\\leq h_{\\mathrm{iso}}$ & Negative second differences can occur,\nbut $D_{k+1}+D_{k+2}\\leq2D_k$ always holds.\\\\\n$h_{\\mathrm{iso}}<h\\leq h_*$ & That compensation law can fail, but\ntwo adjacent second differences cannot both be strictly negative.\\\\\n$h_*<h<2$ & Two adjacent strictly negative second differences can occur\non a real-analytic strongly convex objective while stable descent holds.\\\\\n\\bottomrule\n\\end{tabularx}\n\\end{center}\n\nThe word ``can'' is existential. It does not mean that every objective or\nevery initialization exhibits the behavior. Conversely, the positive\nstatements quantify over every admissible dimension and objective.\nThe counterexamples use three dimensions when $L$ is a certified upper\nbound. If $L$ must equal the optimal global smoothness constant, an\nunused quadratic coordinate makes the same boundaries sharp in four\ndimensions without changing the observed loss sequence.\n\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=\\textwidth]{figures/phase_boundaries.pdf}\n \\caption{The exact local boundaries separate four guarantees. The\n enlarged panel distinguishes failure of the specified compensation\n law from the possibility of two adjacent negative curvatures. Bar\n endpoints are exact theorem boundaries; plotted decimals only locate\n them. The descent guarantee uses the open interval $h<2$.}\n \\label{fig:phase-boundaries}\n\\end{figure}\n\nTheorem~\\ref{thm:compensation} and\nTheorem~\\ref{thm:sharp-compensation} prove the compensation boundary.\nChapter~\\ref{ch:adjacency} proves the adjacent boundary using a positive\nweighted curvature law\n\\[\n r_*\\kappa_k+\\kappa_{k+1}\\geq0,\n \\qquad r_*=1.8798882194\\ldots.\n\\]\nIts coefficient is explicitly algebraic. An exact matrix certificate\nestablishes the full interval; an interpolable analytic construction\nviolates both unweighted curvature signs above the threshold.\n\nThese results do not yet give the exact threshold for every prescribed\nrun length $m\\geq3$. The distinction matters: resolving the first two\ncurvature constraints is a complete local result, not a complete phase\ndiagram for all finite horizons.\n\n\\section{Arbitrarily long runs with exact conditioning}\n\nFor every integer $m\\geq1$, let $M=m+1$ and\n\\begin{equation}\n a=\\frac1{8M},\\qquad \\mu=\\frac1{128M},\\qquad h=2-\\mu.\n \\label{eq:overview-witness}\n\\end{equation}\nStart with the explicit squared-distance objective\n\\[\n H_a(x)=\\frac{1-\\mu}{2}\\dist(x,C_a)^2+\\frac\\mu2\\norm{x}^2,\n \\qquad\n C_a=\\{(u,v):u\\geq0,\\ |v|\\leq au\\}.\n\\]\nIts gradient descent alternates between two projection chambers for the\nentire certified interval. The update is a contraction combined with a\nchange of the active normal direction. Angular bounds certify the\nchambers; a direct calculation certifies every negative second difference.\n\nGaussian convolution at the explicit scale\n\\begin{equation}\n \\sigma=\\frac{a}{2\\sqrt{2\\log(64/a^2)}}\n =\\frac1{16M\\sqrt{2\\log(4096M^2)}}\n \\label{eq:overview-sigma}\n\\end{equation}\ngives a real-analytic function $F_a$. Its optimal smoothness constant is\nexactly one, its optimal strong-convexity constant is exactly $\\mu$, and\nits trajectory from $(0,1)$ satisfies\n\\begin{align}\n 0<D_0<D_1<\\cdots<D_m,\\qquad\n \\kappa_k&\\leq-\\frac3{128M^2}\\quad(0\\leq k<m),\n \\label{eq:overview-longrun}\\\\\n D_0&\\leq\\frac3{64M},\\qquad\n F_a(z_0)-F_a(z_M)\\geq\\frac3{512}.\n \\label{eq:overview-gain}\n\\end{align}\nThe full orbit converges geometrically to a unique minimizer. The\ncondition number is exactly $128M$.\n\nThe analytic scale is obtained by exploiting the local structure of the\nobjective. Within a chamber it is quadratic. Both alternating chambers\nhave the same Hessian trace, so their Gaussian value offsets agree and\ncancel from second differences. Only rare Gaussian crossings of chamber\nboundaries contribute to the local discrepancy. The proof gives exact\ntail moments and an explicit orbit perturbation bound.\n\nFor a decision-maker, \\eqref{eq:overview-gain} has a clear meaning.\nThe first improvement can approach zero along this family, while a fixed\npositive total improvement remains in a subsequent interval. A rule that\ninterprets a small first decrement as a universal certificate of little\nremaining progress is therefore invalid without additional information.\nThe result supplies a controlled stress test; it does not assert that\nevery small decrement warrants more compute.\n\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=\\textwidth]{figures/witness_m32.pdf}\n \\caption{An exact rational trajectory of the unsmoothed family for\n $m=32$. The displayed numbers are decimal renderings of exact fractions.\n The objective decreases while all 32 queried second differences are\n strictly negative. The dashed line is the proved rational upper bound,\n not a fitted trend. Analytic smoothing preserves a slightly weaker\n explicit negative margin.}\n \\label{fig:overview-witness}\n\\end{figure}\n\n\\section{What can be certified from a finite history}\n\nThe finite-transcript framework puts both positive and negative answers\ninto one mathematical language. For a prescribed first-order method,\nall unknown gradients are represented by a positive semidefinite Gram\nmatrix. Pairwise smooth strong interpolation gives affine constraints\non that matrix and the loss values. Feasible data have an explicit global\nrealization; no hidden infinite-dimensional extension condition remains.\n\nAfter a gradient scale is fixed, the feasible set is compact. Every\naffine future query has an attained interval of possible values, and\nevery point of that interval is realizable. A universal threshold\ndecision is justified precisely when the whole interval lies on the\nrequired side of the threshold. Exact algebraic certificates can be\nverified independently of the numerical program that searched for them.\n\nThe basic interpolation and semidefinite formulation are established\nfoundations \\cite{taylor2017,drori2014}. The present treatment makes\ntheir decision semantics, boundedness, realization dimension, residual\nrounding allowance, and strict-margin certificate completeness explicit.\nIt also extends the finite description to deterministic bounded oracle\nerrors and supplies cheaper trajectory-tube bounds when a full\nsemidefinite calculation is unnecessary.\n\n\\section{Why even extensive observations can remain insufficient}\n\nThere is a sharp information barrier within the strongly convex class.\nFor an arbitrarily long common value-and-gradient history ending at\ngradient magnitude $v>0$, the possible remaining optimum gap can fill\nthe whole interval\n\\[\n \\left[\\frac{v^2}{2L},\\frac{v^2}{2\\mu}\\right].\n\\]\nThis is not a failure of a particular forecasting algorithm. All the\nadmissible objectives provide identical inputs to every algorithm using\nthat history. The chapter gives exact continuation formulas and the\nresulting minimax uncertainty and stopping-regret bounds.\n\nRequiring analyticity does not make a finite history identify the future.\nA Gaussian multiplied by a polynomial with sufficiently high-order zeros\ncan perturb an analytic quadratic while preserving every requested\nderivative at every observed point. An explicit global Hessian envelope\nkeeps both perturbed functions within the declared strongly convex\nsmooth class. They follow the same prefix and have different next\nobserved losses. Analytic continuation would require information on an\naccumulation set, not finitely many jets.\n\n\\section{The benefit for recursive learning}\n\nA recursive learning system may propose changes to its model, training\nalgorithm, representation, or proposal generator. The mathematical\nbenefit of this work is a disciplined interface for making and checking\nclaims about those changes.\n\nFirst, trajectory certificates prevent invalid assumptions about progress\nshape from silently controlling compute allocation. Second, exact\nprediction regions distinguish a guaranteed continuation benefit from\nunresolved model uncertainty. Third, information lower bounds explain\nwhy some decisions require additional evaluation rather than a more\nconfident extrapolation. Fourth, a common-task validation gate can accept\nchanges while controlling false improvement claims over all future\nadaptive rounds.\n\nUnder the evaluation contract in Chapter~\\ref{ch:recursive}, accepted\nlower confidence bounds telescope into certified cumulative risk\nreduction. The final incumbent is compared with every examined candidate,\nincluding rejected ones. A resource ledger charges training, validation,\nrejected experiments, and canceled work. If each attempt is reserved\nbefore execution, the ledger proves\n\\[\n \\lambda\\sum_{t=1}^T c_t\n \\leq B_0+R(s_0)-R(s_T).\n\\]\nHere $\\lambda$ is a declared conversion from compute cost to objective\nunits, not a measured speedup. Under fair proposal coverage and vanishing\nevaluation uncertainty, the incumbent risk converges to the best value\nin the declared reachable candidate set.\n\nThese theorems apply even when the accepted state includes the mechanism\nthat makes the next proposal. Their force is conditional: the evaluated\nsystems must remain fixed during each comparison, the evaluation contract\nmust be valid, and favorable proposals must actually be generated.\nNeither convex optimization geometry nor a validation gate proves that\nan arbitrary system will discover indefinitely better algorithms.\n\n\\section{Organization and scientific scope}\n\nThe manuscript is organized so that each central claim has a short path\nto its proof. Chapter~\\ref{ch:foundations} gives the interpolation\nconstruction and basic geometry. Chapters~\\ref{ch:universal}\nand~\\ref{ch:adjacency} establish the exact local boundaries.\nChapters~\\ref{ch:witness} and~\\ref{ch:analytic} give long runs with\ncomplete global guarantees. Chapters~\\ref{ch:transcripts}\nand~\\ref{ch:robustness} provide the certificate machinery.\nChapters~\\ref{ch:information} and~\\ref{ch:recursive} connect its\npossibilities and limitations to learning decisions. The verification\nchapter and appendices specify the executable evidence and reproduction\nprocedure.\n\nThe original consecutive-concavity existence question is answered.\nThe exact compensation and adjacent-concavity boundaries are established.\nSharp boundaries for longer runs, the best possible conditioning for a\ngiven length, and practical acceleration of neural learning remain\nseparate questions. The manuscript states these boundaries of scope\nbecause a theorem becomes more useful when its quantifiers are clear.\n"}
{"format": "latex", "id": "02_foundations", "sha256": "c4e87925ab934b47ec96ebae9286d05cb1dcc9c8557c42ee82f32128790ebe45", "source_path": "research/chapters/02_foundations.tex", "text": "\\chapter{Geometry, interpolation, and exact finite models}\n\\label{ch:foundations}\n\nThis chapter supplies the analytic facts used throughout the monograph.\nThe interpolation theorem is an established foundation of performance\nestimation \\cite{taylor2017,drori2014}; its proof is included because it turns\na finite algebraic witness into a globally defined objective. The distinction\nbetween a feasible list of numbers and a genuine optimization trajectory is\nessential in both the sharpness argument and the forecasting results.\n\n\\section{Function classes and conventions}\n\n\\begin{definition}[Smooth strongly convex class]\nFor $0\\leq\\mu<L<\\infty$, let $\\Fclass_{\\mu,L}(\\R^d)$ consist of differentiable\nfunctions $f:\\R^d\\to\\R$ satisfying, for all $x,y$,\n\\begin{align}\n f(y)&\\geq f(x)+\\ip{\\nabla f(x)}{y-x}+\\frac\\mu2\\norm{y-x}^2,\n \\label{eq:strongdef}\\\\\n \\norm{\\nabla f(y)-\\nabla f(x)}&\\leq L\\norm{y-x}.\n \\label{eq:smoothdef}\n\\end{align}\nThe constants are certified bounds; unless expressly stated, they need not\nbe the best possible constants of the particular function.\n\\end{definition}\n\nAll vector spaces are finite dimensional and use the Euclidean inner\nproduct unless a different metric is specified. A positive semidefinite\nmatrix is denoted by $G\\succeq0$. Function values may be translated by a\ncommon constant. Points may be translated or transformed by an orthogonal\nmap. These operations leave gradient descent and every loss difference\nunchanged after the corresponding coordinate transformation.\n\nFor a fixed step $\\eta>0$, write\n\\[\n x_{k+1}=x_k-\\eta g_k,\\qquad\n g_k=\\nabla f(x_k),\\qquad f_k=f(x_k),\\qquad h=\\eta L.\n\\]\nThe decrement and discrete curvature are\n\\[\n D_k=f_k-f_{k+1},\\qquad\n \\kappa_k=f_k-2f_{k+1}+f_{k+2}=D_k-D_{k+1}.\n\\]\nThus a negative curvature means increasing improvement. It does not mean\nthat the objective itself is nonconvex, that the objective is increasing,\nor that its gradient norm is increasing. These are different properties.\n\n\\section{Descent and interpolation inequalities}\n\n\\begin{lemma}[Descent inequality]\nIf $f$ has an $L$-Lipschitz gradient, then\n\\begin{equation}\n \\abs{f(y)-f(x)-\\ip{\\nabla f(x)}{y-x}}\n \\leq\\frac L2\\norm{y-x}^2.\n \\label{eq:descent-lemma}\n\\end{equation}\nConsequently, for convex $f$ and one gradient step,\n\\begin{equation}\n \\eta(1-h/2)\\norm{g_k}^2\\leq D_k\\leq\\eta\\norm{g_k}^2.\n \\label{eq:drop-sandwich}\n\\end{equation}\n\\end{lemma}\n\n\\begin{proof}\nLet $v=y-x$. The fundamental theorem of calculus gives\n\\[\n f(y)-f(x)-\\ip{\\nabla f(x)}v\n =\\int_0^1\\ip{\\nabla f(x+tv)-\\nabla f(x)}v\\,dt.\n\\]\nThe absolute value of the integrand is at most $Lt\\norm v^2$.\nIntegration proves \\eqref{eq:descent-lemma}. Substitute $v=-\\eta g_k$\nfor the lower bound on $D_k$. Convexity gives\n$f_{k+1}\\geq f_k-\\eta\\norm{g_k}^2$, which is the upper bound.\n\\end{proof}\n\n\\begin{lemma}[Smooth convex interpolation inequality]\n\\label{lem:interpolation-necessary}\nFor $f\\in\\Fclass_{0,L}$ and any $x_i,x_j$,\n\\begin{equation}\n I_{ij}:=f_i-f_j-\\ip{g_j}{x_i-x_j}\n -\\frac1{2L}\\norm{g_i-g_j}^2\\geq0.\n \\label{eq:interp-I}\n\\end{equation}\n\\end{lemma}\n\n\\begin{proof}\nSet $\\phi(x)=f(x)-\\ip{g_j}x$. The function $\\phi$ is convex and\n$L$-smooth, and $x_j$ is a global minimizer because\n$\\nabla\\phi(x_j)=0$. Applying the descent inequality at $x_i$ with\nstep $1/L$ yields\n\\[\n \\phi\\bigl(x_i-L^{-1}\\nabla\\phi(x_i)\\bigr)\n \\leq\\phi(x_i)-\\frac1{2L}\\norm{\\nabla\\phi(x_i)}^2.\n\\]\nThe left side is at least $\\phi(x_j)$. Substitution of the definition of\n$\\phi$ gives \\eqref{eq:interp-I}.\n\\end{proof}\n\nAdding \\eqref{eq:interp-I} and its version with $i,j$ reversed gives\n\\begin{equation}\n \\ip{g_i-g_j}{x_i-x_j}\\geq\\frac1L\\norm{g_i-g_j}^2.\n \\label{eq:cocoercive}\n\\end{equation}\nThis inequality controls gradient differences through displacement. Its\nproof here requires neither a Hessian nor an attained minimum of $f$.\n\n\\begin{proposition}[Nonexpansiveness and gradient decrease]\n\\label{prop:nonexpansive}\nFor $0\\leq\\eta L\\leq2$, the gradient map $T(x)=x-\\eta\\nabla f(x)$ of\na convex $L$-smooth function is nonexpansive. Along its trajectory,\n\\begin{equation}\n \\norm{g_{k+1}}^2\n \\leq\\norm{g_k}^2-\\left(\\frac2h-1\\right)\\norm{g_{k+1}-g_k}^2\n \\qquad(h>0).\n \\label{eq:gradient-decrease}\n\\end{equation}\n\\end{proposition}\n\n\\begin{proof}\nFor $v=x-y$ and $w=\\nabla f(x)-\\nabla f(y)$,\n\\[\n \\norm{T(x)-T(y)}^2\n =\\norm v^2-2\\eta\\ip vw+\\eta^2\\norm w^2\n \\leq\\norm v^2-\\eta(2/L-\\eta)\\norm w^2.\n\\]\nFor the second assertion put $d=g_{k+1}-g_k$. Equation\n\\eqref{eq:cocoercive} and $x_{k+1}-x_k=-\\eta g_k$ imply\n$\\ip d{g_k}\\leq-\\norm d^2/h$. Expanding\n$\\norm{g_k+d}^2$ proves \\eqref{eq:gradient-decrease}.\n\\end{proof}\n\nThe gradient norm can therefore decrease throughout a long interval in\nwhich the decrements increase. The construction in Chapter~\\ref{ch:witness}\nexploits the distinction between these two observables.\n\n\\section{Removing the strong convexity quadratic}\n\n\\begin{lemma}[Class transformation]\n\\label{lem:class-transform}\nFor $0\\leq\\mu<L$, the relation\n\\[\n F(x)=f(x)-\\frac\\mu2\\norm x^2\n\\]\nmaps $\\Fclass_{\\mu,L}$ exactly onto $\\Fclass_{0,L-\\mu}$.\n\\end{lemma}\n\n\\begin{proof}\nConvexity of $F$ follows immediately from \\eqref{eq:strongdef}.\nTo justify its sharper smoothness bound without assuming a classical\nHessian, first smooth $f$ by a compactly supported nonnegative\n$C^\\infty$ kernel of total mass one. Let $f_\\epsilon$ denote this\nconvolution. Integrating \\eqref{eq:strongdef} and\n\\eqref{eq:smoothdef} shows that $f_\\epsilon$ remains\n$\\mu$-strongly convex and $L$-smooth. Its Hessian consequently satisfies\n\\[\n \\mu I\\preceq\\nabla^2 f_\\epsilon(x)\\preceq LI.\n\\]\nThe left inequality follows by twice differentiating the strong convexity\ninequality along any line. The right follows by differentiating the\nLipschitz gradient in a unit direction and using symmetry and convexity\nof the Hessian. Hence\n$0\\preceq\\nabla^2(f_\\epsilon-\\mu\\norm x^2/2)\\preceq(L-\\mu)I$.\nIntegration along line segments gives the $(L-\\mu)$-Lipschitz bound.\nSince $\\nabla f_\\epsilon\\to\\nabla f$ uniformly on compact sets, this\nbound passes to the limit. Conversely, adding $\\mu\\norm x^2/2$ to a\nconvex $(L-\\mu)$-smooth function gives \\eqref{eq:strongdef}, and the\ntriangle inequality gives the $L$-Lipschitz gradient bound.\n\\end{proof}\n\n\\begin{proposition}[Contraction and residual bounds]\n\\label{prop:contraction}\nLet $f\\in\\Fclass_{\\mu,L}$ with $\\mu>0$, and let\n\\[\n q=\\max\\{\\abs{1-\\eta\\mu},\\abs{1-\\eta L}\\}.\n\\]\nThen $T=I-\\eta\\nabla f$ is $q$-Lipschitz. If $0<\\eta<2/L$, it\nhas the unique fixed point $x_*=\\argmin f$, and\n\\begin{align}\n \\norm{x_k-x_*}&\\leq q^k\\norm{x_0-x_*},\\label{eq:distance-contract}\\\\\n \\norm{g_k}&\\leq q^k\\norm{g_0},\\label{eq:gradient-contract}\\\\\n \\frac{\\norm{\\nabla f(x)}^2}{2L}\n &\\leq f(x)-f(x_*)\\leq\n \\frac{\\norm{\\nabla f(x)}^2}{2\\mu}.\n \\label{eq:gap-sandwich}\n\\end{align}\n\\end{proposition}\n\n\\begin{proof}\nFor a smooth approximation as above, the mean Hessian along the segment\nbetween $x,y$ is a symmetric matrix $A$ with spectrum in $[\\mu,L]$,\nand $T(x)-T(y)=(I-\\eta A)(x-y)$. Its norm is at most\n$q\\norm{x-y}$. Passing to the limit proves the first assertion for the\noriginal function. Strong convexity implies coercivity and a unique\nminimizer. Its gradient vanishes, so iteration gives\n\\eqref{eq:distance-contract}. Along a gradient step the same mean-Hessian\nargument gives $g_{k+1}=(I-\\eta A_k)g_k$, proving\n\\eqref{eq:gradient-contract}. The lower bound in\n\\eqref{eq:gap-sandwich} follows from a step of length $1/L$ and the\nminimality of $x_*$. Minimizing the quadratic lower support\n\\eqref{eq:strongdef} over its second argument gives the upper bound.\n\\end{proof}\n\n\\section{Constructive sufficiency of finite interpolation}\n\nA list $(x_i,g_i,f_i)$ is a finite first-order transcript. The next theorem\nstates that the pairwise inequalities already contain all restrictions\nimposed by global smooth convex interpolation. It is not sufficient to\ncheck only neighboring iterates: nonneighboring inequalities can rule out\nan otherwise plausible trajectory.\n\n\\begin{theorem}[Exact convex interpolation]\n\\label{thm:convex-interpolation}\nFor a finite, nonempty index set $\\mathcal I$, data\n$\\{(x_i,g_i,f_i):i\\in\\mathcal I\\}\\subset\\R^d\\times\\R^d\\times\\R$\nare realized by a function $f\\in\\Fclass_{0,L}(\\R^d)$ if and only if\n\\eqref{eq:interp-I} holds for every ordered pair $i,j$.\nOne realizing function is the convex conjugate of\n\\begin{equation}\n \\Psi(g)=\\max_{i\\in\\mathcal I}\n \\left\\{\\ip{g_i}{x_i}-f_i+\\ip{x_i}{g-g_i}\n +\\frac1{2L}\\norm{g-g_i}^2\\right\\}.\n \\label{eq:dual-interpolant}\n\\end{equation}\nIn particular,\n\\begin{equation}\n f(x)=\\sup_{g\\in\\R^d}\\{\\ip xg-\\Psi(g)\\}\n \\label{eq:primal-interpolant}\n\\end{equation}\nis an explicit, finite-description, globally defined interpolant.\n\\end{theorem}\n\n\\begin{proof}\nNecessity is Lemma~\\ref{lem:interpolation-necessary}. For sufficiency,\nlet $\\phi_i=\\ip{g_i}{x_i}-f_i$. At $g_j$, the $i$th expression in\n\\eqref{eq:dual-interpolant} is at most $\\phi_j$, since\n\\[\n \\phi_j-\\phi_i-\\ip{x_i}{g_j-g_i}\n =f_i-f_j-\\ip{g_j}{x_i-x_j}\n \\geq\\frac1{2L}\\norm{g_i-g_j}^2.\n\\]\nThe $j$th expression equals $\\phi_j$. Thus\n$\\Psi(g_j)=\\phi_j$, and the gradient of the active $j$th quadratic is\n$x_j$. It follows directly from the supporting inequality of that\nquadratic that $x_j\\in\\partial\\Psi(g_j)$.\n\nSubtracting $\\norm g^2/(2L)$ from $\\Psi$ leaves the maximum of finitely\nmany affine functions. Therefore $\\Psi$ is $1/L$-strongly convex and\nhas quadratic growth at infinity. For each $x$, the supremum in\n\\eqref{eq:primal-interpolant} has a unique maximizer $g(x)$, characterized\nby $x\\in\\partial\\Psi(g(x))$. Strong convexity, applied in both\ndirections to $g(x),g(y)$, gives\n\\[\n \\ip{x-y}{g(x)-g(y)}\\geq L^{-1}\\norm{g(x)-g(y)}^2.\n\\]\nHence $\\norm{g(x)-g(y)}\\leq L\\norm{x-y}$.\n\nFor completeness, the elementary maximizing inequalities imply\n\\[\n 0\\leq f(y)-f(x)-\\ip{g(x)}{y-x}\n \\leq\\ip{g(y)-g(x)}{y-x}\\leq L\\norm{y-x}^2.\n\\]\nThus $f$ is differentiable with gradient $g(x)$, convex, and\n$L$-smooth. At $x_j$, the maximizing point is $g_j$, because\n$x_j\\in\\partial\\Psi(g_j)$. Consequently\n$f(x_j)=\\ip{x_j}{g_j}-\\phi_j=f_j$ and\n$\\nabla f(x_j)=g_j$.\n\\end{proof}\n\n\\begin{remark}[Attainment and evaluation]\nThe constructed function has a minimizer: any point in the nonempty\nsubdifferential $\\partial\\Psi(0)$ has gradient zero under\n\\eqref{eq:primal-interpolant}. Its value and gradient at any new $x$ can\nbe evaluated by a strongly convex optimization problem in $g$. An explicit\nfunction therefore need not be a single elementary closed-form expression.\nThe finite maximum and conjugation in \\eqref{eq:dual-interpolant}--\n\\eqref{eq:primal-interpolant} specify it without an existence oracle.\n\\end{remark}\n\n\\begin{theorem}[Exact strong interpolation]\n\\label{thm:strong-interpolation}\nFor $0\\leq\\mu<L$, finite data are realizable by a function in\n$\\Fclass_{\\mu,L}$ if and only if, for all $i,j$,\n\\begin{equation}\n f_i-f_j-\\ip{g_j}{x_i-x_j}\n \\geq\\frac\\mu2\\norm{x_i-x_j}^2\n +\\frac{\\norm{g_i-g_j-\\mu(x_i-x_j)}^2}{2(L-\\mu)}.\n \\label{eq:strong-interpolation}\n\\end{equation}\n\\end{theorem}\n\n\\begin{proof}\nTransform the data to\n\\[\n \\widetilde f_i=f_i-\\frac\\mu2\\norm{x_i}^2,\n \\qquad\\widetilde g_i=g_i-\\mu x_i.\n\\]\nLemma~\\ref{lem:class-transform} and Theorem~\\ref{thm:convex-interpolation}\nshow that realizability is equivalent to\n\\[\n \\widetilde f_i-\\widetilde f_j-\n \\ip{\\widetilde g_j}{x_i-x_j}\n \\geq\\frac1{2(L-\\mu)}\\norm{\\widetilde g_i-\\widetilde g_j}^2.\n\\]\nExpanding the left side gives the left side of\n\\eqref{eq:strong-interpolation} minus\n$\\mu\\norm{x_i-x_j}^2/2$. This proves necessity and sufficiency. Applying\nthe explicit conjugate construction to the transformed data and then\nadding $\\mu\\norm x^2/2$ provides the required function.\n\\end{proof}\n\nIf $\\mu=L$, every member of the class is a quadratic with Hessian $LI$.\nFinite interpolation then reduces to\n$g_i-g_j=L(x_i-x_j)$ and\n$f_i-f_j=\\ip{g_j}{x_i-x_j}+L\\norm{x_i-x_j}^2/2$.\nThe denominator $L-\\mu$ in \\eqref{eq:strong-interpolation} must never\nbe used at this endpoint.\n\n\\section{Why quadratic models have much more rigid curves}\n\n\\begin{theorem}[Complete monotonicity of quadratic loss]\n\\label{thm:quadratic-moments}\nLet $f(x)=f_*+\\tfrac12(x-x_*)^TA(x-x_*)$, where\n$0\\prec\\mu I\\preceq A\\preceq LI$, and let $0<\\eta<2/L$.\nThen there are nonnegative weights $w_j$ and numbers\n$t_j=(1-\\eta\\lambda_j)^2\\in[0,1)$ such that\n\\begin{equation}\n y_k:=f(x_k)-f_* =\\sum_j w_jt_j^k.\n \\label{eq:quadratic-mixture}\n\\end{equation}\nFor every integer $r,k\\geq0$,\n\\begin{equation}\n (-1)^r\\Delta^r y_k=\\sum_jw_jt_j^k(1-t_j)^r\\geq0,\n \\label{eq:complete-monotone}\n\\end{equation}\nwhere $\\Delta y_k=y_{k+1}-y_k$ and $0^0=1$ in the moment expression.\nEvery Hankel matrix $(y_{i+j})_{i,j=0}^s$ using available indices is\npositive semidefinite.\n\\end{theorem}\n\n\\begin{proof}\nDiagonalize $A$ orthogonally and let $a_j$ be the initial coordinates\nof $x_0-x_*$. The $j$th coordinate after $k$ steps is\n$(1-\\eta\\lambda_j)^ka_j$. Substituting into the quadratic gives\n\\eqref{eq:quadratic-mixture} with $w_j=\\lambda_ja_j^2/2$.\nFinite differences of $t^k$ give \\eqref{eq:complete-monotone}.\nFor an arbitrary vector $b=(b_0,\\ldots,b_s)$,\n\\[\n \\sum_{i,j}b_ib_jy_{i+j}\n =\\sum_\\ell w_\\ell\\left(\\sum_i b_i t_\\ell^i\\right)^2\\geq0.\n\\]\nThis proves the Hankel assertion.\n\\end{proof}\n\nThe second-order condition $\\kappa_k\\geq0$ is only one consequence of\nthis much stronger structure. A curve can decrease and be convex without\nadmitting any positive mixture of exponential decays. For instance,\n$y_0=1,y_1=3/5,y_2=1/5$ has nonnegative first decrements and zero second\ndifference, but\n\\[\n \\det\\begin{pmatrix}1&3/5\\\\3/5&1/5\\end{pmatrix}=-4/25<0.\n\\]\nIt cannot be the residual-loss prefix of a positive definite quadratic.\nFitting a sum of exponentials to arbitrary smooth convex training is\ntherefore an additional modeling assumption.\n\n\\section{A complete finite quadratic representation theorem}\n\nLet\n\\[\n a=\\min_{\\lambda\\in[\\mu,L]}(1-\\eta\\lambda)^2,\\qquad\n b=\\max_{\\lambda\\in[\\mu,L]}(1-\\eta\\lambda)^2,\n\\]\nwhere $0<\\mu\\leq L$ and $0<\\eta<2/L$. Thus $0\\leq a\\leq b<1$.\nFor $N\\geq0$, write $v_N(t)=(1,t,\\ldots,t^N)$.\n\n\\begin{theorem}[Finite moment characterization]\n\\label{thm:finite-quadratic}\nA vector $y=(y_0,\\ldots,y_N)$ is the residual-loss transcript of some\nquadratic with spectrum in $[\\mu,L]$ under step $\\eta$ if and only if\n\\begin{equation}\n y\\in\\operatorname{cone}\\{v_N(t):t\\in[a,b]\\}.\n \\label{eq:moment-cone}\n\\end{equation}\nWhenever it is realizable, at most $N+1$ quadratic coordinates suffice.\nEquivalently, for every real polynomial\n$p(t)=\\sum_{k=0}^N c_kt^k$ that is nonnegative on $[a,b]$,\n\\begin{equation}\n \\sum_{k=0}^N c_ky_k\\geq0.\n \\label{eq:moment-polar}\n\\end{equation}\n\\end{theorem}\n\n\\begin{proof}\nThe forward implication in \\eqref{eq:moment-cone} follows from\n\\eqref{eq:quadratic-mixture}. A finite positive combination with more\nthan $N+1$ terms can be shortened. The vectors $v_N(t_j)$ all have first\ncoordinate one and lie in an affine space of dimension at most $N$.\nWith more than $N+1$ of them there are real numbers $\\alpha_j$, not all\nzero, such that $\\sum_j\\alpha_jv_N(t_j)=0$ and\n$\\sum_j\\alpha_j=0$. Both signs occur among nonzero $\\alpha_j$.\nFor positive weights $w_j$, choose\n$\\theta=\\min_{\\alpha_j>0}w_j/\\alpha_j$ and replace\n$w_j$ by $w_j-\\theta\\alpha_j$. The combination is unchanged, every\nweight remains nonnegative, and one term disappears. Repetition proves\nthe bound of $N+1$ terms.\n\nConversely, choose for each atom $t_j$ a\n$\\lambda_j\\in[\\mu,L]$ with $(1-\\eta\\lambda_j)^2=t_j$.\nSuch a value exists because the continuous image of $[\\mu,L]$ is exactly\n$[a,b]$. Set $A=\\diag(\\lambda_j)$ and\n$x_{0,j}=\\sqrt{2w_j/\\lambda_j}$. Equation\n\\eqref{eq:quadratic-mixture} realizes the prescribed transcript.\n\nIt remains to prove the polar characterization. The cone in\n\\eqref{eq:moment-cone} is closed. Indeed, for a convergent sequence of\nmoment vectors, the total mass equals its first coordinate and is\nbounded. After the preceding reduction, each vector uses at most $N+1$\natoms. Compactness of $[a,b]$ and of the bounded simplex of weights gives\na convergent subsequence of these representations, which represents the\nlimit. The cone is convex by addition of atomic representations.\nIf $y$ lies outside it, separation from a closed convex cone supplies a\nlinear functional $c$ nonnegative on the cone and negative at $y$.\nNonnegativity on its generators means precisely that\n$p(t)=c\\cdot v_N(t)\\geq0$ on $[a,b]$, contradicting\n\\eqref{eq:moment-polar}. The other implication is immediate by\nintegrating or summing a nonnegative polynomial against the atoms.\n\\end{proof}\n\nThis theorem is a finite version of a classical moment principle, proved\nhere directly for the optimization model. It gives an exact alternative\nto a curve-fit score: a proposed quadratic explanation is either feasible\nor separated by a polynomial inequality. The analogous general smooth\nconvex description uses a Gram matrix rather than scalar spectral atoms;\nthat description is developed in Chapter~\\ref{ch:transcripts}.\n\n\\section{Continuous time does not restore quadratic rigidity}\n\nGradient flow satisfies $\\dot x=-\\nabla f(x)$. For a $C^2$ convex\nfunction, differentiation gives\n\\begin{equation}\n \\frac{d}{dt}f(x(t))=-\\norm{\\nabla f(x(t))}^2,\\qquad\n \\frac{d^2}{dt^2}f(x(t))=\n 2\\nabla f(x(t))^T\\nabla^2f(x(t))\\nabla f(x(t))\\geq0.\n \\label{eq:flow-convex}\n\\end{equation}\nThe general smooth convex result is established in\n\\cite{barzilai2025}. For the analytic class considered here,\n\\eqref{eq:flow-convex} is already a direct proof. Higher alternating\nderivative signs, however, need not hold.\n\n\\begin{example}[An analytic flow curve that is not completely monotone]\nTake\n\\[\n f(x)=x^2-\\sin x-10x.\n\\]\nThen $f''(x)=2+\\sin x\\in[1,3]$, so the function is globally\n$1$-strongly convex and $3$-smooth. It is real analytic. At $x(0)=0$,\nwrite $g=f'(0)=-11$, $H=f''(0)=2$, and $J=f'''(0)=1$.\nFor $F(t)=f(x(t))-\\min f$, direct differentiation in one dimension gives\n\\[\n F'''(0)=-4g^2H^2-2g^3J=-1936+2662=726>0.\n\\]\nA positive mixture of decaying exponentials has nonpositive third\nderivative. Thus even this strongly convex analytic gradient-flow curve\ndoes not belong to that model class. There is no contradiction with\n\\eqref{eq:flow-convex}: its second derivative is positive.\n\\end{example}\n\n\\section{Finite runs and infinite tails}\n\n\\begin{proposition}[No positive increasing tail on a lower-bounded loss]\n\\label{prop:no-infinite-tail}\nSuppose $f_k$ is bounded below and $D_k=f_k-f_{k+1}$.\nThere cannot be an index $K$ for which $D_K>0$ and\n$D_{k+1}\\geq D_k$ for every $k\\geq K$.\n\\end{proposition}\n\n\\begin{proof}\nSuch a tail would satisfy $D_k\\geq D_K>0$, giving\n$f_K-f_{K+n}=\\sum_{j=K}^{K+n-1}D_j\\geq nD_K$.\nThe right side is unbounded, contradicting the lower bound on $f_k$.\n\\end{proof}\n\nThe quantifiers in an arbitrarily long finite-run theorem are therefore\nessential. For every prescribed length there is an objective and a step\nproducing such a run. The objective and step may change with the length.\nThis statement does not produce a single infinite increasing tail and\ndoes not claim that a finite loss budget can finance unlimited positive\nimprovement.\n\n\\section{How the foundations are used}\n\nThere are three different mathematical tasks in the following chapters.\nAn inequality certificate proves that every admissible transcript has a\nproperty. An interpolation witness proves that a particular transcript\ncan actually occur. An explicit analytic construction proves more than\nfinite interpolation alone, by controlling global regularity, condition\nnumber, and the complete orbit. Each task needs its own argument.\n\nThe conjugate construction is particularly useful at a sharp threshold.\nIt allows a finite collection of exact inequalities to replace the\nguessing of a global formula. Conversely, the wedge construction supplies\na simple global formula and uses geometry to certify the finite branch.\nNeither method makes the other redundant. Their combination gives both\nuniversal laws and realizable counterexamples, which is the basis for\nmeaningful model-aware decisions about future learning.\n"}
{"format": "latex", "id": "03_universal", "sha256": "93c5f4ce855699356c2e11717196ab92376da4a8fe5bdf5a7dff9022d79becca", "source_path": "research/chapters/03_universal.tex", "text": "\\chapter{Universal curvature laws}\n\\label{chap:universal}\n\\label{ch:universal}\n\nA loss curve contains more structure than monotonicity, but that structure depends on the step size. This chapter identifies an exact boundary for one such structure. The boundary concerns a two-step average of improvements, and is different from the boundary for convexity of the unfiltered curve. We prove both the positive theorem and its converse. The converse remains valid for real-analytic, strongly convex objectives in three dimensions. We then derive consequences at every horizon, identify the precise sense in which a resulting scalar envelope is sharp, and give error terms for inexact updates.\n\nThroughout the chapter, $f$ is convex and globally $L$-smooth, $L>0$, and the exact iteration is\n\\begin{equation}\n x_{k+1}=x_k-\\eta g_k,\\qquad g_k=\\nabla f(x_k),\\qquad h=\\eta L>0.\n \\label{eq:univ-gd}\n\\end{equation}\nWe write $f_k=f(x_k)$, $D_k=f_k-f_{k+1}$ and\n$\\kappa_k=D_k-D_{k+1}$. All universal statements quantify over the dimension, the objective and the starting point. We first use $L=1$; the displayed general-$L$ forms are obtained by applying the normalized statement to $f/L$.\n\n\\section{The basic inequalities and the classical boundary}\n\nThe smooth convex interpolation inequality is\n\\begin{equation}\n I_{ij}=f_i-f_j-\\ip{g_j}{x_i-x_j}\n -\\frac12\\norm{g_i-g_j}^2\\geq0.\n \\label{eq:univ-I}\n\\end{equation}\nFor completeness, put $\\phi(x)=f(x)-\\ip{g_j}{x}$. Its minimum is attained at $x_j$. The descent inequality at $x_i$, applied to the step $-\\nabla\\phi(x_i)$, gives\n\\[\n \\phi(x_j)\\leq\\phi(x_i-\\nabla\\phi(x_i))\n \\leq\\phi(x_i)-\\tfrac12\\norm{\\nabla\\phi(x_i)}^2.\n\\]\nRearrangement proves \\eqref{eq:univ-I}. Adding the two orientations yields cocoercivity,\n\\begin{equation}\n \\ip{g_i-g_j}{x_i-x_j}\\geq\\norm{g_i-g_j}^2.\n \\label{eq:univ-coco}\n\\end{equation}\nIn particular, with $d_k=g_{k+1}-g_k$, exact gradient descent gives\n\\begin{equation}\n \\norm{g_k}^2-\\norm{g_{k+1}}^2\n \\geq\\left(\\frac2h-1\\right)\\norm{d_k}^2,\n \\qquad 0<h\\leq2.\n \\label{eq:univ-gradient-dissipation}\n\\end{equation}\nIndeed, \\eqref{eq:univ-coco} gives $-h\\ip{d_k}{g_k}\\geq\\norm{d_k}^2$, which is substituted into\n$\\norm{g_{k+1}}^2=\\norm{g_k}^2+2\\ip{g_k}{d_k}+\\norm{d_k}^2$.\nThe gradient norm therefore cannot increase in this interval, even when the loss decrement does increase.\n\nThe standard smoothness estimate also gives\n\\begin{equation}\n D_k\\geq h(1-h/2)\\norm{g_k}^2.\n \\label{eq:univ-descent}\n\\end{equation}\nThus $D_k\\geq0$ for $h\\leq2$, and $D_k>0$ for $h<2$ whenever $g_k\\ne0$. Neither statement by itself implies $D_{k+1}\\leq D_k$.\n\n\\begin{theorem}[Classical sharp convexity boundary]\n\\label{thm:classical-seven-fourths}\nFor $0<h\\leq7/4$, every loss sequence in \\eqref{eq:univ-gd} is discretely convex. For every $h>7/4$ there is a globally $1$-smooth convex objective whose loss sequence is not discretely convex.\n\\end{theorem}\n\n\\begin{proof}\nThe positive statement is the theorem of Barzilai, Shamir and Zamani \\cite{barzilai2025}; we reproduce its certificate because it explains the more general argument below. At the first three iterates,\n\\begin{equation}\n \\frac32I_{01}+\\frac12I_{20}+\\frac12I_{21}\n =\\kappa_0-\\left[\n \\frac{7-4h}{8}\\norm{d_0}^2\n +\\frac12\\norm{d_1+\\tfrac12d_0}^2\\right].\n \\label{eq:univ-classical-certificate}\n\\end{equation}\nThe identity follows by substituting $x_1=x_0-hg_0$ and $x_2=x_0-h(g_0+g_1)$. Every interpolation term is nonnegative. The two square coefficients on the right are nonnegative for $h\\leq7/4$. Translation of the starting index proves the assertion.\n\nHere is the elementary one-dimensional sharpness example used in \\cite{barzilai2025} for $7/4<h<2$:\n\\[\n f(x)=\\begin{cases}x^2/2,&x\\leq1,\\\\x-1/2,&x\\geq1,\\end{cases}\n \\qquad x_0=-\\frac95.\n\\]\nIts derivative is $x$ on $(-\\infty,1]$ and $1$ on $[1,\\infty)$, so it is convex and $1$-Lipschitz. The two subsequent iterates are\n$x_1=\\frac95(h-1)>1$ and $x_2=(4h-9)/5<1$. Direct substitution gives\n\\[\n \\kappa_0=\\frac{16h^2-252h+392}{50}\n =\\frac8{25}(h-\\tfrac74)(h-14)<0.\n\\]\nFor any prescribed $h\\geq2$, choose $t\\in(7/4,2)$ and replace this objective by $(t/h)f$. Its global smoothness bound is at most $1$, while its step-$h$ orbit is the step-$t$ orbit of $f$. Its second difference is multiplied by the positive number $t/h$ and remains negative.\n\\end{proof}\n\nThe last scaling argument uses a global smoothness \\emph{bound}, as universal theorems do. It does not assert that the prescribed bound is the smallest possible Lipschitz constant. When exact attainment of a bound is desired, an orthogonal quadratic coordinate initialized at zero can attain it without changing any observed loss.\n\n\\section{An exact two-step compensation certificate}\n\nDefine\n\\begin{equation}\n h_{\\mathrm{iso}}=\\frac{19-\\sqrt{17}}8,\n \\qquad h_{\\mathrm{iso}}\\approx1.859611796797792.\n \\label{eq:univ-threshold}\n\\end{equation}\nThe notation anticipates an isolation consequence, but the theorem itself concerns a particular weighted inequality.\n\n\\begin{theorem}[Two-step compensation]\n\\label{thm:compensation}\nFor $0<h\\leq h_{\\mathrm{iso}}$, every exact gradient-descent orbit of every convex $L$-smooth objective satisfies\n\\begin{equation}\n D_{k+1}+D_{k+2}\\leq2D_k,\n \\qquad\\text{equivalently}\\qquad\n 2\\kappa_k+\\kappa_{k+1}\\geq0.\n \\label{eq:univ-compensation}\n\\end{equation}\nNo minimizer or strong-convexity assumption is needed.\n\\end{theorem}\n\n\\begin{proof}\nNormalize $L=1$ and set\n\\[\n S=\\frac23f_0-f_1+\\frac13f_3\n =\\frac13(2\\kappa_0+\\kappa_1)\n =\\frac13(2D_0-D_1-D_2).\n\\]\nA finite interpolation certificate is\n\\begin{equation}\n \\frac79I_{01}+\\frac29I_{02}+\\frac13I_{20}\n +\\frac29I_{31}+\\frac19I_{32}=S-Q_h,\n \\label{eq:univ-exact-certificate}\n\\end{equation}\nwhere\n\\begin{align}\n Q_h(d_0,d_1,d_2)={}&\\frac{2-h}{3}\\norm{d_0}^2\n +\\frac{5-2h}{9}\\ip{d_0}{d_1}\n +\\frac{7-2h}{18}\\norm{d_1}^2 \\notag\\\\\n &+\\frac29\\ip{d_1}{d_2}+\\frac16\\norm{d_2}^2.\n \\label{eq:univ-Q}\n\\end{align}\nTo check the identity, expand each $I_{ij}$ using\n$x_i-x_j=-h(\\sum_{r<i}g_r-\\sum_{r<j}g_r)$.\nThe coefficients of the four function values are respectively\n$2/3,-1,0,1/3$. Substituting $g_1=g_0+d_0$, $g_2=g_0+d_0+d_1$ and $g_3=g_0+d_0+d_1+d_2$ cancels every term containing $g_0$. The remaining coefficients are exactly those in \\eqref{eq:univ-Q}.\n\nFor $0<h<2$, completion of squares gives\n\\begin{align}\n Q_h={}&\\frac{2-h}{3}\n \\left\\|d_0+\\frac{5-2h}{6(2-h)}d_1\\right\\|^2\n +\\frac16\\left\\|d_2+\\frac23d_1\\right\\|^2\\notag\\\\\n &+\\frac{8h^2-38h+43}{108(2-h)}\\norm{d_1}^2.\n \\label{eq:univ-square-completion}\n\\end{align}\nThe last numerator has roots $(19\\pm\\sqrt{17})/8$. It is nonnegative for $h\\leq h_{\\mathrm{iso}}$. Equations \\eqref{eq:univ-I} and \\eqref{eq:univ-exact-certificate} now imply $S\\geq0$.\n\nFor a general smoothness bound, apply the normalized proof to $f/L$, whose gradients are $g_i/L$. Multiplication by $L$ gives the useful stronger identity\n\\begin{equation}\n S=\\sum_{(i,j)}\\lambda_{ij}I^{(L)}_{ij}\n +\\frac1L Q_h(d_0,d_1,d_2),\n \\label{eq:univ-general-L}\n\\end{equation}\nwhere $I^{(L)}_{ij}$ has $1/(2L)$ in place of $1/2$ and the five weights are those in \\eqref{eq:univ-exact-certificate}. Translating the starting index proves the theorem.\n\\end{proof}\n\n\\begin{corollary}[Quantitative isolation]\n\\label{cor:quantitative-isolation}\nIn the same step interval, if $\\kappa_k<0$ then\n\\[\n \\kappa_{k+1}\\geq-2\\kappa_k>0.\n\\]\nIf $k\\geq1$, then also $\\kappa_{k-1}\\geq-\\kappa_k/2>0$. In particular, strictly negative second differences cannot occur at adjacent indices.\n\\end{corollary}\n\n\\begin{proof}\nApply \\eqref{eq:univ-compensation} at $k$ and, when available, at $k-1$.\n\\end{proof}\n\nOne should distinguish three assertions: each $\\kappa_k$ is nonnegative; the weighted combination $2\\kappa_k+\\kappa_{k+1}$ is nonnegative; and adjacent $\\kappa_k$ cannot both be negative. The first implies the second, which implies the third. The converses are not logical consequences. The next theorem is sharp for the \\emph{second} assertion.\n\n\\section{Sharpness, including analytic strong convexity}\n\n\\begin{theorem}[Exact boundary for two-step compensation]\n\\label{thm:sharp-compensation}\nFor every $h>h_{\\mathrm{iso}}$ there exists a real-analytic, strongly convex, globally $1$-smooth function on $\\R^3$ and an initial point for which exact step-$h$ gradient descent satisfies\n\\begin{equation}\n D_1+D_2>2D_0.\n \\label{eq:univ-sharp-failure}\n\\end{equation}\nConsequently $h_{\\mathrm{iso}}$ is the largest universal normalized-step threshold for \\eqref{eq:univ-compensation}, even if the function class is restricted to real-analytic strongly convex objectives in three dimensions. The strong-convexity parameter is positive but can depend on $h$.\n\\end{theorem}\n\nThe construction below uses finite interpolation. All its constants are specified by arithmetic operations, finite maxima, finite minima, and one Gaussian convolution. In particular, the proof contains no unspecified small perturbation.\n\n\\subsection{A polynomial family of four-point data}\n\nFirst fix\n\\begin{equation}\n h_{\\mathrm{iso}}<t\\leq\\frac{93}{50}=1.86,\n \\qquad A=\\frac{5-2t}{6(2-t)}.\n \\label{eq:univ-t-A}\n\\end{equation}\nIntroduce four scalars of each of three types:\n\\begin{align*}\n &(u_0,u_1,u_2,u_3)=(0,-A,1-A,\\tfrac13-A),\\\\\n &v_0=0,\\quad\n v_1=\\frac{-9A^2t-9A^2-3At+12A+6t-7}{18t},\\\\\n &v_2=\\frac{-A^2t-A^2+At+2A-1}{2t},\\qquad v_3=-10,\\\\\n &w_0=0,\\quad\n w_1=\\frac{A^2t}{2}+\\frac{At}{6}-\\frac{2A}{3}-\\frac t3+\\frac7{18},\\\\\n &w_2=\\frac{A^2t}{2}+A^2+\\frac{At}{6}-\\frac{5A}{3}-\\frac t3+\\frac89,\\\\\n &w_3=\\frac{A^2t}{2}+2A^2+\\frac{7At}{6}-\\frac{11A}{3}-\\frac{4t}{3}+\\frac{19}{9}.\n\\end{align*}\nLet $U_i=\\sum_{r<i}u_r$ and $V_i=\\sum_{r<i}v_r$, including $U_0=V_0=0$. For a positive parameter $e$ to be specified, define two-dimensional points, gradients and values by\n\\begin{equation}\n X_i=(-ti-te^2V_i,-teU_i),\\quad\n G_i=(1+e^2v_i,eu_i),\\quad\n F_i=-ti+e^2w_i,\\qquad 0\\leq i\\leq3.\n \\label{eq:univ-jet-data}\n\\end{equation}\nThey satisfy the gradient-descent update $X_{i+1}=X_i-tG_i$ exactly. They are provisional data; at this stage we have not asserted that a smooth convex function realizes them.\n\nFor each ordered pair $i\\ne j$, the interpolation residual expands \\emph{exactly} as\n\\begin{equation}\n F_i-F_j-\\ip{G_j}{X_i-X_j}-\\tfrac12\\norm{G_i-G_j}^2\n =e^2J_{ij}+e^4K_{ij},\n \\label{eq:univ-jet-expansion}\n\\end{equation}\nwhere\n\\begin{align}\n J_{ij}={}&w_i-w_j+t\\big[(i-j)v_j+V_i-V_j+u_j(U_i-U_j)\\big]\n -\\tfrac12(u_i-u_j)^2,\\label{eq:univ-J}\\\\\n K_{ij}={}&t v_j(V_i-V_j)-\\tfrac12(v_i-v_j)^2.\n \\label{eq:univ-K}\n\\end{align}\nThere are no omitted higher-order terms: both coordinates in \\eqref{eq:univ-jet-data} are polynomials of degrees at most two in $e$, and the interpolation expression is quadratic in coordinates and gradients.\n\n\\begin{lemma}[Nonnegative leading interpolation residuals]\n\\label{lem:jet-leading-slacks}\nFor every $t$ in \\eqref{eq:univ-t-A}, $J_{ij}\\geq0$ for all $i\\ne j$. The five residuals\n$J_{01},J_{02},J_{20},J_{31},J_{32}$ vanish exactly. The remaining seven are strictly positive.\n\\end{lemma}\n\n\\begin{proof}\nSubstitution in \\eqref{eq:univ-J} gives the following complete list of the nonzero expressions:\n\\begin{align*}\n J_{10}&=\\frac{9A^2(t-1)+3A(t-4)-6t+7}{18},\\\\\n J_{12}&=\\frac{6A(t-1)-3t-1}{9},\\\\\n J_{21}&=\\frac{6t-7-3A(t-1)}9,\\\\\n J_{30}&=\\frac{(A-1)[6t-7-3A(t-1)]}{6},\\\\\n J_{03}&=-\\frac{27A^2(t+1)-3At-42A+54tv_3-12t+23}{18},\\\\\n J_{13}&=-\\frac{9A^2(t+1)-3At-12A+18tv_3-3t+8}{9},\\\\\n J_{23}&=-\\frac{9A^2(t+1)+3At-18A+18tv_3-12t+17}{18}.\n\\end{align*}\nFor transparent bounds, $1859/1000<t\\leq93/50$ and\n$151/100<A<153/100$. The latter follows from\n$A=1/3+1/[6(2-t)]$. Inserting endpoints with the appropriate signs proves\n\\[\n J_{10}>\\tfrac15,\\quad J_{12}>\\tfrac{13}{100},\\quad\n J_{21}>\\tfrac1{50},\\quad J_{30}>\\tfrac3{200},\\quad\n J_{03}>44,\\quad J_{13}>29,\\quad J_{23}>13.\n\\]\nFor example,\n\\[\n 6t-7-3A(t-1)\n \\geq6\\frac{1859}{1000}-7\n -3\\frac{153}{100}\\frac{86}{100}\n =\\frac{1033}{5000}>0,\n\\]\nwhich establishes the claims for $J_{21}$ and $J_{30}$. For $J_{10}$ the same endpoint calculation gives\n$J_{10}\\geq36402631/180000000>1/5$.\nThe last three bounds use $v_3=-10$; even discarding the other negative terms in their numerators leaves the stated strictly positive lower bounds. This covers all twelve ordered pairs.\n\\end{proof}\n\nThe weighted curvature statistic of the provisional data is especially simple:\n\\begin{equation}\n \\frac23F_0-F_1+\\frac13F_3=e^2q(t),\\qquad\n q(t)=\\frac{8t^2-38t+43}{108(2-t)}<0.\n \\label{eq:univ-negative-jet}\n\\end{equation}\nOne can verify the equality directly from the $w_i$. Equivalently, the five vanishing $J_{ij}$ in Lemma~\\ref{lem:jet-leading-slacks} eliminate the interpolation terms in the leading part of \\eqref{eq:univ-exact-certificate}, while the transverse differences are\n$-A,1,-2/3$. Both squares in \\eqref{eq:univ-square-completion} vanish on these differences, leaving precisely $q(t)$.\n\n\\subsection{An exact positive-slack repair}\n\nThe terms $e^4K_{ij}$ in \\eqref{eq:univ-jet-expansion} may have either sign. A third, independent coordinate repairs all of them simultaneously.\nPut $r=1-t/2\\in(0,1)$ and consider the exact scalar quadratic orbit\n\\begin{equation}\n z_i=2r^i,\\qquad b_i=r^i,\\qquad a_i=r^{2i}.\n \\label{eq:univ-repair-orbit}\n\\end{equation}\nThese are respectively the points, gradients and values of $x\\mapsto x^2/4$, under step $t$. Its $1$-smooth interpolation residual is\n\\begin{equation}\n P_{ij}=a_i-a_j-b_j(z_i-z_j)-\\tfrac12(b_i-b_j)^2\n =\\tfrac12(r^i-r^j)^2>0\\quad(i\\ne j).\n \\label{eq:univ-positive-repair}\n\\end{equation}\nChoose the completely explicit positive constants\n\\begin{equation}\n c=1+\\max_{i\\ne j}\\frac{|K_{ij}|}{P_{ij}},\\qquad\n Q=\\frac23-r^2+\\frac13r^6\n =\\frac13(1-r^2)^2(r^2+2)>0,\n \\qquad e^2=\\frac{-q(t)}{2cQ}.\n \\label{eq:univ-repair-parameters}\n\\end{equation}\nThe maxima range over twelve pairs only. Define the final three-dimensional data by\n\\begin{equation}\n x_i=(X_i,\\sqrt c\\,e^2z_i),\\qquad\n g_i=(G_i,\\sqrt c\\,e^2b_i),\\qquad\n f_i=F_i+ce^4a_i.\n \\label{eq:univ-repaired-data}\n\\end{equation}\nTheir update remains exact: $x_{i+1}=x_i-tg_i$.\nMoreover, their interpolation residuals are\n\\begin{equation}\n I_{ij}=e^2J_{ij}+e^4(K_{ij}+cP_{ij})>0\\qquad(i\\ne j),\n \\label{eq:univ-strict-interpolation}\n\\end{equation}\nbecause $K_{ij}+cP_{ij}\\geq P_{ij}>0$.\nTheir curvature statistic is\n\\begin{equation}\n S=\\frac23f_0-f_1+\\frac13f_3\n =e^2q(t)+ce^4Q=\\frac12e^2q(t)<0.\n \\label{eq:univ-exact-negative-S}\n\\end{equation}\nThus the repair preserves the desired strict violation. The use of a separate coordinate is essential to this simple argument: interpolation residuals and loss statistics add under orthogonal direct sums.\n\n\\subsection{Strong interpolation and analytic smoothing}\n\nLet\n\\begin{equation}\n a=\\min_{i\\ne j}I_{ij}>0,\\quad\n b=\\max_{i\\ne j}\\norm{(x_i-x_j)-(g_i-g_j)}^2,\n \\qquad \\mu=\\frac{a}{a+b+1}\\in(0,1).\n \\label{eq:univ-sharp-mu}\n\\end{equation}\nThe strengthened interpolation condition for the class of $\\mu$-strongly convex, $1$-smooth functions is\n\\begin{equation}\n I_{ij}\\geq\\frac{\\mu}{2(1-\\mu)}\n \\norm{(x_i-x_j)-(g_i-g_j)}^2.\n \\label{eq:univ-strong-interpolation-condition}\n\\end{equation}\nFor the data above, its right-hand side is at most\n$ab/[2(b+1)]<a/2$. Hence Theorem~\\ref{thm:strong-interpolation} constructs a globally defined $\\mu$-strongly convex, $1$-smooth function $F:\\R^3\\to\\R$ realizing every pair $(f_i,g_i)$ at $x_i$. In particular its first four exact iterates are exactly the constructed points, and its statistic is \\eqref{eq:univ-exact-negative-S}.\n\nWe next make the function analytic while retaining a quantitatively certified violation. Write $\\delta=-S>0$ and set\n\\begin{equation}\n M=1+\\max_{0\\leq i\\leq3}\\norm{g_i}^2,\\quad\n C=4Mt+3+6t^2,\\qquad\n \\sigma=\\min\\left\\{1,\\frac{\\delta}{2C}\\right\\}.\n \\label{eq:univ-sharp-smoothing-scale}\n\\end{equation}\nDefine\n\\begin{equation}\n F_\\sigma(x)=\\E F(x-Z),\\qquad Z\\sim\\mathcal N(0,\\sigma^2I_3).\n \\label{eq:univ-sharp-smoothing}\n\\end{equation}\nA globally smooth function has at most quadratic growth, so the integral is finite. Gaussian convolution preserves the strong-convexity and smoothness inequalities. It is real analytic: replacing $x$ by $w\\in\\mathbb C^3$ in the Gaussian kernel defines an entire function, since on compact sets in $w$ every complex derivative is bounded by an integrable polynomial times a decaying Gaussian. No conjugates occur in the holomorphic kernel.\n\nLet $y_i$ be the step-$t$ orbit of $F_\\sigma$ starting at $y_0=x_0$. The gradient discrepancy is bounded globally by\n\\[\n \\norm{\\nabla F_\\sigma(x)-\\nabla F(x)}\n \\leq\\E\\norm{Z}\\leq\\sqrt3\\sigma.\n\\]\nFor convex $1$-smooth functions, the step-$t$ map is nonexpansive for $t\\leq2$; expand its squared distance and use \\eqref{eq:univ-coco}. Comparing the two maps therefore yields\n\\begin{equation}\n \\norm{y_i-x_i}\\leq ti\\sqrt3\\sigma.\n \\label{eq:univ-sharp-orbit-error}\n\\end{equation}\nAt a fixed point, Jensen's inequality and the centered smoothness estimate imply\n$0\\leq F_\\sigma(x)-F(x)\\leq3\\sigma^2/2$.\nIt follows, by applying smoothness at $x_i$, that\n\\begin{equation}\n |F_\\sigma(y_i)-f_i|\n \\leq Mt i\\sqrt3\\sigma+\\frac32(t^2i^2+1)\\sigma^2.\n \\label{eq:univ-sharp-value-error}\n\\end{equation}\nHere $\\norm{g_i}\\leq M$, which follows from the definition of $M$.\nTaking the absolute values of the coefficients $2/3,-1,1/3$ at indices $0,1,3$ gives\n\\begin{equation}\n |S_\\sigma-S|\n \\leq2Mt\\sqrt3\\sigma+(3+6t^2)\\sigma^2\n \\leq C\\sigma\\leq\\delta/2.\n \\label{eq:univ-sharp-statistic-error}\n\\end{equation}\nThe last inequality uses $\\sigma\\leq1$, $\\sqrt3<2$ and the definition in \\eqref{eq:univ-sharp-smoothing-scale}. Thus $S_\\sigma\\leq-\\delta/2<0$.\n\n\\begin{proof}[Completion of Theorem~\\ref{thm:sharp-compensation}]\nGiven an arbitrary $h>h_{\\mathrm{iso}}$, choose\n\\[\n t=\\frac12\\left(h_{\\mathrm{iso}}+\\min\\{h,93/50\\}\\right).\n\\]\nThen $h_{\\mathrm{iso}}<t< h$ and $t<93/50$. The preceding construction produces $F_\\sigma$ at step $t$. Set $\\widehat F=(t/h)F_\\sigma$. This function is real analytic, $(t/h)\\mu$-strongly convex and globally $1$-smooth. Its step-$h$ orbit is exactly the step-$t$ orbit of $F_\\sigma$, because $h\\nabla\\widehat F=t\\nabla F_\\sigma$. Its statistic is multiplied by $t/h>0$, and hence remains negative. This proves \\eqref{eq:univ-sharp-failure}.\n\\end{proof}\n\n\\begin{corollary}[Normalization by the optimal smoothness constant]\n\\label{cor:sharp-optimal-smoothness}\nIf the symbol $L$ is required to equal the smallest global gradient-Lipschitz constant, the same compensation boundary is sharp within real-analytic strongly convex functions in dimension four.\n\\end{corollary}\n\n\\begin{proof}\nLet $\\widehat F$ be the three-dimensional function in the completed proof, whose smoothness constant is at most $1$. Define\n$G(x,z)=\\widehat F(x)+z^2/2$ and initialize $z_0=0$. The extra coordinate remains zero at every gradient step, so the loss trace and its violation are unchanged. The added Hessian eigenvalue is exactly $1$ everywhere, while every eigenvalue of the first block is at most $1$. Hence the optimal global smoothness constant of $G$ is exactly $1$. Analyticity and positive strong convexity are preserved.\n\\end{proof}\n\n\\subsection{An exact rational audit case}\n\nAt $t=93/50$ the transverse coefficient is $A=32/21$, and the rational data simplify to\n\\[\n (v_1,v_2,v_3)=\\left(-\\frac{18712}{13671},-\\frac{19393}{41013},-10\\right),\n\\]\n\\[\n (w_1,w_2,w_3)=\\left(\\frac{15268}{11025},\\frac{59161}{22050},\\frac{45797}{11025}\\right),\n \\qquad q(t)=-\\frac1{4725}.\n\\]\nInstead of the exact balancing choice in \\eqref{eq:univ-repair-parameters}, one may use $c=10^{10}$ and $e=10^{-8}$. These rational choices satisfy $K_{ij}+cP_{ij}>0$ for all twelve pairs and $e^2q+ce^4Q<0$. Since $\\sqrt c=10^5$, all three-dimensional points, gradients and values are rational. The supplied exact verifier checks every interpolation inequality, every update and the strict sign of $S$ using fractions. It also checks the strong-convexity margin in \\eqref{eq:univ-strong-interpolation-condition}.\n\nThis rational audit checks a concrete counterexample and the coefficient identities underlying the construction. The proof for every $h>h_{\\mathrm{iso}}$ is the symbolic construction above. Numerical optimization of the four-point interpolation problem can suggest the threshold; it does not establish either the universal positive theorem or this all-parameter converse.\n\n\\section{A convex filtered curve and consequences at every horizon}\n\nThe compensation theorem can be interpreted as ordinary discrete convexity after a fixed local averaging operation. Define\n\\begin{equation}\n V_k=\\frac{2f_k+f_{k+1}}3,\\qquad\n E_k=V_k-V_{k+1}=\\frac{2D_k+D_{k+1}}3.\n \\label{eq:univ-filter}\n\\end{equation}\n\n\\begin{theorem}[Filtered convexity and horizon bounds]\n\\label{thm:filtered-convexity}\nUnder the assumptions of Theorem~\\ref{thm:compensation}, $E_k\\geq0$ and $E_{k+1}\\leq E_k$. Consequently, for every integer $N\\geq1$,\n\\begin{equation}\n 0\\leq V_k-V_{k+N}\\leq NE_k,\n \\label{eq:univ-filter-horizon}\n\\end{equation}\nand\n\\begin{equation}\n \\sum_{j=0}^{N-1}D_{k+j}\n \\leq NE_k+\\frac{D_k-D_{k+N}}3\n \\leq NE_k+\\frac{D_k}3.\n \\label{eq:univ-unfiltered-horizon}\n\\end{equation}\nThe first inequality in \\eqref{eq:univ-unfiltered-horizon} retains the terminal decrement and is therefore stronger than its second inequality.\n\\end{theorem}\n\n\\begin{proof}\nNonnegativity follows from \\eqref{eq:univ-descent}. Direct subtraction gives\n$E_k-E_{k+1}=(2\\kappa_k+\\kappa_{k+1})/3\\geq0$.\nSumming the first $N$ terms of the nonincreasing nonnegative sequence proves \\eqref{eq:univ-filter-horizon}. Finally,\n\\[\n \\sum_{j=0}^{N-1}E_{k+j}\n =\\sum_{j=0}^{N-1}D_{k+j}+\\frac{D_{k+N}-D_k}{3},\n\\]\nwhich rearranges to \\eqref{eq:univ-unfiltered-horizon}.\n\\end{proof}\n\nThe filter has a fixed meaning: the curve $V$ is a $2:1$ average of two consecutive losses. Its convexity does not imply that the original losses are convex. For finite-budget reasoning, however, it gives an exact linear upper bound on the improvement available over any specified future horizon once the two starting decrements are available. It remains a model-class bound and is not an assertion about arbitrary stochastic neural-network training.\n\nOne can also eliminate $D_{k+1}$ from a horizon bound, at the price of a slightly larger coefficient. The resulting coefficient has a complete sharp scalar characterization.\n\n\\begin{theorem}[Sharp scalar envelope implied by compensation]\n\\label{thm:scalar-horizon-envelope}\nLet $(d_j)_{j\\geq0}$ be any nonnegative scalar sequence satisfying\n\\begin{equation}\n d_{j+1}+d_{j+2}\\leq2d_j\\qquad(j\\geq0).\n \\label{eq:univ-scalar-compensation}\n\\end{equation}\nThen for every $N\\geq1$,\n\\begin{equation}\n \\sum_{j=1}^N d_j\\leq B_Nd_0,\\qquad\n B_N=\\frac{2(N+1)}{2+(-1/2)^N}-\\frac23.\n \\label{eq:univ-BN}\n\\end{equation}\nThis coefficient is optimal over precisely the scalar sequence class in \\eqref{eq:univ-scalar-compensation}. In particular, it yields a universal gradient-descent consequence in the compensation interval. Optimality over the smaller class of actual smooth-gradient trajectories is not asserted.\n\\end{theorem}\n\n\\begin{proof}\nDefine $\\lambda_{-1}=0$, $\\lambda_N=0$, and for $0\\leq j<N$ put\n\\begin{equation}\n \\lambda_j=\\frac{(N+1)[2+(-1/2)^j]}{3[2+(-1/2)^N]}\n -\\frac{j+1}{3}.\n \\label{eq:univ-horizon-multipliers}\n\\end{equation}\nThese satisfy $2\\lambda_j-\\lambda_{j-1}-\\lambda_{j-2}=-1$ for $1\\leq j\\leq N$, with the natural omission of $\\lambda_{-1}$ at the first index, and $2\\lambda_0=B_N$. All the $\\lambda_j$ are nonnegative. To see this, the sequence\n\\[\n a_j=\\frac{2+(-1/2)^j}{j+1}\n\\]\nis nonincreasing: the numerator of $a_j-a_{j+1}$, after multiplication by $(j+1)(j+2)$, is\n$2+(3j+5)(-1/2)^j/2$. It is positive at even indices; at $j=1$ it is zero; and at odd $j\\geq3$ its magnitude correction is less than $2$. Hence $a_j\\geq a_N$, exactly the assertion $\\lambda_j\\geq0$.\n\nCoefficient collection now gives the exact identity\n\\begin{equation}\n B_Nd_0-\\sum_{j=1}^N d_j\n =\\sum_{j=0}^{N-1}\\lambda_j(2d_j-d_{j+1}-d_{j+2})\n +\\lambda_{N-1}d_{N+1}.\n \\label{eq:univ-horizon-dual}\n\\end{equation}\nEvery term on its right is nonnegative, proving the inequality.\n\nFor sharpness take $d_0=1$, put $R=(-2)^{-(N+1)}$, and define\n\\[\n d_j=\\frac{1-R(-2)^j}{1-R}\\quad(0\\leq j\\leq N+1),\n \\qquad d_j=0\\quad(j>N+1).\n\\]\nThen $d_{N+1}=0$, and all earlier $d_j$ are nonnegative. For $j\\leq N-1$, they satisfy $2d_j=d_{j+1}+d_{j+2}$ because the characteristic roots of the equality recurrence are $1$ and $-2$. At $j=N$ and afterward the inequality follows from nonnegativity and the zero tail. Therefore the right side of \\eqref{eq:univ-horizon-dual} vanishes and equality is attained.\n\\end{proof}\n\nFor orientation, $B_1=2$, $B_2=2$, $B_3=18/5$, and $B_4=46/11$. The exact expression has asymptotic form\n$B_N=N+1/3+O(N2^{-N})$. Thus compensation alone gives a future gain envelope whose leading coefficient is one, even though an individual future decrement may exceed the current decrement. This is an all-horizon statement, not a new claim that the raw average satisfies $\\sum_{j=1}^ND_{k+j}\\leq ND_k$ for every $N$.\n\n\\section{What isolation says about cumulative curvature}\n\nThere are useful restrictions on the arrangement and mass of negative curvature, not only its immediate neighbor.\n\n\\begin{proposition}[Density and variation restrictions]\n\\label{prop:univ-variation}\nSuppose \\eqref{eq:univ-compensation} holds for every $k\\geq0$. Among any $n$ consecutive second differences, at most $\\lceil n/2\\rceil$ are strictly negative. If $A$ is a finite set of negative-curvature indices, then\n\\begin{equation}\n \\sum_{k\\in A}|\\kappa_k|\n \\leq\\frac12\\sum_{k\\in A}\\kappa_{k+1}.\n \\label{eq:univ-negative-mass}\n\\end{equation}\nThe positive terms on the right have distinct indices. If the sums of the positive and negative parts of the entire curvature sequence are finite, then the total negative part is at most one half of the total positive part.\n\\end{proposition}\n\n\\begin{proof}\nThe count follows because negative indices cannot be adjacent. For each negative index, Corollary~\\ref{cor:quantitative-isolation} gives $|\\kappa_k|\\leq\\kappa_{k+1}/2$, and the map $k\\mapsto k+1$ is injective. Summation proves \\eqref{eq:univ-negative-mass}. Taking an increasing union of finite sets proves the final assertion when the sums are finite.\n\\end{proof}\n\nThe summability qualification is needed: a local compensation law alone does not license subtracting two divergent sums. For actual strongly convex gradient descent with $h<2$, the loss and gradients converge geometrically, so the absolute curvature series is finite and the last conclusion applies. The quantitative isolation bound also explains why a single concave step is not, by itself, evidence of an imminent sustained acceleration regime.\n\n\\section{Exact defect terms for inexact and variable-step updates}\n\nMeasurement noise changes observed values. Update errors change the points at which the function is evaluated. These are different mathematical perturbations and require different allowances.\n\nConsider four points with true gradients $g_i=\\nabla f(x_i)$, but suppose the three updates are\n\\begin{equation}\n x_{i+1}=x_i-\\eta g_i+r_i,\\qquad i=0,1,2,\n \\label{eq:univ-residual-update}\n\\end{equation}\nwhere $r_i$ is an arbitrary displacement residual. Keep $h=\\eta L$ and\n$S=(2/3)f_0-f_1+(1/3)f_3$.\n\n\\begin{theorem}[Compensation with exact update defects]\n\\label{thm:univ-inexact}\nIf $f$ is convex and globally $L$-smooth and $0<h\\leq h_{\\mathrm{iso}}$, then\n\\begin{equation}\n S\\geq\\frac1LQ_h(d_0,d_1,d_2)+\\sum_{i=0}^2\\ip{a_i}{r_i},\n \\label{eq:univ-defect-law}\n\\end{equation}\nwhere\n\\begin{equation}\n a_0=\\frac{3g_0-7g_1-2g_2}{9},\\quad\n a_1=\\frac{3g_0+2g_1-2g_2}{9},\\quad\n a_2=\\frac{2g_1+g_2}{9}.\n \\label{eq:univ-defect-vectors}\n\\end{equation}\nIn particular, if $\\norm{r_i}\\leq\\rho_i$, then\n\\begin{equation}\n S\\geq-\\sum_{i=0}^2\\norm{a_i}\\rho_i.\n \\label{eq:univ-defect-bound}\n\\end{equation}\nIf also $\\norm{g_i}\\leq G$ for $i=0,1,2$ and $\\rho_i\\leq\\rho$, then\n$S\\geq-(22/9)G\\rho$.\n\\end{theorem}\n\n\\begin{proof}\nLet $e_0=0$, $e_i=\\sum_{j<i}r_j$. Relative to the ideal positions\n$x_0-\\eta\\sum_{j<i}g_j$, the actual positions differ by $e_i$. Each actual interpolation residual is therefore the ideal algebraic residual minus $\\ip{g_j}{e_i-e_j}$. Repeating the exact coefficient identity \\eqref{eq:univ-general-L} gives\n\\[\n S=\\sum\\lambda_{ij}I^{(L)}_{ij}+\\frac1LQ_h\n +\\sum\\lambda_{ij}\\ip{g_j}{e_i-e_j}.\n\\]\nCollecting the coefficients of $r_0,r_1,r_2$ in the last sum gives exactly \\eqref{eq:univ-defect-vectors}. Nonnegativity of the actual interpolation residuals proves \\eqref{eq:univ-defect-law}. Cauchy--Schwarz gives \\eqref{eq:univ-defect-bound}. Finally,\n\\[\n \\norm{a_0}\\leq\\frac43G,\\qquad\n \\norm{a_1}\\leq\\frac79G,\\qquad\n \\norm{a_2}\\leq\\frac13G,\n\\]\nwhose sum is $22G/9$.\n\\end{proof}\n\nFor inexact gradients, write $x_{i+1}=x_i-\\eta(g_i+\\xi_i)$. Then $r_i=-\\eta\\xi_i$, and a known bound $\\norm{\\xi_i}\\leq\\varepsilon_i$ gives\n\\[\n S\\geq-\\eta\\sum_{i=0}^2\\norm{a_i}\\varepsilon_i.\n\\]\nThe bound concerns true gradients; replacing them by observed gradients requires its own triangle-inequality allowance.\n\nFor a variable-step run $x_{i+1}=x_i-\\eta_i g_i$, fix any reference $\\eta$ with $\\eta L\\leq h_{\\mathrm{iso}}$ and use\n$r_i=(\\eta-\\eta_i)g_i$ in \\eqref{eq:univ-defect-law}. This yields an exact correction term\n$\\sum_i(\\eta-\\eta_i)\\ip{a_i}{g_i}$. It does not assert that the uncorrected constant-step inequality automatically survives a variable schedule.\n\n\\section{Observable loss diagnostics}\n\nSuppose measured values satisfy deterministic bounds\n$|\\widehat f_i-f_i|\\leq\\nu_i$. Define\n\\begin{equation}\n \\widehat S=\\frac23\\widehat f_0-\\widehat f_1+\\frac13\\widehat f_3,\n \\qquad B_\\nu=\\frac23\\nu_0+\\nu_1+\\frac13\\nu_3.\n \\label{eq:univ-measurement}\n\\end{equation}\nUnder exact updates in the compensation interval,\n\\begin{equation}\n \\widehat S\\geq-B_\\nu.\n \\label{eq:univ-loss-diagnostic}\n\\end{equation}\nUnder the update-error assumptions of Theorem~\\ref{thm:univ-inexact}, the valid necessary condition is instead\n\\begin{equation}\n \\widehat S\\geq-B_\\nu-\\sum_{i=0}^2\\norm{a_i}\\rho_i.\n \\label{eq:univ-combined-diagnostic}\n\\end{equation}\nBoth follow by the triangle inequality. For a common value-error bound $\\nu$, $B_\\nu=2\\nu$. A common additive calibration offset cancels because the three coefficients sum to zero; bounded independent offsets still require the displayed allowance.\n\nA violation of \\eqref{eq:univ-loss-diagnostic} rejects the conjunction of the declared assumptions. It does not identify which assumption failed. A passing statistic is only a necessary condition and cannot establish global convexity, the claimed global smoothness bound, exact gradients, or the absence of distribution shift. The stronger transcript tests elsewhere in this manuscript can use gradients and iterates when available. The attraction of the loss-only test is its small observational requirement, not logical sufficiency.\n\n\\section{Resolved and unresolved threshold questions}\n\nTheorem~\\ref{thm:sharp-compensation} completely resolves the sharp threshold for the universal inequality $D_{k+1}+D_{k+2}\\leq2D_k$. Its negative direction holds under particularly strong regularity and convergence assumptions. This sharpness result neither follows from a numerical semidefinite optimum nor from the existence of long concave runs; it requires the explicit interpolation construction above.\n\nThe prohibition of two adjacent negative second differences is a different question: violating the weighted average law need not make both second differences negative. Chapter~\\ref{ch:adjacency} determines that separate threshold exactly and shows that it is strictly larger than $h_{\\mathrm{iso}}$. The exact threshold for\n$\\sum_{j=1}^ND_{k+j}\\leq ND_k$ for arbitrary $N$ remains separate; Theorem~\\ref{thm:scalar-horizon-envelope} does not identify it. Strong-convexity-sensitive optimal thresholds likewise require retaining a fixed positive condition parameter rather than allowing it to depend on the counterexample. These distinctions prevent the resolved threshold questions from being silently substituted for a complete phase diagram.\n"}
{"format": "latex", "id": "03b_adjacency", "sha256": "0103a4a4ae60ae3c9e6d6244d1221e4637e8975dd74c6dcc0b2a44a23959a17a", "source_path": "research/chapters/03b_adjacency.tex", "text": "\\chapter{The sharp boundary for adjacent concavity}\n\\label{ch:adjacency}\n\nThe sharp compensation threshold of Chapter~\\ref{ch:universal} does not by itself identify when two consecutive negative second differences first become possible. A different positive weighting extends the universal prohibition slightly farther. This chapter determines the exact endpoint and constructs simultaneous-concavity counterexamples immediately above it, including analytic strongly convex realizations.\n\n\\section{An algebraic constant and the complete threshold theorem}\n\nDefine the polynomial\n\\begin{align}\n P(h)={}&5000h^7-33602h^6+97161h^5-175306h^4\\notag\\\\\n &+235672h^3-224304h^2+126080h-33536.\n \\label{eq:adj-polynomial}\n\\end{align}\nLet $h_*$ be its unique root in the rational interval\n\\begin{equation}\n \\ell=\\frac{1863425364}{10^9}<h_*<\n u=\\frac{1863425365}{10^9}.\n \\label{eq:adj-root-interval}\n\\end{equation}\nExistence and uniqueness in this interval are proved below; they are part of the definition's justification. Its numerical value is\n\\[\n h_*=1.863425364363717532\\ldots.\n\\]\nOnly the isolated algebraic definition, not this decimal approximation, enters the proof.\n\n\\begin{theorem}[Sharp threshold for adjacent discrete concavity]\n\\label{thm:sharp-adjacency}\nLet $f$ be convex and globally $L$-smooth, and let constant-step exact gradient descent have normalized step $h=\\eta L>0$.\n\nIf $h\\leq h_*$, no two adjacent second differences can both be strictly negative. More precisely, the positive algebraic constant $r_*$ defined below satisfies, in every dimension,\n\\begin{equation}\n r_*\\kappa_k+\\kappa_{k+1}\\geq0\\qquad(k\\geq0).\n \\label{eq:adj-weighted-law}\n\\end{equation}\n\nFor every $h>h_*$ there is a real-analytic, strongly convex, globally $1$-smooth objective on $\\R^3$ whose step-$h$ trajectory satisfies\n\\begin{equation}\n \\kappa_0<0,\\qquad \\kappa_1<0.\n \\label{eq:adj-negative-pair}\n\\end{equation}\nThus $h_*$ is the exact universal boundary for the possibility of adjacent discrete concavity. The negative statement still holds when normalization uses the optimal global smoothness constant, with dimension four sufficing.\n\\end{theorem}\n\nThe last distinction concerns the normalization convention, not the curve phenomenon: a function with a gradient-Lipschitz constant below $1$ is globally $1$-smooth. An unused orthogonal quadratic coordinate can force the best constant to equal $1$ while leaving the trajectory losses unchanged.\n\n\\section{A one-parameter family of exact interpolation certificates}\n\nNormalize $L=1$, and write $d_i=g_{i+1}-g_i$ as before. For a parameter $r>0$, let\n\\begin{equation}\n M(r,h)=\\begin{pmatrix}\n r(1-h/2)&r(h+2)/12-(h-1)/2&0\\\\\n r(h+2)/12-(h-1)/2&r(h+1)/3-h+1/2&r/6\\\\\n 0&r/6&1/2\n \\end{pmatrix}.\n \\label{eq:adj-matrix}\n\\end{equation}\nFor vector-valued differences, the notation $d^TMd$ means\n$\\sum_{i,j=0}^2M_{ij}\\ip{d_i}{d_j}$. Direct expansion gives\n\\begin{align}\n r\\kappa_0+\\kappa_1={}&\n \\left(\\frac{5r}{3}-1\\right)I_{01}\n +\\left(1-\\frac r6\\right)I_{02}\n +\\frac r2I_{20}\\notag\\\\\n &+\\frac r3I_{31}\n +\\left(1-\\frac r3\\right)I_{32}\n +d^TM(r,h)d.\n \\label{eq:adj-certificate}\n\\end{align}\nEvery interpolation coefficient is nonnegative for $3/5\\leq r\\leq3$. Thus positive semidefiniteness of $M(r,h)$ proves a universal weighted curvature law. Equation~\\eqref{eq:adj-certificate} contains the compensation certificate of the preceding chapter as the case $r=2$, after division by three.\n\nDefine the cubic in $r$\n\\begin{align}\n D(r,h):={}&-288\\det M(r,h)\\notag\\\\\n ={}&4(2-h)r^3+(25h^2-20h-44)r^2\\notag\\\\\n &+(-84h^2+168h-48)r+36(h-1)^2.\n \\label{eq:adj-D}\n\\end{align}\nA tangent zero of this cubic produces the improved boundary. We next isolate it exactly and verify all positivity conditions needed for the matrix argument.\n\n\\section{Exact isolation of the boundary parameters}\n\nFor a cubic $ar^3+br^2+cr+d$, put\n\\[\n H=b^2-3ac,\\qquad\n \\Delta=b^2c^2-4ac^3-4b^3d-27a^2d^2+18abcd.\n\\]\nWhen $H\\ne0$, substitution of $r=(9ad-bc)/(2H)$ gives the elementary identities\n\\begin{align}\n 3ar^2+2br+c&=-\\frac{9a\\Delta}{4H^2},\\label{eq:adj-cubic-derivative}\\\\\n ar^3+br^2+cr+d&=-\\frac{(27a^2d-9abc+2b^3)\\Delta}{8H^3}.\n \\label{eq:adj-cubic-value}\n\\end{align}\nThey are verified by multiplying through by the denominators and expanding. They give a direct algebraic route to a repeated root without a numerical root-selection step.\n\nFor the coefficients of \\eqref{eq:adj-D}, set\n\\begin{equation}\n a(h)=4(2-h),\\quad b(h)=25h^2-20h-44,\\quad\n c(h)=-84h^2+168h-48,\\quad d(h)=36(h-1)^2.\n \\label{eq:adj-abcd}\n\\end{equation}\nTheir discriminant has the exact factorization\n\\begin{equation}\n \\Delta(h)=432(h-2)P(h).\n \\label{eq:adj-discriminant}\n\\end{equation}\nAt $h=h_*$, $a>0$, $b>0$, $c<0$, and hence $H=b^2-3ac>0$. Define\n\\begin{equation}\n r_*=\\frac{9a(h_*)d(h_*)-b(h_*)c(h_*)}{2H(h_*)}.\n \\label{eq:adj-rstar}\n\\end{equation}\nEquations~\\eqref{eq:adj-cubic-derivative}--\\eqref{eq:adj-discriminant} give\n\\begin{equation}\n D(r_*,h_*)=0,\\qquad \\partial_rD(r_*,h_*)=0.\n \\label{eq:adj-tangency}\n\\end{equation}\n\n\\begin{lemma}[Rational isolation and parameter bounds]\n\\label{lem:adj-isolation}\nThe polynomial $P$ has exactly one zero in $(\\ell,u)$ from \\eqref{eq:adj-root-interval}. The parameters above satisfy\n\\begin{equation}\n \\frac{187988820}{10^8}<r_*<\\frac{187988824}{10^8},\n \\qquad\n \\frac{13517}{10000}<A_*<\\frac{6759}{5000},\n \\label{eq:adj-parameter-bounds}\n\\end{equation}\nwhere\n\\begin{equation}\n A_*=\\frac{r_*(h_*+2)-6(h_*-1)}{6r_*(2-h_*)}.\n \\label{eq:adj-Astar}\n\\end{equation}\nIn particular, $3/5<r_*<3$ and $0<h_*<2$.\n\\end{lemma}\n\n\\begin{proof}\nEvery numerical inequality in this proof is between explicit rational numbers. Evaluation at the two endpoints gives\n\\[\n -\\frac{340}{10^9}<P(\\ell)<-\\frac{339}{10^9},\n \\qquad\n \\frac{594}{10^9}<P(u)<\\frac{595}{10^9}.\n\\]\nContinuity therefore gives a zero. For uniqueness, let\n$c_0=(\\ell+u)/2$ and $w=(u-\\ell)/2=1/(2\\cdot10^9)$. The finite Taylor expansion of the polynomial gives\n\\[\n 933<P'(c_0)<934,\n \\qquad\n \\sum_{j=1}^{6}\\frac{|P^{(j+1)}(c_0)|}{j!}w^j<\\frac1{100000}.\n\\]\nConsequently $P'(h)>933-1/100000>0$ throughout $[\\ell,u]$. These two displayed checks require only integer arithmetic: multiply the polynomial and each derivative by the positive common denominator of $c_0$ and $w$ to obtain integer inequalities. The supplied exact verifier performs those operations without floating-point tolerances.\n\nFor explicit parameter bounds, the rational function in \\eqref{eq:adj-rstar} simplifies to\n\\begin{equation}\n R(h)=\\frac{6(175h^4-598h^3+504h^2-4h+40)}\n {625h^4-2008h^3+2232h^2-2848h+3088}.\n \\label{eq:adj-rational-R}\n\\end{equation}\nFor a polynomial $q$, its finite Taylor expansion gives the interval enclosure\n\\[\n q([\\ell,u])\\subseteq\n \\left[q(c_0)-E_q,\\ q(c_0)+E_q\\right],\n \\qquad E_q=\\sum_{j=1}^{\\deg q}\\frac{|q^{(j)}(c_0)|}{j!}w^j.\n\\]\nApply it separately to the numerator and denominator of \\eqref{eq:adj-rational-R}. The denominator is positive; division of the resulting endpoint bounds gives\n\\[\n 1.8798882025<R(h)<1.8798882351\n \\qquad(h\\in[\\ell,u]),\n\\]\nwhere the displayed terminating decimals are rational bounds rounded outward. They imply the looser rational interval for $r_*$ in \\eqref{eq:adj-parameter-bounds}.\n\nThe function\n$A(r,h)=[r(h+2)-6(h-1)]/[6r(2-h)]$\nis increasing in both variables on the isolated rectangle. Its derivative in $r$ is positive because $h>1$, and\n\\[\n \\partial_h A(r,h)=\\frac{4-6/r}{6(2-h)^2}>0\n\\]\nbecause $r>3/2$. Substitution of the rational rectangle endpoints gives, with outward rounding,\n\\[\n 1.35170122<A(r,h)<1.35170131.\n\\]\nThis implies the claimed interval for $A_*$. The elementary signs of $a,b,c,H$ on the same rectangle follow either from their formulas or the same Taylor bound.\n\\end{proof}\n\nThe root interval is intentionally narrow enough to make all later estimates elementary. The exact value is still the isolated root of a degree-seven polynomial. The proof uses no assertion that the displayed decimal approximation is itself exact.\n\n\\section{The universal prohibition up to the boundary}\n\n\\begin{proposition}[A positive matrix certificate on the whole interval]\n\\label{prop:adj-positive}\nFor the fixed algebraic parameter $r_*$, $M(r_*,h)$ is positive semidefinite for every $0<h\\leq h_*$. Consequently \\eqref{eq:adj-weighted-law} holds throughout this interval.\n\\end{proposition}\n\n\\begin{proof}\nAt $h=h_*$, the first and last diagonal entries of $M$ are positive. Eliminating those coordinates by Schur complements leaves the scalar\n\\[\n M_{11}-\\frac{M_{01}^2}{M_{00}}-\\frac{M_{12}^2}{M_{22}}\n =\\frac{\\det M}{M_{00}M_{22}}=0\n\\]\nby \\eqref{eq:adj-tangency}. Thus $M(r_*,h_*)\\succeq0$.\n\nFor a fixed $r$, the matrix is affine in $h$, and\n\\begin{equation}\n M(r,h)=M(r,h_*)+(h_*-h)\n \\begin{pmatrix}\n r/2&1/2-r/12&0\\\\\n 1/2-r/12&1-r/3&0\\\\\n 0&0&0\n \\end{pmatrix}.\n \\label{eq:adj-matrix-monotone}\n\\end{equation}\nThe upper $2\\times2$ block of the second matrix is positive definite at $r=r_*$. Its first diagonal entry is positive and its determinant is\n\\[\n \\frac{-25r_*^2+84r_*-36}{144}>0.\n\\]\nFor example, $1.87<r_*<1.89$, and the concave numerator is greater than $33$ at both endpoints of this larger interval. Hence the second matrix is positive semidefinite. Equation~\\eqref{eq:adj-matrix-monotone} proves the claim for every $h\\leq h_*$, not just for a small interval near the endpoint.\n\nAll interpolation weights in \\eqref{eq:adj-certificate} are nonnegative by Lemma~\\ref{lem:adj-isolation}. The certificate proves $r_*\\kappa_0+\\kappa_1\\geq0$. Translating the start proves the statement at every index. For general $L$, division of the objective by $L$ gives the same curvature inequality, with $h=\\eta L$.\n\\end{proof}\n\nSince $r_*>0$, two strictly negative adjacent curvatures would violate \\eqref{eq:adj-weighted-law}. This proves the positive half of Theorem~\\ref{thm:sharp-adjacency}. The following construction proves that its endpoint cannot be enlarged.\n\n\\section{An explicit simultaneous-concavity counterexample above the boundary}\n\nUse the isolated parameters $(h_*,r_*)$ constructed above, which satisfy $D(r,h)=D_r(r,h)=0$\nwith\n\\[\n 1.863425364<h_*<1.863425365,\\qquad\n 1.87988820<r_*<1.87988824,\n\\]\nwhere $D=-288\\det M$ and\n\\[\n M(r,h)=\n \\begin{pmatrix}\n r(1-h/2)&r(h+2)/12-(h-1)/2&0\\\\\n r(h+2)/12-(h-1)/2&r(h+1)/3-h+1/2&r/6\\\\\n 0&r/6&1/2\n \\end{pmatrix}.\n\\]\nDefine\n\\[\n A_*=\\frac{r_*(h_*+2)-6(h_*-1)}{6r_*(2-h_*)},\n \\qquad B_*=\\frac{r_*}{3}.\n\\]\nThe certified isolating rectangle implies\n\\[\n \\frac{13517}{10000}<A_*<\\frac{6759}{5000},\n \\qquad \\frac58<B_*<\\frac{157}{250}.\n\\]\nFor a new step $t$ with\n\\[\n h_*<t\\leq\\frac{233}{125}=1.864,\n\\]\nkeep $A=A_*$ fixed and set\n\\begin{equation}\n C(t)=3A^2(t-2)+A(7t-4)-8t+1,\\qquad\n B(t)=\\frac{-2+\\sqrt{4-12C(t)}}6.\n \\label{eq:adj-negative-B}\n\\end{equation}\n\n\\begin{lemma}[A uniformly negative common leading curvature]\nThe quantities in \\eqref{eq:adj-negative-B} are real and satisfy\n$5/8<B(t)<157/250$. The two expressions\n\\[\n k_0(t)=-\\frac{3A^2t-6A^2+At+2A+2B(t)-2t-2}{6},\n \\qquad\n k_1(t)=\\frac{2A(t-1)+B(t)^2-2t+1}{2}\n\\]\nare equal. Their common value $k(t)$ obeys\n\\begin{equation}\n k(t)\\leq-\\frac38(t-h_*)<0.\n \\label{eq:adj-leading-negative}\n\\end{equation}\n\\end{lemma}\n\n\\begin{proof}\nThe difference $k_0-k_1$ is\n$-[3B(t)^2+2B(t)+C(t)]/6$, which vanishes by definition.\nTo prove the initial equality $k(h_*)=0$, define\n\\[\n v(r,h)=(-A(r,h),1,-r/3)^\\top,\\qquad\n A(r,h)=\\frac{M_{01}(r,h)}{M_{00}(r,h)}.\n\\]\nEliminating the first and third coordinates gives\n\\[\n v^\\top Mv\n =M_{11}-\\frac{M_{01}^2}{M_{00}}-\\frac{M_{12}^2}{M_{22}}\n =\\frac{\\det M}{M_{00}M_{22}}.\n\\]\nAt $(r_*,h_*)$, this is zero and $Mv=0$. Differentiating the displayed\nidentity with respect to $r$ gives\n$v^\\top M_rv=0$, because $\\partial_r\\det M=-D_r/288=0$ and the terms\ncontaining derivatives of $v$ vanish when $Mv=0$. Direct expansion, for\narbitrary fixed $A,B,h,r$, gives\n\\[\n (-A,1,-B)\\,M(r,h)\\,(-A,1,-B)^\\top\n =r k_0(A,B,h)+k_1(A,B,h).\n\\]\nThe matrix $M$ is affine in $r$. Thus the derivative with fixed $A,B,h$\nis $k_0$, and the preceding zero derivative proves $k_0(A_*,B_*,h_*)=0$.\nThe zero quadratic form then gives $k_1(A_*,B_*,h_*)=0$.\nSince $B_*>0$, it is the positive root of\n$3B^2+2B+C(h_*)=0$. Hence the branch in\n\\eqref{eq:adj-negative-B} starts at $B(h_*)=B_*$.\n\nAll interval bounds used next are rational. Work in the larger rectangle\n\\[\n A\\in[13517/10000,6759/5000],\\qquad\n t\\in[1863/1000,233/125].\n\\]\nBoth partial derivatives of $C$ are positive there:\n$C_A=6A(t-2)+7t-4>0$ and $C_t=3A^2+7A-8>0$.\nEndpoint substitution gives\n\\[\n -\\frac{243421547779}{100000000000}\n \\leq C(t)\\leq\n -\\frac{7582743131}{3125000000}<0.\n\\]\nThe discriminant is therefore positive. At the lower proposed root\n$B=5/8$, even the largest $C$ makes $3B^2+2B+C$ equal to\n$-3595939/781250000<0$. At $B=157/250$, even the smallest $C$\nmakes it equal to $493652221/100000000000>0$.\nSince this polynomial is strictly increasing for $B\\geq0$, the asserted\nroot bounds follow.\n\nDifferentiating $3B^2+2B+C=0$ yields\n\\[\n B'(t)=-\\frac{3A^2+7A-8}{6B(t)+2}.\n\\]\nConsequently\n\\[\n k'(t)=A-1-\\frac{B(t)(3A^2+7A-8)}{6B(t)+2}.\n\\]\nUsing the rational interval endpoints separately gives the valid upper bound\n\\[\n k'(t)\\leq\n \\frac{6759}{5000}-1\n -\\frac{(5/8)[3(13517/10000)^2+7(13517/10000)-8]}\n {6(157/250)+2}\n =-\\frac{369648683}{922880000}<-\\frac38.\n\\]\nIntegration from $h_*$ proves \\eqref{eq:adj-leading-negative}.\n\\end{proof}\n\nWe now construct a four-point first-order transcript with both negative\ncurvatures. In all following formulas $B=B(t)$, $A=A_*$, and $v_3=-10$.\nPut\n\\[\n (u_0,u_1,u_2,u_3)=(0,-A,1-A,1-A-B),\n\\]\n\\begin{align*}\n v_0&=0,\\\\\n v_1&=\\frac{-3A^2t-3A^2-At+4A-2B+2t-1}{6t},\\\\\n v_2&=\\frac{-A^2t-A^2+At+2A-1}{2t},\n\\end{align*}\nand\n\\begin{align*}\n w_0&=0,\\\\\n w_1&=\\frac{A^2t}{2}+\\frac{At}{6}-\\frac{2A}{3}\n +\\frac B3-\\frac t3+\\frac16,\\\\\n w_2&=\\frac{A^2t}{2}+A^2+\\frac{At}{6}-\\frac{5A}{3}\n +\\frac B3-\\frac t3+\\frac23,\\\\\n w_3&=\\frac{A^2t}{2}+2A^2+\\frac{7At}{6}-\\frac{11A}{3}\n +\\frac{B^2}{2}+\\frac B3-\\frac{4t}{3}+\\frac53.\n\\end{align*}\nFor $U_i=\\sum_{j<i}u_j$, $V_i=\\sum_{j<i}v_j$, and a positive parameter $e$,\ndefine\n\\[\n X_i=(-ti-te^2V_i,-teU_i),\\qquad\n G_i=(1+e^2v_i,eu_i),\\qquad\n F_i=-ti+e^2w_i.\n\\]\nThen $X_{i+1}=X_i-tG_i$ exactly. Its interpolation residual is\n\\[\n F_i-F_j-\\langle G_j,X_i-X_j\\rangle-\\frac12\\|G_i-G_j\\|^2\n =e^2J_{ij}+e^4K_{ij},\n\\]\nwith\n\\[\n J_{ij}=w_i-w_j+t[(i-j)v_j+V_i-V_j+u_j(U_i-U_j)]\n -\\frac12(u_i-u_j)^2,\\qquad\n K_{ij}=tv_j(V_i-V_j)-\\frac12(v_i-v_j)^2.\n\\]\nThe five residuals $J_{01},J_{02},J_{20},J_{31},J_{32}$ vanish identically.\nThe remaining seven have the exact expressions\n\\begin{align*}\n J_{10}&=\\frac{3A^2(t-1)+A(t-4)+2B-2t+1}{6},\\\\\n J_{12}&=\\frac{2A(t-1)+B-t-1}{3},\\\\\n J_{21}&=\\frac{2t-1-A(t-1)-2B}{3},\\\\\n J_{30}&=\\frac{A-1}{2}[2t-1-A(t-1)-2B],\\\\\n J_{03}&=-\\frac{\n 3A^2(t+1)+4ABt+2AB-3At-6A+2B^2-2Bt-2B-60t+3}{2},\\\\\n J_{13}&=-\\frac{\n 3A^2(t+1)+6ABt-5At-4A+3B^2-3Bt-4B-59t+4}{3},\\\\\n J_{23}&=-\\frac{\n A^2(t+1)+2ABt-At-2A+2B^2-2Bt-20t+1}{2}.\n\\end{align*}\nThe same rational rectangle used above gives\n\\[\n J_{10}>\\frac3{50},\\quad J_{12}>\\frac3{100},\\quad\n J_{21}>\\frac1{10},\\quad J_{30}>\\frac1{20},\\quad\n J_{03}>42,\\quad J_{13}>26,\\quad J_{23}>13.\n\\]\nFor a fully reproducible endpoint check, lower bounds for the first four are,\nrespectively,\n\\[\n \\frac{36354689221}{600000000000},\\quad\n \\frac{470171}{15000000},\\quad\n \\frac{31463}{312500},\\quad\n \\frac{331966113}{6250000000}.\n\\]\nFor the last three, discarding their negative numerator terms other than\nthe large $-60t,-59t,-20t$ terms gives the positive lower bounds\n\\[\n \\frac{131661001843}{3125000000},\\quad\n \\frac{124279227263}{4687500000},\\quad\n \\frac{42301329921}{3125000000}.\n\\]\nThus all leading interpolation residuals are nonnegative throughout the\nwhole interval, not only at a numerical sample.\n\nThe provisional second differences are exactly\n$F_0-2F_1+F_2=e^2k(t)$ and $F_1-2F_2+F_3=e^2k(t)$.\nTo repair the fourth-order interpolation residuals, put\n$r=1-t/2\\in(0,1)$ and use the scalar quadratic orbit\n$z_i=2r^i$, $b_i=r^i$, $a_i=r^{2i}$. Its interpolation residuals are\n$P_{ij}=\\frac12(r^i-r^j)^2>0$. Define\n\\[\n c=1+\\max_{i\\ne j}\\frac{|K_{ij}|}{P_{ij}},\\qquad\n Q_0=(1-r^2)^2,\\qquad Q_1=r^2Q_0,\\qquad\n e^2=\\frac{-k(t)}{2cQ_0}>0.\n\\]\nSet\n\\[\n x_i=(X_i,\\sqrt c\\,e^2z_i),\\qquad\n g_i=(G_i,\\sqrt c\\,e^2b_i),\\qquad\n f_i=F_i+ce^4a_i.\n\\]\nThe update still satisfies $x_{i+1}=x_i-tg_i$, and every interpolation\nresidual is\n$I_{ij}=e^2J_{ij}+e^4(K_{ij}+cP_{ij})>0$.\nFor $\\ell=0,1$ the second differences satisfy\n\\[\n f_\\ell-2f_{\\ell+1}+f_{\\ell+2}\n =e^2k(t)+ce^4Q_\\ell\n \\leq\\frac12e^2k(t)=:-\\delta<0.\n\\]\nThe two strict inequalities therefore survive the same positive-slack\nrepair simultaneously.\n\nLet\n\\[\n a_0=\\min_{i\\ne j}I_{ij}>0,\\qquad\n b_0=\\max_{i\\ne j}\\|(x_i-x_j)-(g_i-g_j)\\|^2,\\qquad\n \\mu_0=\\frac{a_0}{a_0+b_0+1}.\n\\]\nThe strong-interpolation residual requirement has right-hand side at most\n$a_0b_0/[2(b_0+1)]<a_0/2$. Theorem~\\ref{thm:strong-interpolation} consequently\nconstructs a globally $1$-smooth, $\\mu_0$-strongly convex function\n$F:\\mathbb R^3\\to\\mathbb R$ realizing the full transcript.\n\nFinally put\n\\[\n M_0=1+\\max_{0\\leq i\\leq3}\\|g_i\\|^2,\\qquad\n C_0=16M_0t+6+27t^2,\\qquad\n \\sigma_0=\\min\\{1,\\delta/(2C_0)\\},\n\\]\nand smooth with a centered three-dimensional Gaussian of covariance\n$\\sigma_0^2I_3$. Denote the analytic function by $F_\\sigma$ and its\nstep-$t$ orbit from $x_0$ by $y_i$. Gaussian smoothing preserves both\nglobal curvature bounds. Nonexpansiveness and the uniform gradient\ndiscrepancy $\\sqrt3\\,\\sigma_0$ give\n\\[\n \\|y_i-x_i\\|\\leq ti\\sqrt3\\,\\sigma_0,\\qquad\n |F_\\sigma(y_i)-f_i|\n \\leq M_0ti\\sqrt3\\,\\sigma_0\n +\\frac32(t^2i^2+1)\\sigma_0^2.\n\\]\nThe coefficient weights $(1,-2,1)$ of the first second difference give\nan error at most\n$4M_0t\\sqrt3\\,\\sigma_0+(6+9t^2)\\sigma_0^2$.\nThose of the second give\n$8M_0t\\sqrt3\\,\\sigma_0+(6+27t^2)\\sigma_0^2$.\nBoth are at most $C_0\\sigma_0\\leq\\delta/2$. Hence the analytic orbit\nsatisfies\n\\[\n \\kappa_0(F_\\sigma)\\leq-\\delta/2<0,\\qquad\n \\kappa_1(F_\\sigma)\\leq-\\delta/2<0.\n\\]\nIt is globally convergent because $0<t<2$ and the analytic objective is\nstrongly convex and $1$-smooth.\n\nFor any requested step $h>h_*$, choose\n$t=(h_*+\\min\\{h,233/125\\})/2$ and replace $F_\\sigma$ by\n$(t/h)F_\\sigma$. The step-$h$ orbit is the same, both second differences\nare multiplied by the positive factor $t/h$, and the function remains\nglobally $1$-smooth, analytic, and strongly convex. This proves the\nnegative direction for every $h>h_*$.\n\n\n\\begin{proof}[Completion of Theorem~\\ref{thm:sharp-adjacency}]\nProposition~\\ref{prop:adj-positive} proves the universal prohibition. The construction just given supplies both negative second differences for every $h>h_*$ within a real-analytic strongly convex objective in dimension three, with global smoothness bounded by one. To make this bound optimal, append the coordinate $z^2/2$ initialized at $z=0$. This preserves both curvatures and makes the best global Lipschitz-gradient constant exactly one, in dimension four.\n\\end{proof}\n\n\\section{A convex filter valid up to the exact adjacency boundary}\n\nThe new weighting gives a filtered curve with a correspondingly larger universal step interval. Set\n\\begin{equation}\n V_k=\\frac{r_*f_k+f_{k+1}}{r_*+1},\\qquad\n E_k=V_k-V_{k+1}=\\frac{r_*D_k+D_{k+1}}{r_*+1}.\n \\label{eq:adj-filter}\n\\end{equation}\n\n\\begin{corollary}[Filtered convexity at the exact boundary]\n\\label{cor:adj-filter}\nFor $0<h\\leq h_*$, the sequence $V_k$ is decreasing and discretely convex. For every $N\\geq1$,\n\\begin{align}\n 0\\leq V_k-V_{k+N}&\\leq NE_k,\\label{eq:adj-filter-horizon}\\\\\n \\sum_{j=0}^{N-1}D_{k+j}\n &\\leq NE_k+\\frac{D_k-D_{k+N}}{r_*+1}\n \\leq NE_k+\\frac{D_k}{r_*+1}.\n \\label{eq:adj-loss-horizon}\n\\end{align}\nA negative curvature has positive neighbors, with\n$\\kappa_{k+1}\\geq-r_*\\kappa_k$ and, when $k\\geq1$,\n$\\kappa_{k-1}\\geq-\\kappa_k/r_*$.\n\\end{corollary}\n\n\\begin{proof}\nThe decrements are nonnegative because $h_*<2$. Equation~\\eqref{eq:adj-weighted-law} gives\n$E_k-E_{k+1}=(r_*\\kappa_k+\\kappa_{k+1})/(r_*+1)\\geq0$.\nSummation proves \\eqref{eq:adj-filter-horizon}. The exact telescoping identity\n\\[\n \\sum_{j=0}^{N-1}E_{k+j}\n =\\sum_{j=0}^{N-1}D_{k+j}\n +\\frac{D_{k+N}-D_k}{r_*+1}\n\\]\nproves \\eqref{eq:adj-loss-horizon}. The neighbor bounds follow by applying the weighted law at $k$ and $k-1$.\n\\end{proof}\n\n\\section{Exact checks and the resulting distinctions}\n\nA rational audit need not represent the algebraic endpoint in floating point. The polynomial factorization, interval root isolation, and weighted matrix identities admit exact coefficient checks. For a concrete strictly negative pair, take\n\\[\n t=\\frac{233}{125},\\qquad A=\\frac{13517}{10000},\\qquad\n B=\\frac{313}{500}.\n\\]\nThese are rational interior parameters for the same general jet construction. Their two leading second differences are\n\\[\n k_0=-\\frac{6312087}{25000000000}<0,\\qquad\n k_1=-\\frac{483}{2500000}<0.\n\\]\nThey need not be equal in this independent audit example. The explicit choices $c=10^8$ and $e=1/(2\\cdot10^6)$ give a completely rational three-dimensional transcript with both strict inequalities and all twelve strict interpolation residuals. The companion verifier supplies and checks such a transcript, including its strong-convexity allowance. The Gaussian realization is justified by the explicit analytic perturbation bound, rather than numerical integration.\n\nThe three exact boundaries now have different meanings. At $7/4$, universal convexity of the original loss curve ends. At $(19-\\sqrt{17})/8$, the particular two-step average law $D_{k+1}+D_{k+2}\\leq2D_k$ ends. At $h_*$, the prohibition of two adjacent negative second differences ends. These are three distinct thresholds; replacing one by another changes the theorem.\n\nThe adjacent-concavity theorem does not give the threshold for every longer run length, an optimal condition-number requirement for long runs, or a complete characterization of all realizable curvature sequences. Those questions remain separate. The result here completely identifies the first transition from isolated to potentially consecutive concavity, with both directions proved in explicit function classes.\n"}
{"format": "latex", "id": "04_witness", "sha256": "05591cd176ea890fee10127c73322e2274dca3e17279e6e032ff46330c9b3cc5", "source_path": "research/chapters/04_witness.tex", "text": "\\chapter{An explicit family of long concave runs}\n\\label{ch:witness}\n\n\\section{A quantitative existence theorem}\n\nThe consecutive-concavity question concerns a trajectory, not the curvature of the objective at a single point. It is therefore necessary to specify one globally defined objective, one step size, one initial point, and the complete orbit on the requested time interval. This chapter provides such a construction. The objective is strongly convex, its gradient is globally Lipschitz, and the same orbit converges after the specified interval ends. The construction resolves the finite-run existence question posed in the discussion of optimization curves in \\cite{barzilai2025}; the present theorem also supplies conditioning, curvature, and convergence bounds.\n\nFor an integer $m\\geq1$, put\n\\begin{equation}\n M=m+1,\\qquad a=\\frac1{8M},\\qquad\n \\mu=\\frac a{16}=\\frac1{128M},\\qquad h=2-\\mu.\n \\label{eq:w-parameters}\n\\end{equation}\nThe distinction between $m$ and $M$ is useful: $m$ negative second differences require $M+1=m+2$ loss observations and $M$ positive decrements. Define the closed convex cone\n\\begin{equation}\n C_a=\\{(u,v)\\in\\R^2:u\\geq0,\\ |v|\\leq au\\},\n \\qquad\n H_a(x)=\\frac{1-\\mu}{2}\\dist(x,C_a)^2+\\frac\\mu2\\norm{x}^2.\n \\label{eq:w-objective}\n\\end{equation}\nEvery parameter in this definition is rational. In particular, evaluating the function, its gradient, and its iterates requires only rational arithmetic when the input coordinates are rational.\n\n\\begin{theorem}[Constructive long runs with explicit conditioning]\n\\label{thm:w-main}\nThe function $H_a$ is $C^1$, globally $1$-smooth, and $\\mu$-strongly convex. Its optimal global smoothness and strong-convexity constants are $1$ and $\\mu$, respectively. Starting from $x_0=(0,1)$, exact gradient descent with constant step $h$ satisfies\n\\begin{align}\n \\norm{x_k}&\\leq(1-\\mu)^k &&(k\\geq0),\\label{eq:w-global-decay}\\\\\n H_a(x_k)-2H_a(x_{k+1})+H_a(x_{k+2})\n &\\leq-\\frac{885}{512}a^2 &&(0\\leq k<m).\n \\label{eq:w-rational-margin}\n\\end{align}\nConsequently its decrements obey $0<D_0<D_1<\\cdots<D_m$. Moreover,\n\\begin{equation}\n D_0\\leq\\mu+4a^2\\leq5\\mu,\n \\qquad\n H_a(x_0)-H_a(x_M)\\geq\\frac{885m}{65536M}.\n \\label{eq:w-first-total}\n\\end{equation}\nThe condition number is exactly $128M$. The entire orbit converges to the unique minimizer $0$, including after the alternating projection pattern used in the finite-run proof ends.\n\\end{theorem}\n\nThe theorem is not obtained by concatenating unrelated finite orbits. The same globally defined function generates all the observations. The strong-convexity parameter and step depend on the prescribed run length. The assertion is therefore an arbitrarily long \\emph{finite} run, not an infinite tail of increasing positive decrements. An infinite such tail would contradict lower boundedness of the loss.\n\nThe proof has four parts. Projection geometry gives global regularity and convergence. An angular comparison certifies the active projection face. Elementary inequalities establish a uniform negative second difference. Finally, a quadratic lift gives an exact recurrence for all loss values while the certified face pattern persists.\n\n\\section{Projection facts, proved from the defining minimization}\n\nWe first record the projection facts needed below so that no regularity theorem for squared distances is left implicit.\n\n\\begin{lemma}[Projection and squared distance]\n\\label{lem:w-projection}\nLet $C\\subset\\R^d$ be nonempty, closed, and convex. Each $x\\in\\R^d$ has a unique nearest point $P_Cx$. Write $r_C=I-P_C$. Then\n\\begin{align}\n \\ip{x-P_Cx}{c-P_Cx}&\\leq0\\qquad(c\\in C),\\label{eq:w-proj-vary}\\\\\n \\norm{r_Cx-r_Cy}^2&\\leq\\ip{r_Cx-r_Cy}{x-y}.\n \\label{eq:w-res-firm}\n\\end{align}\nThe function $F_C(x)=\\tfrac12\\dist(x,C)^2$ is convex and continuously differentiable, with gradient $r_C$. Its gradient has Lipschitz constant at most $1$.\n\\end{lemma}\n\n\\begin{proof}\nFix $c_0\\in C$. A minimizing sequence for $\\norm{x-c}$ may be restricted to a closed ball about $x$ whose radius is greater than $\\norm{x-c_0}$. The intersection of that ball with $C$ is compact, so the minimum is attained. If $p\\neq q$ were two minimizers, their midpoint would lie in $C$ and the identity\n\\[\n \\norm{x-\\frac{p+q}{2}}^2\n =\\frac12\\norm{x-p}^2+\\frac12\\norm{x-q}^2-\\frac14\\norm{p-q}^2\n\\]\nwould give a strictly smaller distance. Thus the minimizer is unique.\n\nFor $p=P_Cx$ and any $c\\in C$, the segment $p+t(c-p)$ is feasible for $0\\leq t\\leq1$. The right derivative at $t=0$ of its squared distance to $x$ must be nonnegative. This is precisely \\eqref{eq:w-proj-vary}. Conversely, expanding $\\norm{x-c}^2$ shows that this inequality characterizes a nearest point.\n\nApply \\eqref{eq:w-proj-vary} to $x$ with $c=P_Cy$, and to $y$ with $c=P_Cx$. Addition gives\n\\[\n \\ip{P_Cx-P_Cy}{x-y}\\geq\\norm{P_Cx-P_Cy}^2.\n\\]\nExpanding $r_Cx-r_Cy=(x-y)-(P_Cx-P_Cy)$ yields \\eqref{eq:w-res-firm}. Cauchy--Schwarz then gives $\\norm{r_Cx-r_Cy}\\leq\\norm{x-y}$.\n\nFor convexity, a candidate nearest point at $tx+(1-t)y$ is $tP_Cx+(1-t)P_Cy$. Convexity of the squared norm gives\n\\[\n F_C(tx+(1-t)y)\\leq tF_C(x)+(1-t)F_C(y).\n\\]\nFor differentiability, using $P_Cx$ as a candidate at $x+e$ gives\n\\[\n F_C(x+e)\\leq F_C(x)+\\ip{r_Cx}{e}+\\frac12\\norm{e}^2.\n\\]\nApplying this inequality in the reverse direction and using the Lipschitz property of $r_C$ bounds the same first-order remainder below by $-\\tfrac32\\norm{e}^2$. The two estimates show that the derivative is $r_Cx$. Its continuity and Lipschitz bound have already been proved.\n\\end{proof}\n\nApplying the lemma to \\eqref{eq:w-objective} proves convexity and smoothness of $H_a$. Strong convexity follows by subtracting $\\mu\\norm{x}^2/2$: the remainder is convex. These bounds are globally optimal. In the interior of $C_a$, the objective is exactly $\\mu\\norm{x}^2/2$. In the interior of the polar cone, $P_{C_a}x=0$ and the objective is exactly $\\norm{x}^2/2$. Both interiors are nonempty. The Hessian therefore equals $\\mu I$ and $I$ on nonempty open sets, so neither endpoint constant can be improved.\n\nThe objective is nonnegative and vanishes at $0$. Strong convexity makes this minimizer unique. This is useful later: the trajectory construction cannot fail because the objective lacks a minimum or because the iterates escape toward an unattained infimum.\n\n\\section{A global contraction independent of face switching}\n\nSet\n\\begin{equation}\n q=1-\\mu,\\qquad p=q^2.\n \\label{eq:w-pq}\n\\end{equation}\nThe chosen step has the special identities\n\\begin{equation}\n 1-h\\mu=p,\\qquad 1-h=-q.\n \\label{eq:w-balance}\n\\end{equation}\nIf $P=P_{C_a}$ and $r=I-P$, the gradient is $\\mu P x+r x$, and hence the update is\n\\begin{equation}\n T(x)=x-h\\nabla H_a(x)=pPx-qr x.\n \\label{eq:w-map}\n\\end{equation}\nThe projection and residual of a closed convex cone are orthogonal. Indeed, both $0$ and $2Px$ belong to the cone, so substituting them in \\eqref{eq:w-proj-vary} gives opposite inequalities for $\\ip{x-Px}{Px}$. Thus\n\\begin{equation}\n \\norm{T(x)}^2=p^2\\norm{Px}^2+q^2\\norm{r x}^2\n \\leq q^2\\norm{x}^2.\n \\label{eq:w-radial-contract}\n\\end{equation}\nIteration proves \\eqref{eq:w-global-decay} for every $k\\geq0$.\n\nThis argument deliberately avoids assuming that the orbit remains on a particular pair of projection faces. The finite-run analysis will use an alternating pattern, but the convergence proof remains valid when that pattern changes. It also gives the uniform bound $\\norm{x_k}\\leq1$, which controls every coordinate appearing in the curvature estimates.\n\n\\begin{remark}[Radial decrease and increasing improvements]\nEquation \\eqref{eq:w-radial-contract} does not force $D_k$ to decrease. The objective weighs the tangent and normal coordinates differently, and the active normal rotates when the projection face changes. The norm contracts while the relative allocation between these two coordinates changes. This is the precise mechanism that permits increasing loss decrements under strong convexity.\n\\end{remark}\n\n\\section{Exact coordinates on alternating projection faces}\n\nLet $\\beta=\\arctan a$, and abbreviate\n\\begin{equation}\n c=\\cos(2\\beta)=\\frac{1-a^2}{1+a^2},\n \\qquad s=\\sin(2\\beta)=\\frac{2a}{1+a^2}.\n \\label{eq:w-cs}\n\\end{equation}\nFor the face of sign $(-1)^k$, use the orthonormal tangent and outward normal\n\\begin{equation}\n e_k=(\\cos\\beta,(-1)^k\\sin\\beta),\\qquad\n n_k=(-\\sin\\beta,(-1)^k\\cos\\beta).\n \\label{eq:w-frame}\n\\end{equation}\nWhen $x_k=t_ke_k+d_kn_k$ with $t_k>0$ and $d_k>0$, its nearest point is $t_ke_k$. In these coordinates the objective and update are\n\\begin{align}\n H_a(x_k)&=\\frac12(d_k^2+\\mu t_k^2),\\label{eq:w-loss-coordinate}\\\\\n t_{k+1}&=pc\\,t_k+qs\\,d_k,\\qquad\n d_{k+1}=-ps\\,t_k+qc\\,d_k.\n \\label{eq:w-linear}\n\\end{align}\nThe second line expresses the updated point in the frame of the \\emph{opposite} face. It becomes an actual orbit identity only after positivity of the new coordinates has been established.\n\nThe initial coordinates are $t_0=\\sin\\beta$ and $d_0=\\cos\\beta$, so $t_0/d_0=a$. The following comparison certifies all the faces needed for $m$ second differences.\n\n\\begin{lemma}[Angular comparison and branch certificate]\n\\label{lem:w-angle}\nFor $0\\leq k\\leq M$, the actual orbit has the coordinates in \\eqref{eq:w-linear}, and\n\\begin{align}\n 0<r_k:=\\frac{t_k}{d_k}\n &\\leq\\tan((2k+1)\\beta)<\\frac13,\\label{eq:w-ratio-bound}\\\\\n d_k&\\geq q^k\\cos((2k+1)\\beta),\\qquad d_k^2\\geq\\frac78,\\label{eq:w-normal-bound}\\\\\n t_k&\\geq t_0\\geq\\frac a2,\\qquad \\norm{x_k}\\leq1.\n \\label{eq:w-distance-bound}\n\\end{align}\nIn particular every queried point is at distance at least $a/2$ from the boundary of its quadratic projection chamber.\n\\end{lemma}\n\n\\begin{proof}\nSuppose the representation holds at step $k$, with $0\\leq k<M$. Dividing \\eqref{eq:w-linear} by its normal coordinate gives\n\\begin{equation}\n r_{k+1}=\\frac{qc r_k+s}{c-qs r_k}\n =\\tan\\bigl(\\arctan(qr_k)+2\\beta\\bigr),\n \\label{eq:w-projective}\n\\end{equation}\nprovided the denominator is positive. If $r_k\\leq\\tan((2k+1)\\beta)$, then\n\\[\n c-qs r_k\\geq c-s\\tan((2k+1)\\beta)\n =\\frac{\\cos((2k+3)\\beta)}{\\cos((2k+1)\\beta)}>0.\n\\]\nHere the last inequality follows from\n\\[\n (2M+1)\\beta\\leq(2M+1)a\\leq\\frac5{16}<\\frac\\pi2,\n\\]\nusing $M\\geq2$. The numerator in \\eqref{eq:w-projective} is positive. Thus the next point really projects onto the opposite face. Since $0<q<1$ and tangent is increasing on this angular interval,\n\\[\n r_{k+1}\\leq\\tan(\\arctan(r_k)+2\\beta)\n \\leq\\tan((2k+3)\\beta).\n\\]\nInduction from $r_0=\\tan\\beta$ establishes the representation and angular upper bound.\n\nThe elementary inequalities $\\sin u\\leq u$ and $\\cos u\\geq1-u^2/2$ give\n\\[\n \\tan(5/16)\\leq\\frac{5/16}{1-25/512}\n =\\frac{160}{487}<\\frac13.\n\\]\nThe normal coordinate satisfies\n\\[\n d_{k+1}=q(c-qs r_k)d_k\n \\geq q\\frac{\\cos((2k+3)\\beta)}{\\cos((2k+1)\\beta)}d_k.\n\\]\nThe product telescopes, yielding $d_k\\geq q^k\\cos((2k+1)\\beta)$. Bernoulli's inequality and $M\\mu=1/128$ imply $q^k\\geq127/128$. Therefore\n\\begin{equation}\n d_k\\geq\\frac{127}{128}\\frac{487}{512}\n =\\frac{61849}{65536},\\qquad\n \\left(\\frac{61849}{65536}\\right)^2>\\frac78.\n \\label{eq:w-normal-rational}\n\\end{equation}\n\nIt remains to show that the tangent stays away from zero. The parameter range gives\n\\[\n 1-pc\\leq2\\mu+2a^2\\leq\\frac a4,\n \\qquad qs=\\frac{2qa}{1+a^2}\\geq a.\n\\]\nConsequently\n\\[\n t_{k+1}-t_k=d_k\\bigl(qs-(1-pc)r_k\\bigr)\n \\geq\\frac{11}{12}a d_k>0.\n\\]\nThis proves $t_k\\geq\\sin\\beta=a/\\sqrt{1+a^2}\\geq a/2$. The norm bound comes from \\eqref{eq:w-radial-contract}. In the orthonormal face coordinates, the chamber is the quadrant $t>0,d>0$, whose boundary distance is $\\min\\{t,d\\}$. The preceding estimates give the last assertion.\n\\end{proof}\n\nThe angular comparison is useful beyond this particular choice of $a$. It shows how much of the apparent rotation can accumulate before a face change is no longer certified. A constant step size is retained throughout: the proof does not slow the algorithm down as the end of the requested interval approaches.\n\n\\section{A complete uniform curvature calculation}\n\nFix $0\\leq k<m$. Write $t=t_k$, $d=d_k$, and $r=t/d$. Define\n\\begin{equation}\n u=\\frac{d_{k+1}}{d_k}=qc-psr,\\qquad A=1-u,\n \\qquad v=\\frac{d_{k+2}}{d_k}.\n \\label{eq:w-uAv}\n\\end{equation}\nThe normal coordinate follows the scalar recurrence\n\\[\n d_{k+2}=(p+q)c\\,d_{k+1}-pq\\,d_k.\n\\]\nBecause $p=q^2$ and $q=1-\\mu$, this can be written exactly as\n\\begin{align}\n v&=1-2A-b,\\label{eq:w-v-b}\\\\\n b&=(p+q)(1-c)u+2\\mu^2-\\mu^3-(3\\mu-\\mu^2)A.\n \\label{eq:w-b}\n\\end{align}\nNo approximation occurs in these identities.\n\n\\begin{lemma}[Scalar bounds for the normal component]\n\\label{lem:w-normal-curvature}\nThroughout the certified interval,\n\\begin{equation}\n 0\\leq A\\leq\\frac{41}{48}a,\n \\quad u\\geq\\frac{15}{16},\n \\quad 1-2A\\geq\\frac78,\n \\quad\\frac{27}{8}a^2\\leq b\\leq\\frac92a^2.\n \\label{eq:w-scalars}\n\\end{equation}\nIn particular,\n\\begin{equation}\n 1-2u^2+v^2\\leq-4a^2.\n \\label{eq:w-normal-second}\n\\end{equation}\n\\end{lemma}\n\n\\begin{proof}\nSince $a\\leq1/16$, $\\mu=a/16$, and $0\\leq r\\leq1/3$, we have\n\\[\n A=1-qc+psr\\leq\\mu+2a^2+2ar\n \\leq\\left(\\frac1{16}+\\frac18+\\frac23\\right)a\n =\\frac{41}{48}a.\n\\]\nAlso $A\\geq0$ because $qc\\leq1$ and $psr\\geq0$. Thus\n\\[\n u\\geq1-\\frac{41}{768}\\geq\\frac{15}{16},\n \\qquad 1-2A\\geq1-\\frac{41}{384}\\geq\\frac78.\n\\]\nThe same parameter range implies\n\\[\n p+q\\geq\\frac{63}{32},\\qquad\n 1-c=\\frac{2a^2}{1+a^2}\\geq\\frac{63}{32}a^2.\n\\]\nThe first term of \\eqref{eq:w-b} is therefore at least $59535a^2/16384$. The last, potentially negative, term has magnitude at most\n\\[\n 3\\mu A\\leq\\frac{41}{256}a^2\n =\\frac{2624}{16384}a^2.\n\\]\nThe middle term is nonnegative. Hence\n\\[\n b\\geq\\frac{56911}{16384}a^2\\geq\\frac{27}{8}a^2.\n\\]\nFor the upper bound, drop the nonpositive last term, use $u\\leq1$, and obtain\n\\[\n b\\leq4a^2+2\\mu^2\\leq\\frac92a^2.\n\\]\nFinally, \\eqref{eq:w-v-b} gives the identity\n\\[\n 1-2u^2+v^2=2A^2-2b(1-2A)+b^2.\n\\]\nSubstitution of the established bounds and $a^2\\leq1/256$ yields\n\\begin{align*}\n 1-2u^2+v^2\n &\\leq\\left(\\frac{1681}{1152}-\\frac{189}{32}+\\frac{81}{1024}\\right)a^2\\\\\n &=-\\frac{40255}{9216}a^2\\leq-4a^2.\n\\end{align*}\n\\end{proof}\n\nThe normal part $d_k^2/2$ of the objective has second difference\n\\[\n \\frac{d_k^2}{2}(1-2u^2+v^2)\n \\leq-2d_k^2a^2\\leq-\\frac74a^2.\n\\]\nThe tangent contribution must be bounded separately because it can have the opposite sign. Set\n\\begin{equation}\n E=t_{k+1}-t_k,\\qquad C=t_{k+2}-2t_{k+1}+t_k.\n \\label{eq:w-EC}\n\\end{equation}\nThe already proved inequalities give\n\\begin{align}\n |E|&\\leq\\frac a4\\frac13+2a=\\frac{25}{12}a,\n &|d_{k+1}-d_k|&\\leq\\frac{41}{48}a,\\label{eq:w-E-bound}\\\\\n |C|&\\leq\\frac a4\\frac{25}{12}a+2a\\frac{41}{48}a\n =\\frac{107}{48}a^2\\leq\\frac94a^2.\n \\label{eq:w-C-bound}\n\\end{align}\nThe expression for $C$ used here is $(pc-1)E+qs(d_{k+1}-d_k)$, obtained by subtracting two successive tangent updates. Because $|t_{k+2}|\\leq1/3$,\n\\begin{align}\n t_k^2-2t_{k+1}^2+t_{k+2}^2\n &=2E^2+2t_{k+2}C-C^2\\notag\\\\\n &\\leq\\left(2\\left(\\frac{25}{12}\\right)^2+\\frac23\\frac94\\right)a^2\n =\\frac{733}{72}a^2\\leq11a^2.\n \\label{eq:w-tangent-second}\n\\end{align}\nMultiplying by $\\mu/2\\leq1/512$ and adding the normal part proves\n\\[\n \\kappa_k(H_a)\\leq-\\frac74a^2+\\frac{11}{512}a^2\n =-\\frac{885}{512}a^2,\n\\]\nas claimed in Theorem~\\ref{thm:w-main}.\n\n\\subsection{First decrement and accumulated gain}\n\nThe gradient at $x_0$ is nonzero. The smooth descent inequality, with $0<h<2$, gives $D_0>0$. All subsequent queried decrements are positive and strictly increasing because the second differences are negative.\n\nSince $t_1>t_0$, the tangent term contributes a nonpositive amount to $D_0$. At $r_0=a$,\n\\[\n 1-\\frac{d_1}{d_0}\n =\\mu+\\frac{2a^2(q+p)}{1+a^2}\\leq\\mu+4a^2.\n\\]\nUsing $0<d_1/d_0\\leq1$ and $d_0^2\\leq1$ gives\n\\[\n D_0\\leq\\frac{d_0^2}{2}\\left(1-\\frac{d_1^2}{d_0^2}\\right)\n \\leq\\mu+4a^2\\leq5\\mu.\n\\]\nThe curvature bound implies\n\\[\n D_k\\geq D_0+\\frac{885}{512}k a^2.\n\\]\nSumming for $0\\leq k<M$ gives\n\\[\n H_a(x_0)-H_a(x_M)\n \\geq M D_0+\\frac{885}{1024}a^2mM\n \\geq\\frac{885m}{65536M}.\n\\]\nThis completes the proof of Theorem~\\ref{thm:w-main}.\n\n\\section{An exact rational representation}\n\nAlthough the geometric proof uses trigonometric coordinates, the objective and every iterate are rational. For $z=(u,v)$, the projection is\n\\begin{equation}\n P_{C_a}(u,v)=\n \\begin{cases}\n (u,v),&u\\geq0,\\ |v|\\leq au,\\\\\n (0,0),&u+a|v|\\leq0,\\\\\n (t,\\operatorname{sgn}(v)at),&\\text{otherwise},\n \\end{cases}\n \\qquad t=\\frac{u+a|v|}{1+a^2}.\n \\label{eq:w-explicit-proj}\n\\end{equation}\nThe cases meet continuously. They may be verified directly using \\eqref{eq:w-proj-vary}; an independent evaluation is obtained by comparing the origin, the interior candidate, and the orthogonal projections onto the two boundary rays.\n\nRemove the common factor $\\sqrt{1+a^2}$ from the frame coordinates by defining\n\\[\n T_k=\\sqrt{1+a^2}\\,t_k,\\qquad N_k=\\sqrt{1+a^2}\\,d_k.\n\\]\nThen $T_0=a$, $N_0=1$, the update is the rational matrix in \\eqref{eq:w-linear}, and\n\\begin{equation}\n H_a(x_k)=\\frac{N_k^2+\\mu T_k^2}{2(1+a^2)}.\n \\label{eq:w-rational-loss}\n\\end{equation}\nThe ambient coordinates are recovered by\n\\begin{equation}\n u_k=\\frac{T_k-aN_k}{1+a^2},\\qquad\n v_k=(-1)^k\\frac{aT_k+N_k}{1+a^2}.\n \\label{eq:w-ambient-recover}\n\\end{equation}\nThese formulas allow exact verification without evaluating a sine, cosine, square root, or logarithm.\n\n\\begin{proposition}[Pointwise quadratic certificate]\n\\label{prop:w-curvature-matrix}\nLet\n\\begin{equation}\n B=\\begin{pmatrix}pc&qs\\\\-ps&qc\\end{pmatrix},\\qquad\n Q=\\diag(\\mu,1),\\qquad\n K=Q-2B^\\top QB+(B^2)^\\top QB^2.\n \\label{eq:w-BQK}\n\\end{equation}\nOn the certified branch,\n\\begin{equation}\n \\kappa_k(H_a)=\\frac12\n \\begin{pmatrix}t_k&d_k\\end{pmatrix}\n K\\begin{pmatrix}t_k\\\\d_k\\end{pmatrix}.\n \\label{eq:w-pointwise-certificate}\n\\end{equation}\nWhen $a,\\mu,h$ and the initial coordinates are rational after a common normalization, this equality gives an exact rational decision procedure for every finite list of curvature signs, provided each projection branch is checked.\n\\end{proposition}\n\n\\begin{proof}\nThe loss is the quadratic form $v^\\top Qv/2$ in $v=(t,d)^\\top$. Substitution of $Bv$ and $B^2v$ gives the displayed identity. The branch check is necessary because $B$ is the update in alternating face coordinates, not a global linear representation of the whole objective.\n\\end{proof}\n\n\\section{Losses satisfy an exact third-order recurrence}\n\nThe finite trajectory is compressible in a stronger sense than rational evaluability. Define\n\\begin{equation}\n \\tau=(p+q)c,\\qquad \\Delta=pq=q^3,\n \\qquad A_B=\\tau^2-\\Delta.\n \\label{eq:w-spectral-invariants}\n\\end{equation}\nThese are the trace, determinant, and a quadratic invariant of $B$.\n\n\\begin{theorem}[Exact recurrence for the loss sequence]\n\\label{thm:w-loss-recurrence}\nWhenever the four queried states lie in the alternating branch,\n\\begin{equation}\n f_{k+3}=A_B f_{k+2}-\\Delta A_B f_{k+1}+\\Delta^3f_k,\n \\qquad f_k=H_a(x_k).\n \\label{eq:w-third-order}\n\\end{equation}\nThe decrements and second differences satisfy the same linear recurrence on every index range where the necessary observations are certified.\n\\end{theorem}\n\n\\begin{proof}\nThe update $v\\mapsto Bv$ acts linearly on the three quadratic coordinates $(t^2,td,d^2)$. If $B$ has eigenvalues $\\lambda_1,\\lambda_2$, this lifted action has eigenvalues $\\lambda_1^2$, $\\lambda_1\\lambda_2$, and $\\lambda_2^2$. Their elementary symmetric polynomials are\n\\[\n \\tau^2-\\Delta,\\qquad \\Delta(\\tau^2-\\Delta),\\qquad \\Delta^3.\n\\]\nThe Cayley--Hamilton identity of the lifted $3\\times3$ matrix gives \\eqref{eq:w-third-order} for every linear functional of the quadratic coordinates, including the loss. The matrix identity remains valid when eigenvalues coincide: all its entries are polynomials in the entries of $B$, and matrices with distinct eigenvalues are dense. Equivalently, one may expand the lifted matrix directly and verify the identity without eigenvalues. A fixed linear difference operator commutes with a constant-coefficient recurrence, proving the final assertion.\n\\end{proof}\n\nAn observed sequence satisfying \\eqref{eq:w-third-order} need not come from this objective. The theorem is a forward implication from a known mathematical model. A few matching values do not establish the model, its branch inequalities, or its validity on subsequent steps. This distinction matters for a learning controller: a certified model permits exact extrapolation within a certified interval, whereas an unconstrained fit to a few losses does not.\n\n\\section{The spectral shape of the reflected family}\n\nThe recurrence can be made more explicit. This section gives a general statement for any matrix\n\\begin{equation}\n B=\\begin{pmatrix}pc&qs\\\\-ps&qc\\end{pmatrix},\n \\quad p,q\\in(0,1),\\quad c=\\cos(2\\beta)>0,\\quad s=\\sin(2\\beta)>0,\n \\label{eq:w-general-B}\n\\end{equation}\nas long as the alternating projection branch is valid. For gradient descent on the same cone family, these parameters correspond to $p=1-h\\mu$ and $q=h-1$ when $1<h<2$ and $h\\mu<1$. Thus the analysis is not limited to the balance identity $p=q^2$.\n\n\\begin{proposition}[Oscillatory loss formula]\n\\label{prop:w-oscillatory}\nSuppose $\\tau^2<4\\Delta$, where $\\tau=(p+q)c$ and $\\Delta=pq$. Define\n\\[\n \\rho=\\sqrt\\Delta,\\qquad\n \\theta=\\arccos\\frac{\\tau}{2\\rho}\\in(0,\\pi/2),\\qquad\n J=\\frac{B/\\rho-\\cos\\theta\\,I}{\\sin\\theta}.\n\\]\nThen $J^2=-I$ and\n\\begin{equation}\n B^k=\\rho^k\\bigl(\\cos(k\\theta)I+\\sin(k\\theta)J\\bigr).\n \\label{eq:w-B-power}\n\\end{equation}\nFor any positive definite $Q$ and nonzero initial vector $v$, put\n\\[\n \\alpha=v^\\top Qv,\\qquad\n \\gamma=(Jv)^\\top Q(Jv),\\qquad\n \\xi=v^\\top QJv,\n\\]\nand $A=(\\alpha+\\gamma)/4$, $C_1=(\\alpha-\\gamma)/4$, $C_2=\\xi/2$. The quadratic loss is exactly\n\\begin{equation}\n f_k=\\Delta^k\\bigl(A+C_1\\cos(2k\\theta)+C_2\\sin(2k\\theta)\\bigr).\n \\label{eq:w-oscillatory-loss}\n\\end{equation}\n\\end{proposition}\n\n\\begin{proof}\nThe Cayley--Hamilton identity $B^2-\\tau B+\\Delta I=0$ gives $J^2=-I$ after substitution. The multiplication law for $\\cos u\\,I+\\sin u\\,J$ is the ordinary angle-addition law because $J^2=-I$, proving \\eqref{eq:w-B-power} by induction. Inserting this power into $v^\\top(B^k)^\\top QB^kv/2$ and expanding the double-angle identities gives \\eqref{eq:w-oscillatory-loss}.\n\\end{proof}\n\nThe eigenvalues are complex precisely when\n\\begin{equation}\n \\sin^2(2\\beta)>\\left(\\frac{p-q}{p+q}\\right)^2.\n \\label{eq:w-complex-condition}\n\\end{equation}\nIndeed, $\\tau^2<4pq$ is equivalent to $(p+q)^2(1-s^2)<4pq$. In the constructed family this condition is readily satisfied: $|p-q|/(p+q)=\\mu/(2-\\mu)$, whereas $s=2a/(1+a^2)$ and $\\mu=a/16$.\n\n\\begin{corollary}[A family-specific ceiling on consecutive concavity]\n\\label{cor:w-phase-ceiling}\nUnder Proposition~\\ref{prop:w-oscillatory}, let $U=C_1-iC_2$ and write\n\\[\n U(1-\\Delta e^{2i\\theta})^2=R e^{i\\phi},\\qquad R\\geq0.\n\\]\nThen\n\\begin{equation}\n \\kappa_k=\\Delta^k\\bigl(A(1-\\Delta)^2+R\\cos(2k\\theta+\\phi)\\bigr).\n \\label{eq:w-curvature-phase}\n\\end{equation}\nIf $R\\leq A(1-\\Delta)^2$, no second difference is negative. Otherwise every consecutive run of negative second differences within the branch has length at most\n\\begin{equation}\n \\left\\lceil\n \\frac{\\arccos\\bigl(A(1-\\Delta)^2/R\\bigr)}{\\theta}\n \\right\\rceil\n \\leq\\left\\lceil\\frac\\pi{2\\theta}\\right\\rceil.\n \\label{eq:w-phase-ceiling}\n\\end{equation}\n\\end{corollary}\n\n\\begin{proof}\nApply $f_k-2f_{k+1}+f_{k+2}$ to the constant and complex oscillatory terms of \\eqref{eq:w-oscillatory-loss}. This gives \\eqref{eq:w-curvature-phase}. If $R$ is no larger than the nonnegative offset, the bracket cannot be negative. Otherwise the negative phases form an open interval modulo $2\\pi$ of width\n\\[\n w=2\\arccos\\bigl(A(1-\\Delta)^2/R\\bigr)\\leq\\pi.\n\\]\nSuccessive phases differ by $2\\theta<\\pi$. A sequence of consecutive negative phases cannot jump across the complementary interval, whose width is at least $\\pi$. Hence a run of length $\\ell$ satisfies $(\\ell-1)2\\theta<w$. For an integer $\\ell$, this is equivalent to $\\ell\\leq\\lceil w/(2\\theta)\\rceil$, which is the stated bound.\n\\end{proof}\n\nThis ceiling is a result about the reflected-cone family. It is not a universal theorem for all strongly convex objectives. Its role is to describe the mechanism exactly and to show why the finite interval eventually ends. The global convergence proof did not depend on this spectral calculation.\n\n\\section{General ceilings from strong convexity}\n\nThere are also bounds that do not assume a projection model. They help distinguish arbitrary finite run length across a family from arbitrary length at fixed conditioning and fixed step.\n\n\\begin{proposition}[Two universal finite-run ceilings]\n\\label{prop:w-universal-ceiling}\nLet $f$ be $\\mu$-strongly convex and $L$-smooth, with $0<\\mu\\leq L$, and let $0<\\eta<2/L$. Define\n\\[\n \\gamma=\\mu\\eta(2-\\eta L),\\qquad\n q_\\eta=\\max\\{|1-\\eta\\mu|,|1-\\eta L|\\}<1.\n\\]\nSuppose $D_0>0$ and $D_0\\leq D_1\\leq\\cdots\\leq D_m$. Then\n\\begin{equation}\n m+1\\leq\\frac1\\gamma.\n \\label{eq:w-budget-ceiling}\n\\end{equation}\nIf $q_\\eta>0$, also\n\\begin{equation}\n m\\leq\n \\frac{\\log\\bigl(L/(\\mu\\gamma)\\bigr)}{-2\\log q_\\eta}.\n \\label{eq:w-contract-ceiling}\n\\end{equation}\nIf $q_\\eta=0$, the first update reaches the minimizer and no strictly increasing positive-decrement run of length $m\\geq1$ is possible.\n\\end{proposition}\n\n\\begin{proof}\nThe descent inequality and strong convexity give\n\\[\n D_0\\geq\\eta\\left(1-\\frac{\\eta L}{2}\\right)\\norm{\\nabla f(x_0)}^2\n \\geq\\gamma\\bigl(f(x_0)-f_*\\bigr).\n\\]\nFor completeness, the gradient lower bound follows by minimizing the strong-convexity lower quadratic\n\\[\n f(y)\\geq f(x_0)+\\ip{\\nabla f(x_0)}{y-x_0}+\\frac\\mu2\\norm{y-x_0}^2\n\\]\nover $y$: it yields $f(x_0)-f_*\\leq\\norm{\\nabla f(x_0)}^2/(2\\mu)$. Summing the nondecreasing decrements gives\n\\[\n (m+1)D_0\\leq\\sum_{k=0}^mD_k\n =f(x_0)-f(x_{m+1})\\leq f(x_0)-f_*.\n\\]\nCombining the two inequalities proves \\eqref{eq:w-budget-ceiling}.\n\nThe gradient-descent map is $q_\\eta$-Lipschitz for a smooth strongly convex function. One proof mollifies the function, averages its Hessian along a segment, and notes that the eigenvalues of $I-\\eta A$ lie between $1-\\eta L$ and $1-\\eta\\mu$; passage to the limit preserves the bound. Thus\n\\[\n f(x_m)-f_*\\leq\\frac L2q_\\eta^{2m}\\norm{x_0-x_*}^2\n \\leq\\frac L\\mu q_\\eta^{2m}\\bigl(f(x_0)-f_*\\bigr).\n\\]\nBecause $D_m\\leq f(x_m)-f_*$ and $D_m\\geq D_0\\geq\\gamma(f(x_0)-f_*)$, cancellation gives $q_\\eta^{2m}\\geq\\mu\\gamma/L$. Taking logarithms proves \\eqref{eq:w-contract-ceiling}. The case $q_\\eta=0$ follows directly from the contraction statement.\n\\end{proof}\n\nThese ceilings are sufficient restrictions, not claimed sharp bounds. For the present construction, $L=1$, $h=2-\\mu$, and $q_h=1-\\mu$. The logarithmic ceiling is of order $\\mu^{-1}\\log(1/\\mu)$, whereas the proved construction has run length of order $\\mu^{-1}$. This comparison leaves a logarithmic gap and does not establish an optimal global condition-number tradeoff.\n\n\\section{What the exact audit establishes}\n\nThe accompanying rational verifier computes every trajectory twice: once in ambient coordinates using \\eqref{eq:w-explicit-proj}, and once with \\eqref{eq:w-linear}. A separate face-enumeration projection implementation checks the ambient computation. The verifier then checks the projection inequalities, coordinate bounds, global contraction, scalar curvature identities, the matrix certificate, the loss recurrence, and every claimed rational parameter inequality.\n\nThese finite checks serve two purposes. They detect implementation and transcription errors, and they supply reusable exact examples for testing proposed curve-shape rules. They are not the proof of the all-$m$ theorem: that proof is the preceding symbolic and inequality argument. Nor does an exact rational trajectory verify a Gaussian integral. The analytic extension and its wider smoothing scale are established separately in Chapter~\\ref{ch:analytic}.\n\n"}
{"format": "latex", "id": "05_analytic", "sha256": "9df56e76ba2d1e51377e422736aa44e678aa2158b003b1310b17ec8db6969469", "source_path": "research/chapters/05_analytic.tex", "text": "\\chapter{Analytic realization and transformations}\n\\label{ch:analytic}\n\n\\section{The strengthened analytic theorem}\n\nThe objective in Chapter~\\ref{ch:witness} is globally differentiable but is only piecewise quadratic. Its lack of a classical Hessian on a few rays is unnecessary for the curve phenomenon. This chapter constructs an entire Gaussian realization, proves an explicit smoothing scale, and transfers the result to arbitrary dimension and fixed positive-definite metrics.\n\nKeep the parameters in \\eqref{eq:w-parameters} and define\n\\begin{equation}\n B_a=\\log\\frac{64}{a^2},\\qquad\n \\sigma=\\frac{a}{2\\sqrt{2B_a}},\\qquad\n Z\\sim\\mathcal N(0,\\sigma^2I_2),\\qquad\n F_a(x)=\\E H_a(x-Z).\n \\label{eq:a-definition}\n\\end{equation}\nHere $\\log$ is the natural logarithm. This notation $B_a$ is a scalar tail parameter and is unrelated to the trajectory matrix $B$ in \\eqref{eq:w-BQK}. The smoothing scale is of order $1/(M\\sqrt{\\log M})$. It is specified completely, rather than being chosen after observing a numerical trajectory.\n\n\\begin{theorem}[Analytic, strongly convex long runs]\n\\label{thm:a-main}\nFor every integer $m\\geq1$, the function $F_a$ in \\eqref{eq:a-definition} is real analytic and is the restriction of an entire function on $\\mathbb C^2$. Its optimal global smoothness and strong-convexity constants are exactly $1$ and $\\mu=1/[128(m+1)]$.\n\nStarting at $z_0=(0,1)$ with constant step $h=2-\\mu$, its exact gradient-descent orbit satisfies\n\\begin{equation}\n \\kappa_k(F_a)\\leq-\\frac32a^2\n =-\\frac{3}{128(m+1)^2},\\qquad 0\\leq k<m.\n \\label{eq:a-margin}\n\\end{equation}\nIts decrements satisfy $0<D_0<D_1<\\cdots<D_m$, with\n\\begin{align}\n D_0&\\leq\n \\mu+4a^2+\\frac{a^3}{8}+\\frac{a^4}{64}\n \\leq6\\mu=\\frac{3}{64(m+1)},\\label{eq:a-first}\\\\\n F_a(z_0)-F_a(z_M)&\\geq\\frac{3m}{256M}\\geq\\frac3{512}.\n \\label{eq:a-total}\n\\end{align}\nThere is a unique minimizer $z_*$, and the whole orbit obeys\n\\begin{equation}\n \\norm{z_*}\\leq\\sigma\\sqrt{\\frac2\\mu},\\qquad\n \\norm{z_k-z_*}\\leq(1-\\mu)^k\\norm{z_0-z_*}.\n \\label{eq:a-convergence}\n\\end{equation}\nThe condition number is therefore exactly $128(m+1)$.\n\\end{theorem}\n\nThe analytic theorem does not rely on numerical quadrature. Its quantitative proof compares the integral-defined orbit with an exactly evaluable orbit and bounds the discrepancy by explicit Gaussian tail moments. A smooth loss curve can therefore retain a long strict concavity interval with a negative margin that is certified independently of any plotting precision.\n\n\\section{Why a local quadratic argument matters}\n\nA direct global smoothing estimate gives\n\\[\n \\norm{\\nabla F_a(x)-\\nabla H_a(x)}\n \\leq \\E\\norm Z\\leq\\sqrt2\\,\\sigma.\n\\]\nOver $M$ steps, nonexpansiveness alone then gives a trajectory discrepancy of order $M\\sigma$. Comparing loss second differences by absolute value bounds would require $\\sigma$ to be of order $a^2/M$, which is of order $a^3$ here.\n\nThe certified orbit has additional geometry. Every queried point lies inside a quadratic chamber at distance at least $a/2$ from its boundary. The two visited quadratic forms have the same Hessian trace, $1+\\mu$. Gaussian smoothing of either quadratic adds the same constant $\\sigma^2(1+\\mu)/2$ to its value and makes no change to its gradient. This common constant disappears from every decrement and second difference. The only relevant errors come from Gaussian samples that cross a chamber boundary.\n\nThe next lemma formulates this observation independently of the cone construction. It is useful whenever a smooth function has a finite trajectory contained in well-separated quadratic regions.\n\n\\begin{lemma}[Quadratic-chamber smoothing estimate]\n\\label{lem:a-chamber}\nLet $H:\\R^d\\to\\R$ have a globally $L$-Lipschitz gradient. Suppose that, on the ball of radius $r>0$ about $x$, it agrees with a quadratic function\n\\[\n Q_x(y)=H(x)+\\ip{\\nabla H(x)}{y-x}\n +\\frac12(y-x)^\\top A_x(y-x),\n \\qquad \\norm{A_x}_{\\mathrm{op}}\\leq L.\n\\]\nLet $Z$ be a centered random vector with finite second moment. Then\n\\begin{align}\n \\norm{\\E\\nabla H(x-Z)-\\nabla H(x)}\n &\\leq 2L\\,\\E\\bigl[\\norm Z\\,\\mathbf1_{\\{\\norm Z>r\\}}\\bigr],\n \\label{eq:a-general-gradient-tail}\\\\\n \\left|\\E H(x-Z)-H(x)-\\frac12\\E[Z^\\top A_xZ]\\right|\n &\\leq L\\,\\E\\bigl[\\norm Z^2\\,\\mathbf1_{\\{\\norm Z>r\\}}\\bigr].\n \\label{eq:a-general-value-tail}\n\\end{align}\n\\end{lemma}\n\n\\begin{proof}\nThe gradients of $H$ and $Q_x$ coincide at $x$. Their Lipschitz bounds imply\n\\[\n \\norm{\\nabla H(x-z)-\\nabla Q_x(x-z)}\\leq2L\\norm z.\n\\]\nThe difference is exactly zero for $\\norm z\\leq r$. Also, both functions have the same value and gradient at $x$, so integrating their gradient bounds along the segment from $x$ to $x-z$ gives\n\\[\n |H(x-z)-Q_x(x-z)|\\leq L\\norm z^2.\n\\]\nAgain the difference is zero inside the ball. Taking expectations proves the two tail estimates after noting that\n\\[\n \\E\\nabla Q_x(x-Z)=\\nabla H(x),\\qquad\n \\E Q_x(x-Z)=H(x)+\\frac12\\E[Z^\\top A_xZ].\n\\]\nThese equalities use only $\\E Z=0$.\n\\end{proof}\n\nThe lemma does not require $H$ to be piecewise quadratic everywhere. It requires one local quadratic neighborhood at each point where the estimate is used. Nor does it require the random vector to be Gaussian. Gaussianity is used below to obtain real analyticity and a particularly explicit tail formula.\n\n\\begin{lemma}[Two-dimensional Gaussian tail moments]\n\\label{lem:a-rayleigh}\nIf $Z\\sim\\mathcal N(0,\\sigma^2I_2)$ and $R=\\norm Z$, then for every $r>0$,\n\\begin{align}\n \\Prob(R>r)&=\\exp\\left(-\\frac{r^2}{2\\sigma^2}\\right),\\label{eq:a-tail-prob}\\\\\n \\E[R^2\\mathbf1_{\\{R>r\\}}]\n &=(r^2+2\\sigma^2)\n \\exp\\left(-\\frac{r^2}{2\\sigma^2}\\right),\\label{eq:a-tail-second}\\\\\n \\E[R\\mathbf1_{\\{R>r\\}}]\n &\\leq\\left(r+\\frac{2\\sigma^2}{r}\\right)\n \\exp\\left(-\\frac{r^2}{2\\sigma^2}\\right).\n \\label{eq:a-tail-first}\n\\end{align}\n\\end{lemma}\n\n\\begin{proof}\nPolar integration gives the density $s\\exp[-s^2/(2\\sigma^2)]/\\sigma^2$ for $R$ on $s\\geq0$. Integrating it proves \\eqref{eq:a-tail-prob}. Integrating $s^2$ times this density, or integrating once by parts after the substitution $u=s^2/(2\\sigma^2)$, proves \\eqref{eq:a-tail-second}. On $\\{R>r\\}$, the elementary inequality $R\\leq R^2/r$ gives \\eqref{eq:a-tail-first}.\n\\end{proof}\n\n\\section{Regularity and exact global constants}\n\n\\begin{lemma}[Analytic regularity and convexity parameters]\n\\label{lem:a-regularity}\nThe function $F_a$ in \\eqref{eq:a-definition} is real analytic, globally $1$-smooth, and $\\mu$-strongly convex. Its gradient is\n\\begin{equation}\n \\nabla F_a(x)=\\E\\nabla H_a(x-Z).\n \\label{eq:a-gradient}\n\\end{equation}\nNeither the smoothness constant $1$ nor the strong-convexity constant $\\mu$ can be improved globally.\n\\end{lemma}\n\n\\begin{proof}\nThe function $H_a$ has at most quadratic growth, and its gradient has at most linear growth. Gaussian moments are finite, so differentiation under the integral is justified by domination on compact sets and gives \\eqref{eq:a-gradient}. Averaging the gradient Lipschitz inequality gives\n\\[\n \\norm{\\nabla F_a(x)-\\nabla F_a(y)}\\leq\\norm{x-y}.\n\\]\nSimilarly, average the defining strong-convexity inequality for $H_a$ at $x-Z$ and $y-Z$. This proves that $F_a$ is $\\mu$-strongly convex.\n\nFor analyticity, define for $w\\in\\mathbb C^2$\n\\begin{equation}\n \\mathcal F_a(w)=\\frac1{2\\pi\\sigma^2}\n \\int_{\\R^2}H_a(y)\n \\exp\\left[-\\frac{\\sum_{j=1}^2(w_j-y_j)^2}{2\\sigma^2}\\right]\\,dy.\n \\label{eq:a-entire}\n\\end{equation}\nThere are no complex conjugates in the exponent. If $w=u+iv$, the absolute value of the exponential is\n\\[\n \\exp\\left[-\\frac{\\norm{u-y}^2-\\norm v^2}{2\\sigma^2}\\right].\n\\]\nOn every compact set of $w$, this expression and each of its complex derivatives are bounded by a polynomial in $\\norm y$ times an integrable Gaussian. The same is true after multiplication by $H_a(y)$, which grows at most quadratically. Differentiation under the integral therefore shows that \\eqref{eq:a-entire} is entire in each coordinate, jointly holomorphic, and restricts to $F_a$ on $\\R^2$.\n\nTo prove optimality of the global constants, note that $H_a$ has a bounded Hessian outside the finitely many chamber-boundary rays. Difference quotients of its Lipschitz gradient are bounded. For any fixed $x$, dominated convergence thus gives\n\\[\n \\nabla^2F_a(x)=\\E\\nabla^2H_a(x-Z),\n\\]\nwhere the Hessian of $H_a$ may be defined arbitrarily on the boundary rays, a set hit with probability zero. Choose a unit vector $v$ in the interior of $C_a$. For each fixed $Z$, the point $tv-Z$ lies in the interior of $C_a$ for all sufficiently large $t$. Its Hessian is then $\\mu I$. Bounded convergence gives\n\\[\n \\nabla^2F_a(tv)\\longrightarrow\\mu I\\qquad(t\\to\\infty).\n\\]\nChoose instead $v$ in the interior of the polar cone. The same argument gives $\\nabla^2F_a(tv)\\to I$. A stronger global lower Hessian bound or a smaller global upper Hessian bound would contradict these limits. The optimal constants are consequently exactly $\\mu$ and $1$.\n\\end{proof}\n\nThe limiting argument concerns global constants, not a claim that an analytic Hessian equals an endpoint matrix on an open set. Gaussian averaging generally makes the local Hessian strictly intermediate. Its asymptotic values nonetheless determine the best global bounds.\n\n\\begin{lemma}[Global contraction of the smoothed update]\n\\label{lem:a-contract}\nFor $q=1-\\mu$, the map $T_F(x)=x-h\\nabla F_a(x)$ is $q$-Lipschitz on $\\R^2$.\n\\end{lemma}\n\n\\begin{proof}\nFor any $x,y$, the fundamental theorem of calculus gives\n\\[\n \\nabla F_a(x)-\\nabla F_a(y)=A_{x,y}(x-y),\\qquad\n A_{x,y}=\\int_0^1\\nabla^2F_a(y+t(x-y))\\,dt.\n\\]\nThe matrix $A_{x,y}$ is symmetric and has eigenvalues in $[\\mu,1]$. Hence $I-hA_{x,y}$ has eigenvalues in $[1-h,1-h\\mu]=[-q,q^2]$. Its operator norm is at most $q$, which proves the assertion.\n\\end{proof}\n\n\\section{Explicit orbit comparison at the wider scale}\n\nEvery point $x_k$ of the rational orbit, for $0\\leq k\\leq M$, lies at distance at least $a/2$ from its chamber boundary by Lemma~\\ref{lem:w-angle}. On either visited chamber, the Hessian of $H_a$ has eigenvalues $\\mu,1$, so its trace is $1+\\mu$. Apply the chamber lemma with\n\\[\n r=\\frac a2,\\qquad L=1,\\qquad\n \\frac{r^2}{2\\sigma^2}=B_a,\\qquad e^{-B_a}=\\frac{a^2}{64}.\n\\]\nSince $a\\leq1/16$, we have $B_a>1$. Lemma~\\ref{lem:a-rayleigh} gives the uniform bounds\n\\begin{align}\n \\norm{\\nabla F_a(x_k)-\\nabla H_a(x_k)}\n &\\leq \\varepsilon_g:=\\frac{a^3}{32},\\label{eq:a-eg}\\\\\n \\left|F_a(x_k)-H_a(x_k)-c_\\sigma\\right|\n &\\leq \\varepsilon_v:=\\frac{a^4}{128},\n \\qquad c_\\sigma=\\frac{1+\\mu}{2}\\sigma^2.\n \\label{eq:a-ev}\n\\end{align}\nFor clarity, the first estimate follows from\n\\[\n 2\\left(r+\\frac{2\\sigma^2}{r}\\right)e^{-B_a}\n =2r\\left(1+\\frac1{B_a}\\right)\\frac{a^2}{64}\n \\leq\\frac{a^3}{32},\n\\]\nand the second from\n\\[\n (r^2+2\\sigma^2)e^{-B_a}\n =r^2\\left(1+\\frac1{B_a}\\right)\\frac{a^2}{64}\n \\leq\\frac{a^4}{128}.\n\\]\nThe logarithm has therefore disappeared from the error bounds. It occurs only in the explicit choice of the Gaussian scale.\n\n\\begin{proposition}[Finite orbit and loss discrepancy]\n\\label{prop:a-orbit}\nLet $z_{k+1}=z_k-h\\nabla F_a(z_k)$ with $z_0=x_0$. For $0\\leq k\\leq M$,\n\\begin{align}\n \\norm{z_k-x_k}&\\leq2k\\varepsilon_g\\leq\\frac{a^2}{128},\n \\label{eq:a-orbit-distance}\\\\\n |F_a(z_k)-H_a(x_k)-c_\\sigma|\n &\\leq\\varepsilon_v+4k\\varepsilon_g.\n \\label{eq:a-orbit-values}\n\\end{align}\n\\end{proposition}\n\n\\begin{proof}\nWrite $\\delta_k=\\norm{z_k-x_k}$. Compare the two updates at the base point $x_k$ and use Lemma~\\ref{lem:a-contract}:\n\\[\n \\delta_{k+1}\n \\leq q\\delta_k+h\\norm{\\nabla F_a(x_k)-\\nabla H_a(x_k)}\n \\leq\\delta_k+2\\varepsilon_g.\n\\]\nSince $\\delta_0=0$, induction gives $\\delta_k\\leq2k\\varepsilon_g$. The parameter identity $Ma=1/8$ implies\n\\[\n 2M\\varepsilon_g=\\frac{Ma^3}{16}=\\frac{a^2}{128}.\n\\]\nThis proves \\eqref{eq:a-orbit-distance}.\n\nAt $x_k$, the gradient norm of $H_a$ is at most $1$, because $\\nabla H_a(0)=0$, the function is $1$-smooth, and $\\norm{x_k}\\leq1$. Therefore\n\\[\n \\norm{\\nabla F_a(x_k)}\\leq1+\\varepsilon_g.\n\\]\nAt any point on the segment from $x_k$ to $z_k$, the gradient norm of $F_a$ is at most $1+\\varepsilon_g+\\delta_k<2$. Integration along that segment gives\n\\[\n |F_a(z_k)-F_a(x_k)|\\leq2\\delta_k\\leq4k\\varepsilon_g.\n\\]\nCombining with \\eqref{eq:a-ev} proves \\eqref{eq:a-orbit-values}.\n\\end{proof}\n\nThe comparison uses the local approximation at the known points $x_k$, not at unknown points of the analytic orbit. Global contraction transports that approximation to the actual orbit. This avoids a circular assumption that the smoothed iterates remain in the same projection chambers. The bound in fact implies that they do remain within a small neighborhood of those chambers during the queried interval, but that additional fact is unnecessary for the proof.\n\n\\subsection{Transfer of the negative second differences}\n\nSet\n\\[\n e_k=F_a(z_k)-H_a(x_k)-c_\\sigma.\n\\]\nThe common quadratic offset cancels:\n\\[\n \\kappa_k(F_a)-\\kappa_k(H_a)=e_k-2e_{k+1}+e_{k+2}.\n\\]\nFor $0\\leq k<m=M-1$, Proposition~\\ref{prop:a-orbit} gives\n\\begin{align}\n |\\kappa_k(F_a)-\\kappa_k(H_a)|\n &\\leq4\\varepsilon_v+(16k+16)\\varepsilon_g\\notag\\\\\n &\\leq4\\varepsilon_v+16M\\varepsilon_g\\notag\\\\\n &\\leq\\frac{a^4}{32}+\\frac{a^2}{16}\n =\\frac{2+a^2}{32}a^2\n \\leq\\frac{513}{8192}a^2.\n \\label{eq:a-curvature-error}\n\\end{align}\nCombining this with \\eqref{eq:w-rational-margin} gives the slightly stronger intermediate bound\n\\begin{equation}\n \\kappa_k(F_a)\\leq-\\frac{13647}{8192}a^2\n \\leq-\\frac32a^2.\n \\label{eq:a-raw-margin}\n\\end{equation}\nThe simpler last constant is used in the theorem.\n\n\\subsection{Positive first decrement and nonvanishing total gain}\n\nAt $x_0=z_0$, the normal component of $\\nabla H_a(x_0)$ is $d_0=\\cos\\beta>1/2$. Equation \\eqref{eq:a-eg} shows that $\\nabla F_a(z_0)\\neq0$. Since $0<h<2$ and $F_a$ is $1$-smooth, its first decrement is strictly positive. Equation \\eqref{eq:a-margin} then makes every later queried decrement positive and increasing.\n\nThe error in the first decrement is at most $2\\varepsilon_v+4\\varepsilon_g$. Thus\n\\[\n D_0(F_a)\\leq\\mu+4a^2+\\frac{a^4}{64}+\\frac{a^3}{8}.\n\\]\nThe inequalities $4a^2\\leq4\\mu$ and $a^4/64+a^3/8\\leq\\mu$, valid for $a\\leq1/16$, give \\eqref{eq:a-first}. Summing\n\\[\n D_k(F_a)\\geq D_0(F_a)+\\frac32ka^2\n\\]\nover $0\\leq k<M$ proves\n\\[\n F_a(z_0)-F_a(z_M)\n \\geq M D_0(F_a)+\\frac34a^2mM\n \\geq\\frac{3m}{256M}\\geq\\frac3{512}.\n\\]\nThe initial decrement tends to zero as $m$ grows, while the guaranteed accumulated gain over the increasing-decrement interval remains bounded below by a positive constant.\n\n\\section{Minimizer location and full-orbit convergence}\n\nStrong convexity makes $F_a$ coercive. A continuous coercive function attains its minimum, and strong convexity makes the minimizer unique. The nonnegativity of $H_a$ implies $F_a\\geq0$. Also,\n\\[\n F_a(0)=\\E H_a(-Z)\\leq\\frac12\\E\\norm Z^2=\\sigma^2,\n\\]\nbecause $H_a(y)\\leq\\norm y^2/2$. Strong convexity at the minimizer gives\n\\[\n F_a(0)\\geq F_a(z_*)+\\frac\\mu2\\norm{z_*}^2\n \\geq\\frac\\mu2\\norm{z_*}^2.\n\\]\nTherefore $\\norm{z_*}\\leq\\sigma\\sqrt{2/\\mu}$. This bound is smaller than the generic estimate obtained by dividing a uniform gradient discrepancy by $\\mu$; it uses the nonnegative quadratic growth of the explicit construction.\n\nFinally $T_F(z_*)=z_*$. Lemma~\\ref{lem:a-contract} gives\n\\[\n \\norm{z_k-z_*}\\leq q^k\\norm{z_0-z_*},\n\\]\nfor every $k\\geq0$. In particular the orbit converges globally and geometrically even after the finite concavity interval ends. This completes the proof of Theorem~\\ref{thm:a-main}.\n\n\\begin{remark}[No inference of an infinite accelerating tail]\nThe finite interval does not imply continuing acceleration. At every fixed $m$, the positive decrements are summable because $F_a$ is lower bounded. Eventually they must become small. The explicit convergence bound identifies a global mechanism for this transition without asserting that every subsequent second difference has a particular sign.\n\\end{remark}\n\n\\section{Scaling and arbitrary dimension}\n\nThe normalized theorem extends to other loss scales and state dimensions without changing the mechanism.\n\n\\begin{proposition}[Loss and coordinate scaling]\n\\label{prop:a-scaling}\nLet $F:\\R^d\\to\\R$ be a $1$-smooth, $\\mu$-strongly convex objective with a known gradient-descent orbit $z_k$ at step $h$. For $L>0$, $R>0$, and a constant $b\\in\\R$, set\n\\[\n G(x)=LR^2F(x/R)+b,\\qquad x_0=Rz_0,\\qquad \\eta=h/L.\n\\]\nThen $G$ is $L$-smooth and $L\\mu$-strongly convex, $x_k=Rz_k$ for every $k$, and\n\\[\n G(x_k)-G(x_{k+1})=LR^2D_k(F),\\qquad\n \\kappa_k(G)=LR^2\\kappa_k(F).\n\\]\nAnalyticity is preserved.\n\\end{proposition}\n\n\\begin{proof}\nThe gradient is $\\nabla G(x)=LR\\nabla F(x/R)$. Substitution in the update gives\n\\[\n x_{k+1}=Rz_k-\\frac hL LR\\nabla F(z_k)=Rz_{k+1}.\n\\]\nThe smoothness and strong-convexity constants follow from the same scaling, and the scalar loss identities are immediate. Affine composition and multiplication preserve analyticity.\n\\end{proof}\n\nThe additive constant is arbitrary because the trajectory and all differences are unchanged. Thus a negative or positive offset in reported training loss has no bearing on the curvature statements. The coordinate scale $R$ separately controls the magnitude of the loss changes while leaving the condition number and normalized step unchanged.\n\n\\begin{corollary}[Analytic examples in every dimension at least two]\n\\label{cor:a-dimension}\nFor any $d\\geq2$, define\n\\[\n F_{a,d}(x)=F_a(x_1,x_2)+\\frac\\mu2\\sum_{j=3}^d x_j^2.\n\\]\nStarting from $(0,1,0,\\ldots,0)$, its loss sequence is exactly the sequence of Theorem~\\ref{thm:a-main}. It is analytic, $1$-smooth, and $\\mu$-strongly convex, with condition number exactly $1/\\mu$.\n\\end{corollary}\n\n\\begin{proof}\nThe additional coordinates start at zero and remain zero under the update $x_j\\mapsto(1-h\\mu)x_j$. The Hessian is block diagonal, so its eigenvalues lie in $[\\mu,1]$. The two-dimensional block retains the optimal endpoint constants established in Lemma~\\ref{lem:a-regularity}.\n\\end{proof}\n\nThe dimension statement does not claim that one-dimensional examples have the same behavior. It establishes that the finite-run phenomenon is not confined to a function whose parameter dimension is exactly two: the same loss traces appear inside higher-dimensional analytic, strongly convex models.\n\n\\section{Fixed metrics and preconditioned gradient descent}\n\nA fixed change of metric preserves the trajectory interpretation when the algorithm changes with it. Let $P$ be a symmetric positive-definite $d\\times d$ matrix and use the norm $\\norm x_P=\\sqrt{x^\\top Px}$.\n\n\\begin{proposition}[Metric transfer]\n\\label{prop:a-metric}\nLet $F$ be a $1$-smooth, $\\mu$-strongly convex objective in Euclidean coordinates. Define\n\\[\n G_P(x)=F(P^{1/2}x),\\qquad\n x_{k+1}=x_k-hP^{-1}\\nabla G_P(x_k).\n\\]\nIf $x_0=P^{-1/2}z_0$, then $x_k=P^{-1/2}z_k$, where $z_k$ is ordinary gradient descent on $F$ with step $h$. In particular the loss sequences, decrements, and second differences coincide exactly.\n\nThe function $G_P$ is $1$-smooth and $\\mu$-strongly convex in the $P$-metric. In Euclidean coordinates it is at most $\\lambda_{\\max}(P)$-smooth and at least $\\mu\\lambda_{\\min}(P)$-strongly convex. If $F=F_a$ in dimension two, these Euclidean constants are globally optimal.\n\\end{proposition}\n\n\\begin{proof}\nThe ordinary Euclidean gradient is\n\\[\n \\nabla G_P(x)=P^{1/2}\\nabla F(P^{1/2}x).\n\\]\nMultiply the preconditioned update by $P^{1/2}$ to obtain\n\\[\n P^{1/2}x_{k+1}\n =P^{1/2}x_k-h\\nabla F(P^{1/2}x_k).\n\\]\nThis is exactly the update for $z_k$. The $P$-metric isometry $x\\mapsto P^{1/2}x$ transfers the smoothness and strong-convexity inequalities. In Euclidean coordinates the Hessian is\n\\[\n \\nabla^2G_P(x)=P^{1/2}\\nabla^2F(P^{1/2}x)P^{1/2},\n\\]\nwhose eigenvalues are bounded below by $\\mu\\lambda_{\\min}(P)$ and above by $\\lambda_{\\max}(P)$. For $F_a$, the Hessian limits $\\mu I$ and $I$ proved in Lemma~\\ref{lem:a-regularity} become $\\mu P$ and $P$. These limits establish optimality of the two Euclidean constants.\n\\end{proof}\n\nThis result covers a fixed positive-definite preconditioner. A changing preconditioner, momentum, or an adaptive optimizer is a different algorithm and requires separate trajectory constraints. Applying the metric theorem to such an algorithm without checking its update equation would be an invalid extension.\n\n\\section{Direct sums and curvature budgets}\n\nThe construction can be combined with other objectives while retaining explicit control of its curve shape. The cleanest combination uses independent blocks.\n\n\\begin{proposition}[Heterogeneous block composition]\n\\label{prop:a-direct-sum}\nFor $1\\leq j\\leq r$, let $F_j:\\R^{d_j}\\to\\R$ be $1$-smooth and $\\mu_j$-strongly convex, and suppose its orbit at step $h_j\\in(0,2)$ is $z_k^{(j)}$. Fix any common step $\\eta>0$, put $c_j=h_j/\\eta$, and define\n\\[\n G(x^{(1)},\\ldots,x^{(r)})=\\sum_{j=1}^r c_jF_j(x^{(j)}).\n\\]\nGradient descent on $G$ at step $\\eta$, initialized by the block initial points, has the block orbit $x_k^{(j)}=z_k^{(j)}$. It is $L_G$-smooth and $\\mu_G$-strongly convex for\n\\[\n L_G=\\max_jc_j,\\qquad \\mu_G=\\min_jc_j\\mu_j.\n\\]\nIts normalized step obeys $\\eta L_G=\\max_jh_j<2$, and its second differences are\n\\[\n \\kappa_k(G)=\\sum_{j=1}^r c_j\\kappa_k(F_j).\n\\]\nIf each component has $\\kappa_k(F_j)\\leq-\\Gamma_j$ on a common interval, then $\\kappa_k(G)\\leq-\\sum_jc_j\\Gamma_j$ there. Analyticity is preserved when each component is analytic.\n\\end{proposition}\n\n\\begin{proof}\nThe gradient in block $j$ is $c_j\\nabla F_j(x^{(j)})$. Since $\\eta c_j=h_j$, each block follows its declared orbit independently. The Hessian is block diagonal, so the displayed constants follow by taking the largest upper block bound and smallest lower block bound. Loss differences commute with finite summation. Analyticity of a finite sum is immediate.\n\\end{proof}\n\nHeterogeneous blocks allow several prescribed finite-run mechanisms to coexist at one common algorithmic step. Their losses add, so one must budget the curvature of every added component. A component whose second differences are positive can weaken or remove the desired negative interval.\n\n\\begin{proposition}[Adding nonzero quadratic coordinates]\n\\label{prop:a-quadratic-budget}\nLet $F_a$ be the analytic witness, and consider\n\\[\n G(z,y)=F_a(z)+\\frac12\\sum_{j=1}^s\\lambda_jy_j^2,\n \\qquad \\lambda_j\\in[\\mu,1].\n\\]\nAt step $h=2-\\mu$, the extra coordinates satisfy\n\\[\n y_{j,k}=(1-h\\lambda_j)^ky_{j,0}.\n\\]\nTheir contribution to the second difference is\n\\begin{equation}\n Q_k=\n \\frac12\\sum_{j=1}^s\n \\lambda_j\\bigl[h\\lambda_j(2-h\\lambda_j)\\bigr]^2\n (1-h\\lambda_j)^{2k}y_{j,0}^2\\geq0.\n \\label{eq:a-quadratic-curvature}\n\\end{equation}\nConsequently, if\n\\begin{equation}\n \\frac12\\sum_{j=1}^s\n \\lambda_j\\bigl[h\\lambda_j(2-h\\lambda_j)\\bigr]^2y_{j,0}^2\n \\leq\\Gamma<\\frac32a^2,\n \\label{eq:a-quadratic-budget}\n\\end{equation}\nthen $G$ has all the same $m$ strictly negative second differences, with\n\\[\n \\kappa_k(G)\\leq-\\frac32a^2+\\Gamma<0.\n\\]\n\\end{proposition}\n\n\\begin{proof}\nThe formula for $y_{j,k}$ follows from the scalar quadratic update. Its loss is\n\\[\n q_{j,k}=\\frac{\\lambda_j}{2}(1-h\\lambda_j)^{2k}y_{j,0}^2.\n\\]\nTaking a second difference gives\n\\[\n q_{j,k}-2q_{j,k+1}+q_{j,k+2}\n =\\frac{\\lambda_j}{2}(1-h\\lambda_j)^{2k}\n \\bigl[1-(1-h\\lambda_j)^2\\bigr]^2y_{j,0}^2,\n\\]\nwhich is \\eqref{eq:a-quadratic-curvature}. Since $0<h\\lambda_j<2$, the factor $(1-h\\lambda_j)^{2k}$ is at most $1$. The budget assumption bounds $Q_k$ by $\\Gamma$ uniformly, and Theorem~\\ref{thm:a-main} supplies the negative contribution from the first block.\n\\end{proof}\n\nThis last statement supplies a quantitative neighborhood of nontrivial higher-dimensional examples, rather than merely appending coordinates initialized at their minima. It also explains why the phenomenon is absent from a sum of independent convex quadratics alone: every term in \\eqref{eq:a-quadratic-curvature} is nonnegative. The switching geometry of the first block is essential to the present construction.\n\n\\section{A reusable analytic realization principle}\n\nThe proof yields a general finite-horizon transfer criterion that can be used without the special numerical constants above.\n\n\\begin{proposition}[Common-trace quadratic realization criterion]\n\\label{prop:a-general-realization}\nLet a $1$-smooth, $\\mu$-strongly convex function $H$ have an exact constant-step orbit $x_0,\\ldots,x_N$ at a step $h\\in(0,2)$ for which the Gaussian-smoothed update is nonexpansive. Suppose every $x_k$ lies at distance at least $r$ inside a region where $H$ equals a quadratic with the same Hessian trace $t_0$. Suppose also that $\\norm{\\nabla H(x_k)}\\leq G$.\n\nLet $F=\\E H(\\,\\cdot-Z)$ with $Z\\sim\\mathcal N(0,\\sigma^2I_d)$, and define\n\\[\n b_g=2\\E[\\norm Z\\,\\mathbf1_{\\{\\norm Z>r\\}}],\\qquad\n b_v=\\E[\\norm Z^2\\,\\mathbf1_{\\{\\norm Z>r\\}}].\n\\]\nIf $z_0=x_0$ and $z_k$ is the smoothed orbit, then\n\\begin{align}\n \\norm{z_k-x_k}&\\leq khb_g,\\label{eq:a-general-orbit}\\\\\n \\left|F(z_k)-H(x_k)-\\frac{t_0\\sigma^2}{2}\\right|\n &\\leq b_v+(G+b_g)khb_g+\\frac12(khb_g)^2.\n \\label{eq:a-general-loss}\n\\end{align}\nThus any finite collection of strict linear inequalities in the losses transfers whenever its margin exceeds the corresponding coefficient-weighted sum of these error bounds.\n\\end{proposition}\n\n\\begin{proof}\nLemma~\\ref{lem:a-chamber} bounds the gradient discrepancy at each base point by $b_g$ and the value discrepancy after subtraction of the common trace offset by $b_v$. Nonexpansiveness gives the recursion $\\delta_{k+1}\\leq\\delta_k+hb_g$, so $\\delta_k\\leq khb_g$. The gradient norm of $F$ at $x_k$ is at most $G+b_g$. Smoothness along the segment from $x_k$ to $z_k$ gives\n\\[\n |F(z_k)-F(x_k)|\\leq(G+b_g)\\delta_k+\\frac12\\delta_k^2.\n\\]\nThis proves \\eqref{eq:a-general-loss}. For a linear functional $\\sum_k\\alpha_k f_k$, the discrepancy is bounded by the sum of $|\\alpha_k|$ times the displayed errors, after canceling the trace offset when $\\sum_k\\alpha_k=0$. If the coefficients do not sum to zero, the known constant offset must be retained explicitly. A strict inequality with a larger margin therefore remains strict.\n\\end{proof}\n\nThe condition on the smoothed update holds, in particular, for a convex $1$-smooth objective with $h\\in(0,2)$, since the averaged Hessian eigenvalues then lie in $[0,1]$. Strong convexity is retained in the statement to match the main application. The trace condition identifies exactly which quadratic smoothing effects cancel from a trajectory functional. When the traces differ, the resulting deterministic offsets can still be tracked, but they cannot be discarded.\n\n\\section{Transfer of every finite strict trajectory condition}\n\nThe explicit constructions above retain quantitative curvature and conditioning bounds. A more general approximation argument shows that analyticity and an unspecified positive strong-convexity parameter impose no additional obstruction to a finite collection of strict trajectory conditions. This is a transfer consequence of regularization, Gaussian smoothing, and finite-orbit continuity; it does not supply a fixed condition-number guarantee.\n\nFor a globally defined differentiable objective $f$, initial point $x_0$, step $h>0$, and finite $N\\geq0$, write\n\\[\n \\mathcal T_N(f;x_0,h)\n =\\bigl(x_k,f(x_k),\\nabla f(x_k)\\bigr)_{k=0}^N\n\\]\nfor its first-order gradient-descent transcript.\n\n\\begin{theorem}[Analytic strong-convex transfer of finite strict conditions]\n\\label{thm:a-strict-transfer}\nLet $f:\\R^d\\to\\R$ be globally convex and $1$-smooth. Fix $x_0$, $h>0$, and finite $N$. Let $\\Phi_1,\\ldots,\\Phi_s$ be continuous real functions of the finite transcript, and suppose\n\\[\n \\Phi_j\\bigl(\\mathcal T_N(f;x_0,h)\\bigr)>0\n \\qquad(1\\leq j\\leq s).\n\\]\nFor every $\\bar\\mu>0$ and every prescribed positive accuracy for that transcript, there exist $0<\\eps<\\min\\{\\bar\\mu,1/2\\}$ and a real-analytic function $F_\\eps:\\R^d\\to\\R$ that is $\\eps$-strongly convex and $(1-\\eps)$-smooth, hence also $1$-smooth, such that its transcript at the same $x_0$ and same step $h$ has the prescribed accuracy and satisfies all the same strict inequalities.\n\nThe function can be taken to be the restriction of an entire function. If optimal global smoothness exactly equal to $1$ is required, the same strict conditions can be realized in dimension $d+1$ by adjoining an unused quadratic coordinate.\n\\end{theorem}\n\n\\begin{proof}\nFor $0<\\eps<1/2$, define\n\\[\n \\Psi_\\eps(x)=(1-2\\eps)f(x)+\\frac\\eps2\\norm{x}^2,\n \\qquad\n F_\\eps(x)=\\E\\Psi_\\eps(x-\\eps Z),\n \\qquad Z\\sim\\mathcal N(0,I_d).\n\\]\nThe nonnegative coefficient $1-2\\eps$ preserves convexity of the first term. The second term makes $\\Psi_\\eps$ $\\eps$-strongly convex, and its gradient Lipschitz constant is at most\n$(1-2\\eps)+\\eps=1-\\eps$.\nGaussian averaging preserves both bounds. A globally smooth function has at most quadratic growth, so the entire Gaussian-kernel argument in Lemma~\\ref{lem:a-regularity} applies in dimension $d$ and proves the claimed analyticity.\n\nThe regularized functions satisfy $\\Psi_\\eps\\to f$ and $\\nabla\\Psi_\\eps\\to\\nabla f$ uniformly on every compact set. The centered Gaussian and smoothness give the uniform estimates\n\\[\n 0\\leq F_\\eps(x)-\\Psi_\\eps(x)\n \\leq\\frac d2\\eps^2,\\qquad\n \\norm{\\nabla F_\\eps(x)-\\nabla\\Psi_\\eps(x)}\n \\leq\\sqrt d\\,\\eps.\n\\]\nFor the value inequality, integrate the smoothness estimate at $x$; the linear term has zero expectation, and convexity supplies the lower inequality. For the gradient inequality, average the Lipschitz estimate and use $\\E\\norm Z\\leq\\sqrt d$. Thus $F_\\eps\\to f$ in $C^1$ uniformly on compact sets.\n\nWe verify that this convergence transfers to the finite orbit. Let $x_k$ be the orbit of $f$ and let $y_k^\\eps$ be the orbit of $F_\\eps$, both from $x_0$. Choose a closed ball $K$ containing every $x_k$, $0\\leq k\\leq N$, with distance at least one from its boundary, and put\n\\[\n \\alpha_\\eps=\\sup_{x\\in K}\n \\norm{\\nabla F_\\eps(x)-\\nabla f(x)}\\longrightarrow0.\n\\]\nWhile $y_k^\\eps\\in K$, the $1$-Lipschitz property of $\\nabla f$ gives\n\\[\n \\norm{y_{k+1}^\\eps-x_{k+1}}\n \\leq(1+h)\\norm{y_k^\\eps-x_k}+h\\alpha_\\eps.\n\\]\nStarting with zero error, induction bounds this discrepancy by\n$\\alpha_\\eps[(1+h)^k-1]$. For sufficiently small $\\eps$, this bound is less than one for every $k\\leq N$, which justifies the induction within $K$. Therefore all queried states converge. Uniform compact $C^1$ convergence and continuity of $f,\\nabla f$ then show that all queried values and gradients converge as well.\n\nThe finitely many strict continuous inequalities define an open neighborhood of the limiting transcript. Every sufficiently small positive $\\eps$ places the new transcript in that neighborhood and within any prescribed accuracy. The same choice may be made below $\\bar\\mu$.\n\nFor the final assertion define\n$\\widehat F_\\eps(x,u)=F_\\eps(x)+u^2/2$\nand initialize at $(x_0,0)$. The extra coordinate remains zero under every update. The first $d$ coordinates and all losses are unchanged, while the added quadratic makes the optimal global smoothness constant exactly $1$. The original conditions are interpreted on the first $d$ coordinates; strong convexity is retained.\n\\end{proof}\n\n\\begin{corollary}[Equality of feasible finite strict curvature patterns]\n\\label{cor:a-finite-patterns}\nFix a dimension, initial point, positive constant step, and finite horizon. Any finite strict sign pattern of loss decrements and second differences realizable by a globally convex $1$-smooth objective is also realizable by a real-analytic, strongly convex, globally $1$-smooth objective, when its positive strong-convexity parameter is not fixed in advance.\n\\end{corollary}\n\n\\begin{proof}\nEach requested positive or negative sign is a strict linear inequality in finitely many loss values, hence a continuous strict condition on the transcript. Apply Theorem~\\ref{thm:a-strict-transfer}. The reverse inclusion is immediate because every analytic strongly convex $1$-smooth objective belongs to the convex $1$-smooth class.\n\\end{proof}\n\nThis theorem does not preserve exact jets, exact interpolation equalities, or a prescribed lower bound on the optimal strong-convexity constant. It also does not transfer an infinite collection of conditions by one compactness argument: the regularization needed for a longer horizon may be smaller. In the stable range $0<h<2$, the resulting strongly convex objectives have globally convergent gradient descent, but their rates may become arbitrarily slow as $\\eps$ decreases. The explicit earlier constructions remain necessary when a useful condition-number or curvature-margin bound is part of the claim.\n\n\\section{Scope of the analytic conclusion}\n\nThe original existence question asks whether consecutive concavity can occur under the stated optimization assumptions. Theorem~\\ref{thm:a-main} provides an affirmative construction in a much narrower function class: entire real restrictions, global strong convexity, a finite explicit condition number, and globally convergent constant-step gradient descent. These restrictions rule out explanations based solely on nonsmoothness, nonconvexity, divergence, or a nonexistent minimizer.\n\nThe theorem does not give a universal shape characterization for every analytic objective. The exact boundary for two adjacent negative second differences is established separately in Chapter~\\ref{ch:adjacency}; sharp thresholds for every longer prescribed run are not established here. Nor does the present construction prove that $128(m+1)$ is the optimal condition-number requirement. The spectral ceiling in Chapter~\\ref{ch:witness} concerns the reflected family, while the general ceilings there leave a gap. These remaining boundaries are distinct from the all-finite-length existence theorem proved above.\n\nFor recursive learning, the immediate consequence is a rigorous stress test for continuation and stopping rules. A rule that treats a small initial decrement as evidence of uniformly diminishing future gains fails on this explicit family. The conditional decision theory elsewhere in this manuscript instead bases guarantees on declared function classes and verifiable trajectory constraints. The analytic construction supplies exact counterexamples against which such declarations can be tested.\n\n"}
{"format": "latex", "id": "06_transcripts", "sha256": "91045541d43906271d6968ad47c64a7baa79733211156afc75b809a76d72a0d6", "source_path": "research/chapters/06_transcripts.tex", "text": "\\chapter{Complete certificates for finite learning trajectories}\n\\label{ch:transcripts}\n\nThe preceding chapters give closed-form laws and explicit counterexamples.\nThis chapter gives a complete finite representation of the underlying\ndecision problem. A controller receives a finite transcript, declares a\nfunction class and an update rule, and asks which future observations are\nstill possible. The answer is a convex feasibility problem when the ambient\ndimension is unrestricted and the first-order coefficients are fixed.\n\nThe representation is an application of smooth convex interpolation and\nperformance estimation \\cite{taylor2017,drori2014}. It is not a claim to\nhave invented those methods. The purpose here is to make their role in\ntrajectory certification explicit, to prove the relevant completeness and\ndimension statements, and to provide a small exact certificate checker\nthat can be separated from a numerical search procedure.\n\n\\section{Prescribed first-order algorithms}\n\nFix a horizon $N$ and real coefficients $s_{ij}$ with $s_{ij}=0$ for\n$j\\geq i$. Consider\n\\begin{equation}\n x_i=x_0+\\sum_{j<i}s_{ij}g_j,\n \\qquad g_j=\\nabla f(x_j),\\qquad 0\\leq i\\leq N.\n \\label{eq:fixed-first-order}\n\\end{equation}\nConstant-step gradient descent has $s_{ij}=-\\eta$ for $j<i$.\nVariable-step descent has $s_{ij}=-\\eta_j$. Prescribed momentum methods\ncan also be expanded in this form. Their coefficients may be large or\nunstable: the exact finite representation does not require convergence.\n\nTranslate the initial point to $x_0=0$ and the initial loss to $f_0=0$.\nLet $e_i$ be the $i$th coordinate vector of $\\R^{N+1}$, and let\n$s_i=(s_{i0},\\ldots,s_{iN})^T$. If $V$ is the matrix whose columns are\n$g_0,\\ldots,g_N$, then $x_i=Vs_i$ and $g_i=Ve_i$.\nAll relevant inner products are encoded by\n\\begin{equation}\n G=V^TV\\succeq0.\n \\label{eq:gram-def}\n\\end{equation}\n\nFor $a_{ij}=s_i-s_j$ and\n$b_{ij}=e_i-e_j-\\mu a_{ij}$, define the affine functional\n\\begin{align}\n \\mathcal I_{ij}(G,f)\n ={}& f_i-f_j-e_j^TGa_{ij}\n -\\frac\\mu2a_{ij}^TGa_{ij}\n -\\frac1{2(L-\\mu)}b_{ij}^TGb_{ij}.\n \\label{eq:gram-interpolation}\n\\end{align}\nEvery apparently quadratic expression here is linear in the entries of\n$G$. The coefficients $a_{ij},b_{ij}$ are fixed data, not decision\nvariables. This distinction is what makes the formulation semidefinite.\n\n\\begin{theorem}[Exact finite transcript representation]\n\\label{thm:finite-transcript}\nLet $0\\leq\\mu<L$ and fix \\eqref{eq:fixed-first-order}.\nA proposed loss vector $(f_0,\\ldots,f_N)$ with $f_0=0$ is realizable\nin some finite dimension by $f\\in\\Fclass_{\\mu,L}$ if and only if\nthere is $G\\succeq0$ satisfying\n\\begin{equation}\n \\mathcal I_{ij}(G,f)\\geq0\\qquad(0\\leq i,j\\leq N).\n \\label{eq:finite-feasible}\n\\end{equation}\nEvery feasible pair $(G,f)$ has a realization in dimension\n$\\rank G\\leq N+1$.\n\\end{theorem}\n\n\\begin{proof}\nAn actual trajectory gives \\eqref{eq:gram-def}, and substituting its\ninner products into \\eqref{eq:strong-interpolation} gives\n\\eqref{eq:finite-feasible}. Conversely, factor a feasible matrix as\n$G=V^TV$ with $V\\in\\R^{r\\times(N+1)}$, where $r=\\rank G$.\nDefine $g_i=Ve_i$ and $x_i=Vs_i$. The inequalities\n\\eqref{eq:finite-feasible} are exactly the strong interpolation\ninequalities for these vectors and the specified values. Theorem\n\\ref{thm:strong-interpolation} constructs a global function with these\nvalues and gradients. By construction its iterates obey the prescribed\nalgorithm. This verifies the update rule as well as interpolation.\n\\end{proof}\n\n\\begin{remark}[Fixed dimension]\nIf a problem must live in dimension $d<N+1$, the additional condition\n$\\rank G\\leq d$ is required. Dropping it is a relaxation of that fixed\ndimension problem. It is exact for dimension-free guarantees and for\ndimensions at least $N+1$. A positive semidefinite feasibility result\nalone must not be described as an exact one-dimensional realization.\n\\end{remark}\n\n\\section{Measurements, normalization, and compactness}\n\nThe homogeneous model needs a scale. Suppose\n$\\norm{g_0}\\leq R_g$, where $R_g<\\infty$ is declared. This is the\nlinear Gram constraint $G_{00}\\leq R_g^2$. Exact observations of inner\nproducts are linear equalities in $G$. Interval observations of losses,\nloss differences, and inner products are linear inequalities. For example,\n\\[\n \\widehat D_i-\\nu_i\\leq f_i-f_{i+1}\\leq\\widehat D_i+\\nu_i.\n\\]\nThese may be imposed without asserting that measurement errors at\ndifferent times are independent.\n\nFor raw losses with an uncertain common offset, introduce a scalar $b$\nand impose $\\abs{f_i+b-\\widehat f_i}\\leq\\nu_i$. Since $f_0=0$,\nthe interval at the initial point bounds $b$. Subtracting a noisy initial\nreading and then pretending all resulting errors are independent would\nlose the common-error structure; this formulation retains it.\n\n\\begin{lemma}[Explicit bounds for every feasible transcript]\n\\label{lem:compact-bounds}\nDefine recursively\n\\begin{align}\n C_0&=R_g,\\qquad\n R_i=\\sum_{j<i}|s_{ij}|C_j,\\qquad C_i=R_g+LR_i\\quad(i\\geq1),\n \\label{eq:transcript-C}\\\\\n B_i&=R_gR_i+\\frac L2R_i^2,\\qquad\n T=\\sum_{i=0}^N C_i^2,\n \\label{eq:transcript-BT}\n\\end{align}\nwith $R_0=B_0=0$. Every feasible pair with $G_{00}\\leq R_g^2$\nsatisfies\n\\[\n G_{ii}\\leq C_i^2,\\qquad \\tr G\\leq T,\\qquad |f_i|\\leq B_i.\n\\]\n\\end{lemma}\n\n\\begin{proof}\nUse the realizing function from Theorem~\\ref{thm:finite-transcript}.\nInduction in \\eqref{eq:fixed-first-order} gives\n$\\norm{x_i}\\leq R_i$ and\n$\\norm{g_i}\\leq\\norm{g_0}+L\\norm{x_i}\\leq C_i$.\nThe descent inequality between $0$ and $x_i$, together with $f_0=0$,\ngives $|f_i|\\leq R_gR_i+LR_i^2/2$. Summing the squared gradient bounds\ngives the trace bound.\n\\end{proof}\n\nFor stable constant-step gradient descent a sharper choice is\n$C_i=R_g$, $R_i=i\\eta R_g$ and\n$-i\\eta R_g^2\\leq f_i\\leq0$, by\n\\eqref{eq:gradient-decrease} and \\eqref{eq:drop-sandwich}.\nThe general bounds are retained because they also cover fixed momentum\ncoefficients and steps outside the stable range over a finite horizon.\n\n\\begin{corollary}[Compact convex model set]\n\\label{cor:compact-model}\nThe set $\\mathcal K$ of feasible $(G,f)$ obeying the scale bound and\nany closed affine observation constraints is a compact convex set,\npossibly empty. A bounded common offset may be included as an extra\ncoordinate without changing this conclusion.\n\\end{corollary}\n\n\\begin{proof}\nAll defining constraints are closed and convex. Lemma\n\\ref{lem:compact-bounds} bounds the diagonal of $G$ and all the values.\nFor a positive semidefinite matrix,\n$|G_{ij}|\\leq\\sqrt{G_{ii}G_{jj}}$, so every matrix entry is bounded.\nClosed bounded subsets of finite-dimensional Euclidean space are compact.\n\\end{proof}\n\nCompactness resolves several otherwise subtle issues. Forecast extrema\nare attained. A strict violated law has an actual finite-dimensional\nwitness. Approximate feasibility cannot escape to an unbounded gradient\nsequence under the scale constraint. None of these conclusions follows\nmerely from writing down an unnormalized semidefinite problem.\n\n\\section{Exact prediction regions and optimal logical decisions}\n\nLet a scalar query be affine in the finite data:\n\\begin{equation}\n Q(G,f)=c^Tf+\\tr(CG)+c_0,\n \\label{eq:linear-query}\n\\end{equation}\nwhere $C$ is symmetric. Examples are cumulative future improvement,\none second difference, a weighted combination of curvatures, a squared\ngradient norm, or a fixed quadratic combination of the iterates.\n\n\\begin{theorem}[Complete prediction interval]\n\\label{thm:complete-prediction}\nSuppose $\\mathcal K\\neq\\varnothing$. Then\n\\[\n q_-:=\\min_{(G,f)\\in\\mathcal K}Q(G,f),\\qquad\n q_+:=\\max_{(G,f)\\in\\mathcal K}Q(G,f)\n\\]\nexist, and the set of all realizable query values is exactly\n$[q_-,q_+]$. Every value in this interval has a global smooth convex\nrealization of dimension at most $N+1$.\n\\end{theorem}\n\n\\begin{proof}\nCompactness and continuity give both extrema. If $z_-,z_+$ are\nminimizers and maximizers, the segment\n$(1-t)z_-+tz_+$ belongs to $\\mathcal K$. The affine query takes every\nvalue $(1-t)q_-+tq_+$. Theorem~\\ref{thm:finite-transcript} realizes\neach point of the segment. Conversely, every admissible trajectory has a\npair in $\\mathcal K$, so its query lies between the extrema.\n\\end{proof}\n\nFor vector queries, the same argument gives a compact convex prediction\nregion. Its support function is obtained by optimizing scalar linear\ncombinations of the queries. This is useful when a controller must compare\nimprovement and gradient size jointly rather than reduce them to one\nnumber prematurely.\n\n\\begin{corollary}[Maximal sound threshold rule]\n\\label{cor:maximal-rule}\nFor the claim $Q\\geq\\gamma$, the available transcript and model imply\nthe claim if and only if $q_-\\geq\\gamma$. They imply its strict failure\nif and only if $q_+<\\gamma$. If $q_-<\\gamma\\leq q_+$, there are\nadmissible objectives on opposite sides of the decision boundary.\nNo rule using only the same information can assert the same binary answer\ncorrectly on all these objectives.\n\\end{corollary}\n\n\\begin{proof}\nThe first two assertions follow from the exact interval. In the third\ncase its attained endpoints give two indistinguishable admissible\ntranscripts with different truth values for the claim. A rule receiving\nthe same input must return the same answer, so one of these objectives\nwould refute an unconditional binary assertion.\n\\end{proof}\n\nFor a training allocation decision, set $Q=f_t-f_{t+r}$ and let\n$\\gamma$ be a declared value of the additional compute. The lower\nendpoint can justify continuation under the model; the upper endpoint can\njustify rejecting that continuation. An interval crossing $\\gamma$\nrecords genuine model uncertainty. It is not resolved by replacing the\ninterval with its midpoint and calling that number a guarantee.\n\n\\section{Monotonicity under additional information}\n\n\\begin{proposition}[Information can only narrow a fixed query]\nIf new valid observations replace $\\mathcal K$ by a nonempty\n$\\mathcal K'\\subseteq\\mathcal K$, the prediction interval for the same\nquery satisfies\n\\[\n [q_-',q_+']\\subseteq[q_-,q_+].\n\\]\n\\end{proposition}\n\n\\begin{proof}\nMinimization over a smaller feasible set cannot decrease its minimum;\nmaximization cannot increase its maximum.\n\\end{proof}\n\nThe qualification ``the same query'' matters. A moving-horizon controller\nmay ask a different question after each observation, and its interval\nwidth need not decrease. Similarly, changing $L$, changing the algorithm,\nor changing the objective creates a different model set. Those changes\nmust be recorded instead of being treated as extra information about an\nunchanged model.\n\n\\section{A sound algebraic certificate language}\n\nCollect all scalar affine constraints as\n$\\ell_j(G,f)\\geq0$. These include interpolation, normalization, and\nmeasurement constraints. Equalities can either be eliminated or represented\nby a pair of opposite inequalities. A certificate of $Q\\geq b$ is an\nidentity\n\\begin{equation}\n Q(G,f)-b=\\sum_j\\lambda_j\\ell_j(G,f)+\\tr(PG),\n \\qquad \\lambda_j\\geq0,\\quad P\\succeq0.\n \\label{eq:exact-dual-certificate}\n\\end{equation}\nEvery term on the right is nonnegative on the model set. This proves the\nbound without solving the optimization problem during verification.\nThe universal compensation identity in Chapter~\\ref{ch:universal} is\nan especially small certificate of this form after switching from\ngradients to gradient differences.\n\nNumerical search often returns coefficients with rounding error. Dropping\nthat error is not a proof. A residual version of the certificate is both\nsimple and useful.\n\n\\begin{theorem}[Residual-tolerant exact verification]\n\\label{thm:residual-cert}\nAssume $G\\succeq0$, $\\tr G\\leq T$, and $|f_i|\\leq B_i$ on the\nmodel set. Suppose exact coefficient comparison gives\n\\begin{equation}\n Q-b-\\sum_j\\lambda_j\\ell_j=\\tr(PG)+r^Tf+s,\n \\label{eq:residual-identity}\n\\end{equation}\nwhere $\\lambda_j\\geq0$ and $P+\\delta I\\succeq0$ for some\n$\\delta\\geq0$. Then every feasible transcript satisfies\n\\begin{equation}\n Q\\geq b+s-\\delta T-\\sum_i|r_i|B_i.\n \\label{eq:residual-bound}\n\\end{equation}\n\\end{theorem}\n\n\\begin{proof}\nPositive semidefiniteness gives\n$\\tr(PG)\\geq-\\delta\\tr G\\geq-\\delta T$.\nAlso $r^Tf\\geq-\\sum_i|r_i|B_i$. Substitute these inequalities into\n\\eqref{eq:residual-identity}, using nonnegativity of the constraint terms.\n\\end{proof}\n\nFor rational problem data and rational certificate coefficients, all parts\nof this check can use exact fractions. Positive semidefiniteness can be\nchecked by exact symmetric elimination: a positive diagonal pivot permits\na Schur complement; a zero pivot requires its remaining row and column\nto vanish; a negative pivot refutes positive semidefiniteness. The\ncompanion implementation follows exactly these rules and never treats a\nsmall negative floating-point pivot as zero.\n\n\\section{Completeness away from a decision boundary}\n\nIt is useful to know what this certificate language can express. Exact\nboundary attainment by a particular semidefinite dual is delicate. The\nfollowing result avoids assuming a strict feasible point or silently\nassuming strong duality.\n\n\\begin{theorem}[Arbitrarily accurate certificates on a bounded model]\n\\label{thm:certificate-completeness}\nLet\n\\[\n K_0=\\{(G,f):G\\succeq0,\\ \\tr G\\leq T,\\ |f_i|\\leq B_i\\}\n\\]\nand let the nonempty model set be\n$\\mathcal K=\\{z\\in K_0:\\ell_j(z)\\geq0\\ \\forall j\\}$.\nFor every $b<\\min_{z\\in\\mathcal K}Q(z)$, there exist\n$\\lambda_j\\geq0$ such that\n\\begin{equation}\n \\min_{z\\in K_0}\\left(Q(z)-\\sum_j\\lambda_j\\ell_j(z)\\right)>b.\n \\label{eq:strict-complete}\n\\end{equation}\nIf all problem data are rational, the multipliers and a residual\ncertificate proving a lower bound strictly greater than $b$ may be chosen\nrational. Thus every strict positive decision margin has a finite exact\ncertificate. At a zero margin, certificates approaching the optimum from\nbelow are guaranteed; an exact finite certificate at the boundary is not\nasserted by this theorem.\n\\end{theorem}\n\n\\begin{proof}\nWrite $m$ for the number of scalar constraints and form the convex set\n\\[\n C=\\{(\\ell_1(z),\\ldots,\\ell_m(z),Q(z)):z\\in K_0\\}\n +\\bigl(\\R_{\\leq0}^{m}\\times\\R_{\\geq0}\\bigr).\n\\]\nIt is closed: the first summand is compact, so in any convergent sequence\nof sums a subsequence of the compact summands converges, and the residual\nconverges in the closed cone. The point $(0,\\ldots,0,b)$ lies outside\n$C$, since membership would imply a feasible $z$ with $Q(z)\\leq b$.\n\nStrict separation of a point from a closed convex set supplies a linear\nfunctional separating this point from $C$. One elementary construction\nuses the nearest point of $C$ to the external point and the line joining\nthem; the projection inequality gives the strict separation. Because\n$C$ is closed under decreasing its first $m$ coordinates and increasing\nits last coordinate, a separating normal has the form\n$(-\\lambda_1,\\ldots,-\\lambda_m,\\beta)$ with\n$\\lambda_j\\geq0$ and $\\beta\\geq0$. Thus\n\\[\n -\\sum_j\\lambda_j\\ell_j(z)+\\beta Q(z)>\\beta b\n \\qquad(z\\in K_0).\n\\]\nThe coefficient $\\beta$ cannot be zero: a feasible $z$ would then make\nthe left side nonpositive. Dividing by $\\beta$ proves\n\\eqref{eq:strict-complete}.\n\nFor rational data, approximate the nonnegative multipliers by nonnegative\nrational numbers. Every $\\ell_j$ is bounded on compact $K_0$, so\nsufficiently accurate approximations preserve a strict margin. The\nresulting affine expression has the form $\\tr(PG)+r^Tf+s$ with\nrational coefficients. Its minimum over $K_0$ is exactly\n\\[\n s-T\\max\\{0,-\\lambda_{\\min}(P)\\}-\\sum_iB_i|r_i|.\n\\]\nIndeed, the matrix minimum is attained either at $G=0$ or at a rank-one\nmatrix of trace $T$ in a minimum-eigenvalue direction; each scalar box\nminimum is attained at an endpoint. Choose a rational $\\delta$ slightly\nlarger than $\\max\\{0,-\\lambda_{\\min}(P)\\}$, while preserving the\nstrict margin. Then $P+\\delta I\\succeq0$ has an exact rational\nsemidefinite check, and Theorem~\\ref{thm:residual-cert} certifies a lower\nbound greater than $b$.\n\\end{proof}\n\nThe statement is about existence of verifiable certificates, not a\nuniform complexity bound for finding them. Near an exact threshold the\nrequired precision may be large. The sharp constant in\nChapter~\\ref{ch:universal} is resolved by a symbolic identity and an\nexplicit counterexample construction, not by rounding a numerical\noptimizer's output.\n\n\\section{Bounded oracle error remains finitely representable}\n\nSuppose the update uses $v_i=g_i+e_i$ with either absolute bounds\n$\\norm{e_i}\\leq\\epsilon_i$ or relative bounds\n$\\norm{e_i}\\leq\\delta_i\\norm{g_i}$. Replace\n\\eqref{eq:fixed-first-order} by\n\\[\n x_i=x_0+\\sum_{j<i}s_{ij}(g_j+e_j).\n\\]\nUse a Gram matrix of the $N+1$ true gradients and the $N$ error vectors.\nThe update coefficients remain fixed linear combinations of these\nvectors. Interpolation still concerns the true gradients $g_i$.\nThe error restrictions become affine Gram inequalities,\n\\[\n \\langle e_i,e_i\\rangle\\leq\\epsilon_i^2,\n \\qquad\\text{or}\\qquad\n \\langle e_i,e_i\\rangle\\leq\\delta_i^2\\langle g_i,g_i\\rangle.\n\\]\n\n\\begin{proposition}[Exact finite bounded-error model]\n\\label{prop:error-gram}\nWith the augmented Gram matrix, the finite interpolation constraints,\nupdate equations, and stated error bounds are jointly necessary and\nsufficient for a realized inexact first-order transcript. Dimension at\nmost $2N+1$ suffices. With bounded initial gradient and finite error\nparameters the feasible set is compact after fixing the initial loss.\n\\end{proposition}\n\n\\begin{proof}\nFactor a feasible augmented Gram matrix and recover the vectors\n$g_i,e_i$. Form the points from the update coefficients. The exact\ninterpolation theorem realizes the values and true gradients, and the\nGram error constraints realize the declared deterministic oracle errors.\nConversely an actual inexact trajectory gives those constraints. For\ncompactness, the absolute-error induction becomes\n\\[\n R_i=\\sum_{j<i}|s_{ij}|(C_j+\\epsilon_j),\\qquad\n C_i=R_g+LR_i.\n\\]\nFor relative errors it becomes\n$R_i=\\sum_{j<i}|s_{ij}|(1+\\delta_j)C_j$.\nThese finite recursions bound every vector, then every Gram entry and\nloss value as before.\n\\end{proof}\n\nThis is a worst-case model for realized error vectors. It does not encode\nindependence, unbiasedness, or a noise distribution. Such probabilistic\nassumptions require separate constraints or separate concentration\narguments. The deterministic robustness bounds in the next chapter often\ngive cheaper certificates when the full augmented semidefinite program\nwould be unnecessary.\n\n\\section{Limits of the finite representation}\n\nSeveral restrictions are structural. The coefficients in\n\\eqref{eq:fixed-first-order} must be fixed when the certificate is\nchecked. A future adaptive policy may require separate feasible sets for\nits decision branches. An optimizer with coordinatewise nonlinear\nnormalization is not automatically represented by fixed scalar\ncoefficients. A nonconvex neural network loss is not a member of\n$\\Fclass_{\\mu,L}$ merely because a short transcript is feasible.\n\nFeasibility has an existential meaning: at least one function in the\ndeclared class explains the observations. A violated necessary\ncertificate rejects the conjunction of the assumptions. A passing\ncertificate does not establish that the real training system satisfies\nthem globally. This logical distinction is retained when the mathematical\nmodel is used as a component of a recursive learning controller.\n\nWithin its declared scope, however, the framework is complete. It does\nnot leave hidden infinite-dimensional function choices behind the\nfinite problem. A feasible point constructs a function, a valid algebraic\ncertificate proves a bound, and a strict counterexample to a bound has a\nfinite-dimensional realization. This gives a precise interface between\noptimization theory, exact verification, and decisions based on partial\nlearning histories.\n"}
{"format": "latex", "id": "07_robustness", "sha256": "77b229cc9b0d34a66c4a18985bd2bef2d037b96ad6cd67dd6cbe0766a55a43d0", "source_path": "research/chapters/07_robustness.tex", "text": "\\chapter{Robust laws for imperfect and changing iterations}\n\\label{ch:robustness}\n\nAn exact optimization law is useful only when its assumptions can be\nconnected to the process being observed. Loss measurement error, gradient\nerror, step variation, and objective drift are different perturbations.\nThis chapter gives a general transfer theorem that keeps their effects\nseparate. It complements the sharper local update-residual identity in\nChapter~\\ref{ch:universal} and the exact augmented Gram model in\nChapter~\\ref{ch:transcripts}.\n\n\\section{A reference orbit and a perturbed orbit}\n\nFix $f\\in\\Fclass_{\\mu,L}$, with $0\\leq\\mu<L$, and a constant step\n$0<\\eta<2/L$. Let the exact reference orbit be\n\\[\n x_{k+1}=T(x_k),\\qquad T=I-\\eta\\nabla f.\n\\]\nThe observed states obey\n\\begin{equation}\n y_{k+1}=T(y_k)+r_k,\\qquad \\norm{r_k}\\leq\\rho_k.\n \\label{eq:perturbed-orbit}\n\\end{equation}\nThe residual $r_k$ has units of parameter displacement. For an additive\ngradient error $e_k$, it is $r_k=-\\eta e_k$ and\n$\\rho_k=\\eta\\epsilon_k$ when $\\norm{e_k}\\leq\\epsilon_k$.\n\nSet $q=1$ when $\\mu=0$, and otherwise set\n$q=\\max\\{|1-\\eta\\mu|,|1-\\eta L|\\}<1$.\nSuppose $\\norm{y_0-x_0}\\leq E_0$, and define\n\\begin{equation}\n E_{k+1}=qE_k+\\rho_k,\n \\qquad\n E_k=q^kE_0+\\sum_{j=0}^{k-1}q^{k-1-j}\\rho_j.\n \\label{eq:tube-recurrence}\n\\end{equation}\n\n\\begin{theorem}[Deterministic trajectory and loss tube]\n\\label{thm:trajectory-tube}\nFor every $k$, the two trajectories satisfy\n\\begin{align}\n \\norm{y_k-x_k}&\\leq E_k,\\label{eq:tube-position}\\\\\n |f(y_k)-f(x_k)|&\\leq G_kE_k+\\frac L2E_k^2,\n \\label{eq:tube-values}\n\\end{align}\nwhere any valid upper bound $G_k\\geq\\norm{\\nabla f(x_k)}$ may be used.\nIn particular $G_k=\\norm{g_0}$ is valid for the convex class, and\n$G_k=q^k\\norm{g_0}$ is valid when $\\mu>0$.\n\\end{theorem}\n\n\\begin{proof}\nThe reference map is $q$-Lipschitz by\nPropositions~\\ref{prop:nonexpansive} and \\ref{prop:contraction}. Therefore\n\\[\n \\norm{y_{k+1}-x_{k+1}}\n \\leq q\\norm{y_k-x_k}+\\rho_k.\n\\]\nInduction proves \\eqref{eq:tube-position} and expansion proves the\nexplicit convolution formula in \\eqref{eq:tube-recurrence}. Apply\n\\eqref{eq:descent-lemma} between $x_k$ and $y_k$ to obtain\n\\eqref{eq:tube-values}. The final gradient bounds were proved in\nChapter~\\ref{ch:foundations}.\n\\end{proof}\n\nFor constant $\\rho$, the displacement bound is $E_k\\leq E_0+k\\rho$\nin the merely convex case and\n$E_k\\leq q^kE_0+\\rho(1-q^k)/(1-q)$ in the strongly convex case.\nThe first bound is a finite-horizon statement; the second exposes an\nerror floor. A convergence claim for the perturbed process must account\nfor this floor rather than reuse the exact-orbit convergence theorem.\n\n\\section{Transferring any finite linear law}\n\nSuppose an exact law on the reference trajectory is\n\\begin{equation}\n \\sum_{i=0}^Nc_if(x_i)\\geq b.\n \\label{eq:reference-law}\n\\end{equation}\nLet measured values satisfy\n$|\\widehat\\ell_i-f(y_i)|\\leq\\nu_i$.\nDefine\n\\[\n A_i=\\nu_i+G_iE_i+\\frac L2E_i^2.\n\\]\n\n\\begin{theorem}[Finite-law robustness]\n\\label{thm:law-transfer}\nUnder these assumptions,\n\\begin{equation}\n \\sum_{i=0}^Nc_i\\widehat\\ell_i\n \\geq b-\\sum_{i=0}^N|c_i|A_i.\n \\label{eq:robust-linear-law}\n\\end{equation}\nConsequently a strict violation of \\eqref{eq:robust-linear-law} rejects\nthe conjunction of the reference model, residual bounds, and measurement\nbounds.\n\\end{theorem}\n\n\\begin{proof}\nThe trajectory tube and measurement bound give\n$|\\widehat\\ell_i-f(x_i)|\\leq A_i$. Multiply the errors by $c_i$,\nsum their absolute bounds, and use \\eqref{eq:reference-law}.\n\\end{proof}\n\nThe measurement part $\\sum_i|c_i|\\nu_i$ is sharp for independently\nbounded deterministic value errors: choose each error with the sign\nopposite to $c_i$. The trajectory part is a general upper bound and\nneed not be sharp for a particular construction. In particular, the\nanalytic smoothing theorem of Chapter~\\ref{ch:analytic} uses common\nquadratic structure to obtain substantially better bounds than a generic\nLipschitz perturbation argument.\n\nFor the compensation law take\n$(c_0,c_1,c_2,c_3)=(2/3,-1,0,1/3)$ and $b=0$.\nIf the comparison restarts at the observed state, $E_0=0$. Only three\nloss observations are required. A common additive offset cancels because\n$\\sum_i c_i=0$. The method also transfers any weighted curvature\ncertificate produced by the finite transcript framework.\n\n\\begin{example}[A complete three-step error budget]\nNormalize $L=1$, take $h=\\eta=9/5$, assume\n$\\norm{g_0}\\leq1$, and suppose each gradient error is at most\n$10^{-5}$ and each loss measurement error at most $10^{-6}$.\nThe nonexpansive bound gives $E_i=(9/5)i\\,10^{-5}$.\nThe allowance for the compensation statistic is exactly bounded by\n\\begin{align*}\n B&=2\\cdot10^{-6}\n +E_1+\\tfrac12E_1^2\n +\\tfrac13(E_3+\\tfrac12E_3^2)\\\\\n &=0.000038000648.\n\\end{align*}\nA measured statistic of $-10^{-4}$ contradicts the declared assumptions.\nA statistic of $-3\\cdot10^{-5}$ does not. These conclusions use the\nspecific step, gradient bound, error model, and measurement tolerance;\nthey are not claims about every training pipeline with those raw losses.\n\\end{example}\n\n\\section{Variable steps and objective drift}\n\nSuppose the observed iteration uses\n\\begin{equation}\n y_{k+1}=y_k-\\eta_k\\bigl(\\nabla f_k(y_k)+e_k\\bigr),\n \\label{eq:drifting-iteration}\n\\end{equation}\nwhere $f$ remains the reference objective and each $\\eta_k\\geq0$.\nAssume at the observed state\n\\[\n \\norm{\\nabla f_k(y_k)-\\nabla f(y_k)}\\leq\\delta_k,\n \\qquad\\norm{e_k}\\leq\\epsilon_k.\n\\]\nPut $a_k=|\\eta_k-\\eta|$. Subtracting the reference update gives\n\\[\n \\norm{r_k}\\leq a_k\\norm{\\nabla f(y_k)}\n +\\eta_k(\\delta_k+\\epsilon_k).\n\\]\nSince $\\norm{\\nabla f(y_k)}\\leq G_k+LE_k$, a closed error recursion is\n\\begin{equation}\n E_{k+1}=(q+La_k)E_k+a_kG_k\n +\\eta_k(\\delta_k+\\epsilon_k).\n \\label{eq:drift-tube}\n\\end{equation}\n\n\\begin{proposition}[Finite robustness under step and gradient drift]\nWith $E_0$ bounding the initial displacement,\n\\eqref{eq:drift-tube} bounds $\\norm{y_k-x_k}$ for every finite $k$.\nIf measurements are within $\\nu_k$ of $f_k(y_k)$ and\n$|f_k(y_k)-f(y_k)|\\leq\\beta_k$, Theorem~\\ref{thm:law-transfer}\nholds after replacing $\\nu_k$ by $\\nu_k+\\beta_k$.\n\\end{proposition}\n\n\\begin{proof}\nThe preceding residual decomposition and the Lipschitz estimate give\n\\eqref{eq:drift-tube} by the same induction as\nTheorem~\\ref{thm:trajectory-tube}. The value claim is the triangle\ninequality followed by Theorem~\\ref{thm:law-transfer}.\n\\end{proof}\n\nNo convergence conclusion is drawn when $q+La_k$ repeatedly exceeds one.\nThe formula remains a correct finite bound. A controller may shorten the\nreference block or choose another certified reference model when the\nallowance becomes too wide. It may not set the allowance to zero merely\nbecause the step changes appear small on a plot.\n\nObjective drift has two distinct effects. Gradient drift moves the\ntrajectory; value drift changes what the displayed loss measures. A\nconstant offset in $f_k$ has no gradient effect but can arbitrarily alter\nraw loss differences across rounds. This is why comparisons between\nsuccessive models in recursive learning must use a common evaluation\ncriterion or explicitly account for its change.\n\n\\section{Local restarts prevent irrelevant error accumulation}\n\nSuppose a long process is tested at times $s=0,1,\\ldots$. For the test\nbeginning at $s$, create a fresh exact reference trajectory with\n$x_0^{(s)}=y_s$, the same reference objective, and the declared step.\nThen its initial discrepancy is zero, regardless of errors before $s$.\n\n\\begin{corollary}[Blockwise certification]\nA finite-horizon law valid for every initial point can be checked on each\nblock using only the residual and measurement bounds inside that block.\nErrors before its start do not enter the trajectory allowance.\n\\end{corollary}\n\n\\begin{proof}\nThe law is valid from the new initial point. Apply\nTheorem~\\ref{thm:law-transfer} with $E_0=0$ to the residual sequence of\nthat block.\n\\end{proof}\n\nThis is particularly useful for the four-state compensation certificate.\nIt avoids an artificial error bound growing with the total age of a\ntraining run. It does not erase actual measurement error or drift within\nthe four queried states.\n\n\\section{Persistence of a strict concave run under adversarial error}\n\nThe long-run construction is a theorem about exact descent. Its strict\nmargin also supplies a nonzero tolerance to arbitrary bounded errors.\nThe following bound is intentionally uniform; it does not use the special\nprojection geometry.\n\n\\begin{proposition}[A robust finite-run margin]\n\\label{prop:robust-long-run}\nLet $f$ be convex and $1$-smooth, let $0<\\eta<2$, and suppose its exact\norbit from $x_0$ has $\\norm{\\nabla f(x_0)}\\leq2$ and\n\\[\n f_k-2f_{k+1}+f_{k+2}\\leq-\\gamma\n \\quad(0\\leq k<m),\\qquad 0<\\gamma\\leq1.\n\\]\nPut $M=m+1$. Start $y_0=x_0$, use additive gradient errors at most\n$\\epsilon$, and measure each loss with error at most $\\nu$.\nIf\n\\begin{equation}\n \\epsilon\\leq\\frac\\gamma{64M},\\qquad\n \\nu\\leq\\frac\\gamma{32},\n \\label{eq:run-noise-tolerance}\n\\end{equation}\nthen every measured second difference in the same interval is at most\n$-\\gamma/2$.\n\\end{proposition}\n\n\\begin{proof}\nFor $0\\leq i\\leq M$, the nonexpansive tube gives\n$E_i\\leq2M\\epsilon$. Gradient norms on the exact orbit are at most two,\nso each value error is at most\n$\\nu+4M\\epsilon+2M^2\\epsilon^2$.\nA second difference has coefficient absolute sum four, hence its error\nis at most\n\\[\n 4\\nu+16M\\epsilon+8M^2\\epsilon^2\n \\leq\\frac\\gamma8+\\frac\\gamma4+\\frac{\\gamma^2}{512}\n \\leq\\frac{193}{512}\\gamma<\\frac\\gamma2.\n\\]\nAdding this to the exact upper bound $-\\gamma$ proves the result.\n\\end{proof}\n\nFor the analytic witness, one may take\n$\\gamma=3/[128M^2]$, and its initial gradient is bounded by two.\nThus the explicit sufficient gradient-error tolerance is\n$3/[8192M^3]$. This generic adversarial tolerance is much smaller than\nthe Gaussian smoothing width proved earlier. The two perturbations have\ndifferent structure: an arbitrary error sequence need not preserve the\nshared quadratic trace or have exponentially rare chamber crossings.\n\nThe proposition certifies strict curvature, not positivity of each\nmeasured decrement. If positive measured improvement is also required,\napply the two-point version of Theorem~\\ref{thm:law-transfer} and use a\nseparate lower bound on the smallest decrement.\n\n\\section{From probabilistic errors to deterministic envelopes}\n\nSometimes an oracle supplies a tail model instead of an absolute bound.\nFor example, suppose each coordinate $e_{k,j}$ has, conditionally on the\npast, the sub-Gaussian moment bound with a nonnegative\n$\\mathcal H_k$-measurable scale $\\sigma_k$,\n\\[\n \\E[\\exp(\\lambda e_{k,j})\\mid\\mathcal H_k]\n \\leq\\exp(\\lambda^2\\sigma_k^2/2)\n \\quad(\\lambda\\in\\R).\n\\]\nNo independence between the coordinates is required for the following\nconservative envelope. For any $\\alpha_k\\in(0,1)$, Chernoff's argument\nand a union bound give\n\\begin{equation}\n \\Prob\\left\\{\\norm{e_k}>\n \\sigma_k\\sqrt{2d\\log(2d/\\alpha_k)}\n \\,\\middle|\\,\\mathcal H_k\\right\\}\\leq\\alpha_k.\n \\label{eq:subg-envelope}\n\\end{equation}\nIndeed, optimize\n$\\exp(-\\lambda u+\\lambda^2\\sigma_k^2/2)$ over $\\lambda>0$ for\neach upper tail, repeat for the lower tail, and union-bound the $2d$\nevents. If every coordinate is at most\n$\\sigma_k\\sqrt{2\\log(2d/\\alpha_k)}$ in absolute value, its Euclidean\nnorm is at most the displayed threshold.\nWhen $\\sigma_k=0$, the conditional moment bound forces every coordinate\nto vanish almost surely, and the same conclusion holds directly.\n\nChoosing a deterministic summable sequence with\n$\\sum_k\\alpha_k\\leq\\alpha$ gives\none event of probability at least $1-\\alpha$ on which all these\ndeterministic error envelopes hold. On that event every tube and every\nfinite-law certificate above is valid simultaneously. This converts a\ndeclared tail assumption into a pathwise guarantee; it does not prove\nthat the assumption holds for a particular optimizer.\n\n\\section{What robustness can and cannot establish}\n\nThere is no general reason for an inexact algorithm to retain the exact\nloss-curvature sign at every step. The literature already contains\nrelative-error impossibility results even for one-dimensional quadratics\n\\cite{hieu2025}. A useful robust theorem therefore carries an explicit\ndefect allowance or uses a strict margin. Replacing a signed inequality\nby a tolerance-aware inequality is the mathematically appropriate change.\n\nThese bounds have a practical interpretation for recursive learning.\nA proof-producing component may certify a training subroutine in a\ndeclared smooth convex model. A monitor can then compare real telemetry\nwith that certificate and its error budget. A violation can trigger a\nmodel recheck or a different subroutine. Acceptance of a new model on an\nexternal task still requires a separate evaluation guarantee, developed\nin Chapter~\\ref{ch:recursive}. Optimization consistency and task\nimprovement answer different questions, and both are needed when an\nalgorithm proposes changes to its own learning process.\n"}
{"format": "latex", "id": "08_information", "sha256": "59c40fa61dda5bc2fdb106f1bfc0cfd7d3386edf7fe3f4738abe167d652b39fb", "source_path": "research/chapters/08_information.tex", "text": "\\chapter{Information Limits of Learning-Curve Decisions}\n\\label{ch:information}\n\nThe shape of a measured optimization curve is evidence about one observed\ntrajectory. It is not a description of the objective away from that\ntrajectory. This distinction becomes operational when a controller decides\nwhether to continue training, abandon a candidate, or replace its own search\nprocedure. The controller must compare a future benefit with a resource\ncost, although the future benefit is usually not directly observed.\n\nThis chapter makes that information limitation quantitative. An arbitrarily\nlong, exact first-order gradient-descent transcript can leave the entire\nclassical condition-number uncertainty in the remaining objective gap.\nEven real analyticity and access to finitely many higher derivatives do not\nmake the next unobserved loss identifiable. These are constructive\nstatements with explicit functions. They complement the positive\ncertificate results: a valid interval for future performance can be\ncomputed under a declared model, while an exact point forecast generally\nrequires additional information.\n\n\\section{What a transcript determines}\n\nFix $0<\\mu<L$ and a step $\\eta>0$. A first-order transcript through time $t$\nis\n\\begin{equation}\n \\mathcal T_t=\\{(x_i,f_i,g_i):0\\leq i\\leq t\\},\n \\qquad x_{i+1}=x_i-\\eta g_i\\quad(0\\leq i<t).\n \\label{eq:info-transcript}\n\\end{equation}\nAn objective is compatible with this transcript if its value and gradient\nat every $x_i$ equal the recorded $f_i$ and $g_i$. The parameters $\\mu$ and\n$L$ are part of the information available to the controller. An unknown\nadditive constant in the objective is irrelevant to all improvement\nquantities considered here.\n\nThe remaining improvement at a point $x$ is\n\\begin{equation}\n R_f(x)=f(x)-\\min_y f(y).\n \\label{eq:info-gap}\n\\end{equation}\nStrong convexity implies that the minimum exists and is unique for the\nglobally defined objectives used below. It also provides a universal\nupper bound for this remaining improvement. Smoothness supplies a lower\nbound.\n\n\\begin{lemma}[Gradient bounds for the remaining gap]\n\\label{lem:info-gap}\nIf $f$ is globally $\\mu$-strongly convex and $L$-smooth, then for every $x$,\n\\begin{equation}\n \\frac{\\norm{\\nabla f(x)}^2}{2L}\n \\leq R_f(x)\n \\leq \\frac{\\norm{\\nabla f(x)}^2}{2\\mu}.\n \\label{eq:info-gap-sandwich}\n\\end{equation}\n\\end{lemma}\n\\begin{proof}\nWrite $g=\\nabla f(x)$. Smoothness evaluated at $x-g/L$ gives\n\\[\n f(x-g/L)\\leq f(x)-\\frac{\\norm g^2}{2L}.\n\\]\nThe global minimum is no larger than the left-hand side. This proves the\nfirst inequality. Strong convexity gives, for every $y$,\n\\[\n f(y)\\geq f(x)+\\ip g{y-x}+\\frac\\mu2\\norm{y-x}^2\n \\geq f(x)-\\frac{\\norm g^2}{2\\mu},\n\\]\nwhere the last expression is the minimum of the displayed quadratic in\n$y$. Taking the infimum over $y$ proves the upper bound.\n\\end{proof}\n\nThe ratio of the two bounds is $\\kappa=L/\\mu$. The next theorem shows that\na finite history, however long, need not reduce this uncertainty at all.\nThe statement concerns a worst-case transcript; many particular\ntranscripts contain additional useful information.\n\n\\section{A sharp ambiguity after any finite history}\n\n\\begin{theorem}[A common transcript with the full gap interval]\n\\label{thm:info-full-interval}\nSuppose $0<\\eta\\mu<1$ and $0<\\eta L<2$. Let $t\\geq0$ be an integer,\n$v>0$, and $C\\in\\R$. Put $q=1-\\eta\\mu$ and\n\\begin{equation}\n x_i=\\frac v\\mu\\bigl(q^{i-t}-1\\bigr),\\qquad 0\\leq i\\leq t.\n \\label{eq:info-prefix}\n\\end{equation}\nThere is a one-parameter family of globally $\\mu$-strongly convex,\n$L$-smooth $C^1$ objectives, all producing exactly the same transcript\n\\eqref{eq:info-transcript}, for which $x_t=0$, $g_t=v$, and the set of\npossible remaining gaps is exactly\n\\begin{equation}\n \\left[\\frac{v^2}{2L},\\frac{v^2}{2\\mu}\\right].\n \\label{eq:info-exact-interval}\n\\end{equation}\nThe family can be chosen so that $\\mu$ and $L$ are the optimal global\nstrong-convexity and smoothness constants for every member.\n\\end{theorem}\n\n\\begin{proof}\nFor $a\\in[\\mu,L]$, begin with\n\\begin{equation}\n f_a(x)=\n \\begin{cases}\n C+vx+\\dfrac a2x^2,&x\\leq0,\\\\[3pt]\n C+vx+\\dfrac\\mu2x^2,&x\\geq0.\n \\end{cases}\n \\label{eq:info-piecewise}\n\\end{equation}\nThe two values and first derivatives agree at zero. The derivative is\ncontinuous and piecewise affine, with slopes in $[\\mu,L]$. Consequently,\nfor $x<y$,\n\\[\n \\mu(y-x)\\leq f_a'(y)-f_a'(x)\\leq L(y-x).\n\\]\nIntegration proves strong convexity, while the upper inequality proves\nglobal gradient Lipschitz continuity.\n\nAll points in \\eqref{eq:info-prefix} are nonnegative, with equality only\nat $i=t$. On this half-line every member of the family has the same\nvalue and derivative. Moreover,\n\\begin{align*}\n x_i-\\eta(v+\\mu x_i)\n &=q x_i-\\eta v\\\\\n &=\\frac v\\mu\\bigl(q^{i+1-t}-q\\bigr)\n -\\frac v\\mu(1-q)\n =x_{i+1}.\n\\end{align*}\nThis proves that the complete first-order transcript is shared.\n\nThe unique minimizer of $f_a$ is $-v/a$, and\n\\[\n f_a(0)-f_a(-v/a)=\\frac{v^2}{2a}.\n\\]\nAs $a$ runs continuously from $\\mu$ to $L$, these gaps fill precisely\n\\eqref{eq:info-exact-interval}. Lemma~\\ref{lem:info-gap} shows that no\ncompatible objective in the declared class can have a gap outside this\ninterval. Thus the interval is also the exact identified set over the\nwhole class, not merely a list of two examples.\n\nTo make both class constants optimal, choose any $B>x_0$ and replace the\ncommon right-hand branch for $x\\geq B$ by\n\\begin{equation}\n f_a(B)+f_a'(B)(x-B)+\\frac L2(x-B)^2.\n \\label{eq:info-tail}\n\\end{equation}\nThis common modification preserves the transcript and the minimizer.\nThere remains a nonempty interval of curvature $\\mu$ inside $(0,B)$,\nand there is now a nonempty interval of curvature $L$ in $(B,\\infty)$.\nNo stronger global lower curvature or smaller global upper curvature is\npossible.\n\\end{proof}\n\nThe initial point in this construction depends on $t$. In particular,\nthe theorem does not maintain a separately prescribed upper bound on\ninitial distance while the history length grows. If such a distance\nbound is available, it should be included among the compatibility\nconstraints. The claim here is exact for the information explicitly\nstated in the theorem.\n\n\\begin{corollary}[Optimal point-estimation uncertainty]\n\\label{cor:info-minimax}\nFor the transcript in Theorem~\\ref{thm:info-full-interval}, every real\npoint estimate $\\widehat R$ of the remaining gap has worst-case absolute\nerror at least\n\\begin{equation}\n \\frac{v^2}{4}\\left(\\frac1\\mu-\\frac1L\\right).\n \\label{eq:info-minimax-absolute}\n\\end{equation}\nThe midpoint of \\eqref{eq:info-exact-interval} attains this error.\nAmong positive estimates, the smallest possible worst-case symmetric\nmultiplicative factor is $\\sqrt{L/\\mu}$, attained by\n\\begin{equation}\n \\widehat R_{\\mathrm{mult}}=\\frac{v^2}{2\\sqrt{\\mu L}}.\n \\label{eq:info-minimax-multiplicative}\n\\end{equation}\n\\end{corollary}\n\\begin{proof}\nWrite $A=v^2/(2L)$ and $B=v^2/(2\\mu)$. At least one of\n$|\\widehat R-A|$ and $|\\widehat R-B|$ is at least $(B-A)/2$; otherwise\nthe triangle inequality would give $B-A<B-A$. The midpoint has this\ndistance from both endpoints and no larger distance from any interior\npoint. For a positive estimate $u$, the worst multiplicative factor is\nat least $\\max\\{u/A,B/u\\}$. The product of these two expressions is\n$B/A$, so their maximum is at least $\\sqrt{B/A}$. Equality holds at\n$u=\\sqrt{AB}$.\n\\end{proof}\n\nThis result identifies the limitation of a point forecast, rather than\narguing against forecasting. A controller can report an interval, use\nadditional structural information, or pay for further observations. It\ncannot infer information that the transcript does not contain.\n\n\\section{Finite-horizon benefit and stopping regret}\n\nThe same family provides exact continuation curves. It therefore\nsupports a decision-theoretic statement about compute allocation,\nwithout replacing the finite-horizon question by an asymptotic one.\n\n\\begin{proposition}[Exact continuation on indistinguishable objectives]\n\\label{prop:info-continuation}\nFor every $a\\in[\\mu,L]$, the next $n\\geq1$ gradient steps from the common\nterminal point $x_t=0$ remain in $(-\\infty,0]$ and satisfy\n\\begin{align}\n x_{t+j}^{(a)}&=\\frac va\\bigl((1-\\eta a)^j-1\\bigr),\\qquad j\\geq0,\n \\label{eq:info-future-points}\\\\\n G_n(a)&:=f_a(0)-f_a(x_{t+n}^{(a)})\n =\\frac{v^2}{2a}\\bigl[1-(1-\\eta a)^{2n}\\bigr].\n \\label{eq:info-future-gain}\n\\end{align}\nFor fixed $n$, $G_n(a)$ is strictly decreasing on $0<a<2/\\eta$.\n\\end{proposition}\n\n\\begin{proof}\nThe affine iteration on the left branch is\n$x^+=(1-\\eta a)x-\\eta v$. Its closed form from zero is\n\\eqref{eq:info-future-points}. Since $|1-\\eta a|<1$, every power in\nthis expression is at most one, so the points are indeed nonpositive.\nThus the branch calculation is valid for the entire future orbit.\nCompleting the square in \\eqref{eq:info-piecewise} gives\n\\eqref{eq:info-future-gain}.\n\nFor monotonicity, put $h=\\eta a$ and $z=1-h$. Apart from the positive\nconstant $v^2\\eta/2$, the expression is\n$b_n(h)=[1-(1-h)^{2n}]/h$. Its derivative has the sign of\n\\[\n 2nhz^{2n-1}-1+z^{2n}.\n\\]\nIf $0<h\\leq1$, then $0\\leq z<1$ and\n\\[\n 1-z^{2n}=h\\sum_{j=0}^{2n-1}z^j\n >2nhz^{2n-1}.\n\\]\nThe inequality is strict because $n\\geq1$ and $z<1$. If $1<h<2$,\nthen $-1<z<0$, so $2nhz^{2n-1}<0$ and $-1+z^{2n}<0$ separately.\nThe derivative is therefore negative throughout the claimed interval.\n\\end{proof}\n\nSuppose a controller has exactly two actions: stop now, with utility\nzero, or purchase $n$ further steps at cost $c$, with utility\n$G_n(a)-c$. The cost is measured in the same declared units as objective\nimprovement. The controller may use the entire transcript and the\ncorrect constants $\\mu,L,\\eta$.\n\n\\begin{theorem}[Unavoidable regret of a transcript-only stopping choice]\n\\label{thm:info-stopping}\nIf $G_n(L)<c<G_n(\\mu)$, no stopping rule based on the common transcript\ncan be optimal on every objective compatible with it. Write\n\\[\n A=c-G_n(L)>0,\\qquad B=G_n(\\mu)-c>0.\n\\]\nEvery deterministic rule has worst-case regret at least $\\min\\{A,B\\}$.\nEvery randomized rule has worst-case expected regret at least\n\\begin{equation}\n \\frac{AB}{A+B}.\n \\label{eq:info-regret}\n\\end{equation}\nThe latter bound is exact for the decision problem restricted to the\ntwo endpoint objectives.\n\\end{theorem}\n\\begin{proof}\nThe transcript is identical under both endpoint objectives. A randomized\nrule must therefore continue with the same probability $p$ in both\nworlds. On the $L$ endpoint the optimal action is to stop and the\nexpected regret is $pA$. On the $\\mu$ endpoint the optimal action is to\ncontinue and the expected regret is $(1-p)B$. Their maximum is minimized\nwhen $pA=(1-p)B$, namely at $p=B/(A+B)$, giving\n\\eqref{eq:info-regret}. A deterministic rule has $p=0$ or $p=1$, whose\nworst regrets are $B$ and $A$, respectively.\n\\end{proof}\n\nThe theorem concerns the stated decision before new information arrives.\nA sequential strategy can purchase observations and revise its decision;\nthe cost of those observations is then part of a different decision\nproblem. This distinction is useful for recursive learning systems:\nuncertainty should have an explicit acquisition cost rather than be\nsilently replaced by a curve-fitting assumption.\n\n\\begin{example}[The same six observations, opposite continuation choices]\n\\label{ex:info-stopping}\nTake $\\mu=1/16$, $L=1$, $\\eta=1$, $v=1/4$, and $t=5$. The shared\ninitial point is $x_0=1156804/759375$, approximately $1.52336$, and\nthe sixth queried point is zero. The complete observed values and\ngradients are identical for $f_{1/16}$ and $f_1$. Their remaining gaps\nare $1/2$ and $1/32$. For $n=4$ further steps,\n\\[\n G_4(1/16)=\\frac{1732076671}{8589934592}\\approx0.2016403,\n \\qquad G_4(1)=\\frac1{32}=0.03125.\n\\]\nAt a cost of $1/10$, continuation has positive net value on the first\nobjective and negative net value on the second. The signs and rational\nformulas are checked by the accompanying exact-arithmetic program.\n\\end{example}\n\n\\section{Real analyticity does not identify the next loss}\n\nThe preceding family is $C^1$ and piecewise quadratic. One might ask\nwhether analyticity removes the ambiguity. Analytic functions that\nagree on an interval agree everywhere on a connected domain, but a\nfinite transcript is not an interval. A finite number of derivative\nobservations leaves a constructive analytic freedom.\n\nFor an integer $r\\geq1$, the order-$r$ jet of a one-dimensional function\nat $x$ means $(f(x),f'(x),\\ldots,f^{(r)}(x))$. The next result allows\nan arbitrarily large but finite $r$.\n\n\\begin{lemma}[An explicit global derivative envelope]\n\\label{lem:info-gaussian-envelope}\nLet $P$ be a real polynomial and put\n\\begin{align}\n \\phi(x)&=e^{-x^2}P(x),\\\\\n Q_P(x)&=P''(x)-4xP'(x)+(4x^2-2)P(x)=\\sum_j q_jx^j,\\\\\n M(P)&=\\sum_j |q_j|\\,\\bigl\\lceil j/2\\bigr\\rceil!.\n \\label{eq:info-M}\n\\end{align}\nThen $\\phi''(x)=e^{-x^2}Q_P(x)$ and\n$\\sup_{x\\in\\R}|\\phi''(x)|\\leq M(P)$. If $P$ is nonzero, then\n$M(P)>0$. Rational coefficients in $P$ produce a rational $M(P)$.\n\\end{lemma}\n\\begin{proof}\nTwo differentiations give the displayed polynomial $Q_P$. To bound a\nmonomial, set $m=\\lceil j/2\\rceil$. For $|x|\\leq1$,\n$|x|^je^{-x^2}\\leq1\\leq m!$. For $|x|\\geq1$, use\n$|x|^j\\leq |x|^{2m}$ and the exponential-series term\n$e^{x^2}\\geq x^{2m}/m!$ to obtain the same bound. Summing the absolute\ncoefficient bounds proves the envelope. If $M(P)=0$, then\n$\\phi''=0$, so $\\phi$ is affine. But $e^{-x^2}P(x)$ tends to zero at\nboth ends of the real line, so this affine function is zero. That would\nforce $P=0$, a contradiction.\n\\end{proof}\n\n\\begin{theorem}[Exact finite-jet indistinguishability in the analytic class]\n\\label{thm:info-analytic}\nLet $0<\\mu<L$, $0<\\eta L<2$, and choose $a\\in(\\mu,L)$ with\n$\\eta a\\ne1$. For any integers $t\\geq0$ and $r\\geq1$, there are two\nglobally $\\mu$-strongly convex, $L$-smooth real-analytic objectives\nwhose order-$r$ jets agree at every point of a common GD transcript\nthrough time $t$, but whose losses at the common next iterate differ.\nEvery parameter and a valid global Hessian bound can be given explicitly.\n\\end{theorem}\n\n\\begin{proof}\nPut $q=1-\\eta a$, choose $x_0\\ne0$, and let $x_i=q^ix_0$. The\nassumptions give $-1<q<1$ and $q\\ne0$, so these points are pairwise\ndistinct. Define\n\\begin{equation}\n P(x)=\\prod_{i=0}^t(x-x_i)^{2r+2},\\qquad\n \\phi(x)=e^{-x^2}P(x),\\qquad\n \\eps_0=\\frac{\\min\\{a-\\mu,L-a\\}}{2M(P)}.\n \\label{eq:info-analytic-perturbation}\n\\end{equation}\nFor any $0<\\eps\\leq\\eps_0$, set\n\\begin{equation}\n f_\\pm(x)=\\frac a2x^2\\pm\\eps\\phi(x).\n \\label{eq:info-analytic-pair}\n\\end{equation}\nThese functions extend to entire functions of a complex variable.\nLemma~\\ref{lem:info-gaussian-envelope} yields\n\\[\n a-\\eps M(P)\\leq f_\\pm''(x)\\leq a+\\eps M(P),\n\\]\nand the chosen $\\eps$ places this interval strictly inside $[\\mu,L]$.\nThus both objectives satisfy the required global real-domain curvature\nbounds.\n\nAt each observed point $x_i$, $P$ has a zero of order $2r+2$.\nMultiplication by the nonvanishing analytic Gaussian preserves that\norder. Hence $\\phi^{(j)}(x_i)=0$ for every $j\\leq2r+1$, in particular\nfor every derivative available to the controller. The values and\ngradients of $f_+$ and $f_-$ at $x_i$ therefore agree with those of\n$ax^2/2$. Induction gives the common trajectory $x_i=q^ix_0$ through\ntime $t+1$.\n\nThe next iterate $x_{t+1}$ is distinct from every factor root of $P$.\nAll powers in \\eqref{eq:info-analytic-perturbation} are even, so\n$\\phi(x_{t+1})>0$. Consequently,\n\\begin{equation}\n f_+(x_{t+1})-f_-(x_{t+1})\n =2\\eps\\phi(x_{t+1})>0.\n \\label{eq:info-next-loss-separation}\n\\end{equation}\nThe two next loss decrements differ by exactly the same positive\nquantity, with the reverse sign. If the chosen $x_i$ and $a,\\mu,L$\nare rational, the bound $M(P)$ and a permitted $\\eps$ are rational\nand can be certified by exact polynomial arithmetic.\n\\end{proof}\n\nThis theorem establishes nonidentification, not an order-one lower\nbound on prediction error. Its explicit separation may be small when\nthe number of observations or derivative orders is large. It does,\nhowever, rule out an exact universal next-loss predictor based only on\nfinitely many jets, even within the analytic strongly convex class.\nThe same construction works for any prescribed finite set of distinct\nquery points: the product simply uses that set. The GD specialization\nensures that the indistinguishable observations form an actual\noptimization history.\n\nA finite match to a linear recurrence has the same limitation. Such a\nmatch can validate the recorded values against the recurrence, but it\ncannot establish that an unknown objective belongs globally to the\nfamily from which the recurrence was derived. Model membership or a\nfuture branch certificate remains a separate premise.\n\n\\section{A finite budget for accelerating decrements}\n\nThere is also a general restriction that requires no convexity,\nsmoothness, or particular algorithm. If an objective is bounded below,\nthen increasing positive decrements cannot continue forever. A useful\nversion keeps track of the amount of negative discrete curvature.\n\n\\begin{theorem}[Curvature expenditure along a finite run]\n\\label{thm:info-curvature-budget}\nLet $f_k$ be any sequence bounded below by $f_*$, let\n$D_k=f_k-f_{k+1}$, and suppose, for some $m\\geq1$,\n\\[\n D_{k+j+1}-D_{k+j}=a_j\\geq0,\\qquad 0\\leq j<m.\n\\]\nThen\n\\begin{equation}\n (m+1)D_k+\\sum_{j=0}^{m-1}(m-j)a_j\n =f_k-f_{k+m+1}\\leq f_k-f_*.\n \\label{eq:info-curvature-budget}\n\\end{equation}\nIn particular, if $D_k>0$, a nondecreasing positive tail is impossible.\nIf every $a_j\\geq\\delta>0$, and $R=f_k-f_*$, then with $n=m+1$,\n\\begin{equation}\n nD_k+\\frac\\delta2n(n-1)\\leq R,\n \\qquad\n n\\leq\\frac{-(D_k-\\delta/2)\n +\\sqrt{(D_k-\\delta/2)^2+2\\delta R}}{\\delta}.\n \\label{eq:info-run-bound}\n\\end{equation}\n\\end{theorem}\n\\begin{proof}\nFor $0\\leq\\ell\\leq m$,\n$D_{k+\\ell}=D_k+\\sum_{j=0}^{\\ell-1}a_j$. Summing over $\\ell$ counts\n$a_j$ exactly $m-j$ times. The sum of the decrements telescopes to\n$f_k-f_{k+m+1}$, proving \\eqref{eq:info-curvature-budget}. On an\ninfinite nondecreasing tail with $D_k>0$, its first $n$ decrements\nsum to at least $nD_k$, contradicting the finite upper bound $R$ for\nlarge $n$. If $a_j\\geq\\delta$, their weighted sum is at least\n$\\delta m(m+1)/2$. Solving the resulting quadratic inequality in\n$n\\geq0$ gives \\eqref{eq:info-run-bound}.\n\\end{proof}\n\nFor a strongly convex objective, Lemma~\\ref{lem:info-gap} permits the\nobservable substitute $R\\leq\\norm{g_k}^2/(2\\mu)$ in the upper bound.\nThe long concave runs constructed earlier are consistent with this\nbudget: both their initial decrement and their curvature margin shrink\nwith the prescribed run length. No theorem in this manuscript asserts\nan infinite accelerating tail of positive improvement on a lower-bounded\nloss.\n\n\\section{Consequences for a learning controller}\n\nThe results distinguish three pieces of information a controller might\npossess. A transcript supplies observations. A function class supplies\nconstraints on compatible continuations. A model certificate supplies\nadditional structure, such as a valid branch of a declared recurrence.\nOnly the latter two can reduce future ambiguity beyond what the\nobservations themselves determine.\n\nThe sharp interval theorem shows why a small recent decrement need not\njustify stopping when the condition number is large. The analytic\nconstruction shows why increasing regularity alone does not justify a\nunique extrapolation. The stopping-regret theorem identifies a region\nwhere further information has decision value. Finally, the curvature\nbudget prevents finite examples of increasing progress from being\ninterpreted as evidence for an unlimited increase in positive gains.\n\nFor recursive self-improvement, these statements support an explicit\nseparation between candidate generation and candidate acceptance. A\ngenerator may use a learned extrapolator, a formal certificate, or an\naggressive heuristic to propose its next action. The acceptance rule\nshould certify the proposition it actually needs: improvement in a\ncommon objective at an explicit resource cost. The following chapter\ngives a complete rule of that kind, including its statistical and\nresource assumptions.\n"}
{"format": "latex", "id": "09_recursive", "sha256": "732dafa67f0c63335cadbe489be245b600431a6be90af233c5cefb84538b81c9", "source_path": "research/chapters/09_recursive.tex", "text": "\\chapter{Certified Recursive Learning and Resource Allocation}\n\\label{ch:recursive}\n\nAn optimization theorem becomes useful to a recursive learning system\nonly after its decision contract is specified. A training objective may\nchange when a model, optimizer, or representation changes. Improvement\nin that objective is then not automatically improvement in a common\ntask. This chapter defines a common evaluation objective and proves\nguarantees for a controller whose proposals may depend arbitrarily on\nits previous outcomes. The proposal mechanism can itself be replaced\nby an accepted candidate.\n\nThe resulting guarantees have three components. A validation rule\ncontrols false improvement claims across all proposal rounds and all\nchosen sample stopping times. An acceptance rule provides cumulative\nimprovement and a finite comparison with every candidate actually\nexamined. A resource ledger charges all attempted work and prevents the\ncontroller from treating failed experiments as free. These results\napply to a declared model of recursive learning; they do not establish\na hardware speedup or a universal solution to recursive\nself-improvement.\n\n\\section{A common objective across changing implementations}\n\nLet $\\mathcal S$ be a measurable space of candidate systems. A state\n$s\\in\\mathcal S$ may specify a predictor, its architecture and\nparameters, an optimizer, or a complete procedure including its own\nproposal generator. Let $Z$ describe a fresh task instance and any\nevaluation randomness. A fixed loss function\n\\begin{equation}\n \\ell:\\mathcal S\\times\\mathcal Z\\longrightarrow[0,1],\\qquad\n R(s)=\\E_{Z\\sim P}\\ell(s,Z)\n \\label{eq:rsi-risk}\n\\end{equation}\ndefines the common risk. The bounded range is a theorem assumption.\nKnown bounded ranges can be rescaled; clipping an unbounded loss would\ninstead change the objective and must be declared as such.\n\nAt proposal round $t\\geq1$, the incumbent is $s_{t-1}$ and the proposed\nreplacement is $a_t$. Both may depend on all earlier proposals,\ntraining runs, validation observations, accepted changes, and internal\nrandomness. They are fixed before the current validation observations\nare collected. The true improvement for this comparison is\n\\begin{equation}\n \\Delta_t=R(s_{t-1})-R(a_t).\n \\label{eq:rsi-delta}\n\\end{equation}\nThe same fresh task is used for both systems, giving paired observations\n\\begin{equation}\n Y_{t,j}=\\ell(s_{t-1},Z_{t,j})-\\ell(a_t,Z_{t,j})\\in[-1,1].\n \\label{eq:rsi-paired}\n\\end{equation}\n\n\\begin{assumption}[Evaluation contract]\n\\label{ass:rsi-evaluation}\nConditional on the history before round $t$, the two compared systems\nare fixed. For every $j$, the conditional mean of $Y_{t,j}$ given all\ninformation available before that observation equals $\\Delta_t$.\nThe observation remains in $[-1,1]$. Fresh independent draws from the\nfixed distribution $P$ satisfy these conditions.\n\\end{assumption}\n\nThe conditional-mean formulation is slightly more general than\nindependent sampling. It is also the precise point at which evaluator\nfeedback matters. A candidate changed in response to its current\nvalidation observations no longer satisfies the stated contract for\nthe original $\\Delta_t$. It can be entered as a new proposal and\nevaluated under a fresh comparison.\n\nThe problem of adaptive reuse of data has a substantial literature,\nincluding reusable-holdout methods~\\cite{dwork2015}. The rule below\nuses the simpler fresh-evaluation contract. It is not a claim that\nordinary repeated reuse of a fixed public benchmark remains valid.\n\n\\section{One probability guarantee for every future comparison}\n\nThe concentration tools here are classical. They are proved directly\nso that the exact accounting used by the controller can be audited.\nFor broader and sharper confidence-sequence constructions, see\nHoward et al.~\\cite{howard2021}; the bounded-variable exponential\nmethod originates with Hoeffding~\\cite{hoeffding1963}.\n\n\\begin{lemma}[Conditional bounded-variable concentration]\n\\label{lem:rsi-hoeffding}\nIf $Y\\in[-1,1]$ and its conditional mean is $m$, then for every real\n$\\lambda$,\n\\begin{equation}\n \\E[\\exp\\{\\lambda(Y-m)\\}\\mid\\mathcal F]\n \\leq \\exp(\\lambda^2/2).\n \\label{eq:rsi-hoeffding-mgf}\n\\end{equation}\nConsequently, under Assumption~\\ref{ass:rsi-evaluation}, for each\nfixed $n$ and $r>0$,\n\\begin{equation}\n \\Prob\\left(\\left|\\frac1n\\sum_{j=1}^nY_{t,j}-\\Delta_t\\right|>r\n \\,\\middle|\\,\\mathcal F_{t,0}\\right)\n \\leq2e^{-nr^2/2}.\n \\label{eq:rsi-hoeffding-tail}\n\\end{equation}\n\\end{lemma}\n\\begin{proof}\nFor a bounded variable, the second derivative of its log moment\ngenerating function is its variance under the exponentially tilted\ndistribution. Every distribution supported on an interval of length\ntwo has variance at most one: the mean minimizes expected squared\ndistance, and squared distance from the interval midpoint is at most\none. Thus the centered log moment generating function $K$ satisfies\n$K''(\\lambda)\\leq1$, $K(0)=0$, and $K'(0)=0$. Integrating twice gives\n$K(\\lambda)\\leq\\lambda^2/2$. The same argument applies conditionally,\nproving \\eqref{eq:rsi-hoeffding-mgf}.\n\nIterating conditional expectations bounds the moment generating\nfunction of $\\sum_{j=1}^n(Y_{t,j}-\\Delta_t)$ by\n$e^{n\\lambda^2/2}$. Markov's inequality and the choice $\\lambda=r$\nbound its upper tail at $nr$ by $e^{-nr^2/2}$. Apply the same argument\nto the negative sum and add the two tail probabilities.\n\\end{proof}\n\nFix one error budget $\\delta\\in(0,1)$ for the entire run, not a new\nunaccounted budget for each proposed improvement. Define\n\\begin{equation}\n r_{t,n}(\\delta)=\n \\sqrt{\\frac2n\\log\\frac{2t(t+1)n(n+1)}{\\delta}},\n \\qquad t,n\\geq1.\n \\label{eq:rsi-radius}\n\\end{equation}\n\n\\begin{theorem}[Simultaneous validity under adaptive proposals and stopping]\n\\label{thm:rsi-simultaneous}\nUnder Assumption~\\ref{ass:rsi-evaluation}, with probability at least\n$1-\\delta$, the event\n\\begin{equation}\n \\mathcal E=\\left\\{\n \\left|\\widehat\\Delta_{t,n}-\\Delta_t\\right|\n \\leq r_{t,n}(\\delta)\\text{ for every }t,n\\geq1\n \\right\\},\\qquad\n \\widehat\\Delta_{t,n}=\\frac1n\\sum_{j=1}^nY_{t,j},\n \\label{eq:rsi-simultaneous}\n\\end{equation}\nholds. In particular, it holds at sample sizes chosen by any stopping\nrule based on the observations collected so far.\n\\end{theorem}\n\\begin{proof}\nSubstitution in \\eqref{eq:rsi-hoeffding-tail} makes the failure\nprobability for a particular pair $(t,n)$ at most\n\\[\n \\frac{\\delta}{t(t+1)n(n+1)}.\n\\]\nThis bound is conditional on the history that selected the candidate,\nso it also holds unconditionally after averaging over that history.\nThe telescoping series\n$\\sum_{t\\geq1}1/[t(t+1)]=1$ and\n$\\sum_{n\\geq1}1/[n(n+1)]=1$ show that the sum of all pairwise failure\nprobabilities is at most $\\delta$. The countable union bound proves\nthe simultaneous event. Selecting a data-dependent finite $n$ cannot\nleave an event that already holds for every $n$.\n\\end{proof}\n\nThe bound is intentionally elementary. Its width is of order\n$\\sqrt{[\\log(t/\\delta)+\\log n]/n}$. A sharper time-uniform boundary\ncan improve validation cost, but is unnecessary for the correctness of\nthe acceptance and resource theorems below.\n\n\\begin{corollary}[Bounded numerical uncertainty]\n\\label{cor:rsi-measurement}\nSuppose the recorded values satisfy\n$|\\widehat Y_{t,j}-Y_{t,j}|\\leq e_{t,j}$ with certified\n$e_{t,j}\\geq0$. The simultaneous confidence radius for their recorded\naverage is\n\\begin{equation}\n \\widetilde r_{t,n}=r_{t,n}(\\delta)+\\frac1n\\sum_{j=1}^ne_{t,j}.\n \\label{eq:rsi-numeric-error}\n\\end{equation}\n\\end{corollary}\n\\begin{proof}\nThe difference between the true and recorded empirical averages is\nat most $n^{-1}\\sum_j e_{t,j}$. Add this deterministic bound to\n\\eqref{eq:rsi-simultaneous} by the triangle inequality.\n\\end{proof}\n\nThis allowance covers errors in recorded evaluation values. It does\nnot cover an unmodeled change in the task distribution or an evaluation\nprocedure that no longer measures \\eqref{eq:rsi-risk}.\n\n\\section{Acceptance, cumulative gain, and comparison with rejected candidates}\n\nAt a finite chosen validation size $n_t$, let $r_t$ denote a valid\nradius, including any numerical-error allowance, and define\n\\begin{equation}\n L_t=\\widehat\\Delta_t-r_t,\n \\qquad U_t=\\widehat\\Delta_t+r_t.\n \\label{eq:rsi-bounds}\n\\end{equation}\nChoose a nonnegative required improvement $\\gamma_t$. The controller\naccepts the proposal if and only if $L_t\\geq\\gamma_t$; otherwise it\nkeeps the incumbent. Write $I_t=1$ for acceptance and $I_t=0$ for\nrejection. The acceptance rule is thus\n\\begin{equation}\n s_t=\\begin{cases}a_t,&L_t\\geq\\gamma_t,\\\\s_{t-1},&L_t<\\gamma_t.\n \\end{cases}\n \\label{eq:rsi-accept}\n\\end{equation}\nAny early-rejection policy to which the comparison theorem is applied\nmust also end with $L_t<\\gamma_t$; rejecting a certifiably acceptable\nproposal for an unrelated reason does not meet that theorem's rule.\n\n\\begin{theorem}[A recursive acceptance guarantee and a finite oracle inequality]\n\\label{thm:rsi-oracle}\nOn the simultaneous event $\\mathcal E$, rule \\eqref{eq:rsi-accept}\nsatisfies, for every finite $T$,\n\\begin{align}\n R(s_T)&\\leq R(s_0)-\\sum_{t=1}^T I_t\\gamma_t,\n \\label{eq:rsi-cumulative}\\\\\n R(s_T)&\\leq\\min\\left\\{R(s_0),\n \\min_{1\\leq t\\leq T}\\bigl[R(a_t)+\\gamma_t+2r_t\\bigr]\n \\right\\}.\n \\label{eq:rsi-oracle}\n\\end{align}\nThese statements remain true when each accepted candidate changes the\nmechanism that generates later proposals, provided the common\nevaluation contract is maintained.\n\\end{theorem}\n\n\\begin{proof}\nAt an accepted round, simultaneous validity gives\n$\\Delta_t\\geq L_t\\geq\\gamma_t\\geq0$. At a rejected round the state\ndoes not change. Therefore\n\\[\n R(s_t)=R(s_{t-1})-I_t\\Delta_t\n \\leq R(s_{t-1})-I_t\\gamma_t.\n\\]\nTelescoping proves \\eqref{eq:rsi-cumulative}; it also shows that the\nincumbent risk is nonincreasing.\n\nFix any round $t\\leq T$. If its candidate was accepted, then\n$s_t=a_t$ and monotonicity gives $R(s_T)\\leq R(a_t)$. If it was\nrejected, then $L_t<\\gamma_t$. Since\n$\\Delta_t\\leq\\widehat\\Delta_t+r_t=L_t+2r_t$, one has\n\\[\n R(s_{t-1})<R(a_t)+\\gamma_t+2r_t.\n\\]\nAgain monotonicity gives the same upper bound for $R(s_T)$. This\nholds for every candidate, including rejected candidates, and for\nthe initial incumbent. Taking the minimum yields\n\\eqref{eq:rsi-oracle}.\n\nThe proof is pathwise on $\\mathcal E$. It imposes no structural\nrestriction on the history-dependent proposal generator. Replacing\nthat generator changes the later $a_t$, but not any step of the\nargument or the conditional validation guarantee.\n\\end{proof}\n\nThe phrase ``oracle inequality'' in \\eqref{eq:rsi-oracle} has a precise\nscope. It compares the result with candidates that were actually\nexamined, allowing their stated confidence widths and acceptance\nmargins. It supplies no bound on the quality of candidates that the\ngenerator never proposes.\n\n\\begin{corollary}[A finite evaluation budget for a candidate tournament]\n\\label{cor:rsi-finite-tournament}\nSuppose $K$ proposal rounds are planned, each candidate is fixed before\nits fresh validation sample, and each round uses a fixed sample size\n\\begin{equation}\n n\\geq\\frac{2}{\\eps^2}\\log\\frac{2K}{\\delta}.\n \\label{eq:rsi-tournament-samples}\n\\end{equation}\nUsing the fixed-sample radius $\\eps$ and a common margin $\\gamma\\geq0$\nin \\eqref{eq:rsi-accept} gives, with probability at least $1-\\delta$,\n\\begin{equation}\n R(s_K)\\leq\\min_{1\\leq t\\leq K}R(a_t)+\\gamma+2\\eps,\n \\qquad R(s_K)\\leq R(s_0).\n \\label{eq:rsi-tournament}\n\\end{equation}\nThe total evaluation count is $Kn$ paired tasks.\n\\end{corollary}\n\\begin{proof}\nFor each of the $K$ conditionally selected candidates,\n\\eqref{eq:rsi-hoeffding-tail} and \\eqref{eq:rsi-tournament-samples}\ngive failure probability at most $\\delta/K$. The finite union bound\nand the proof of Theorem~\\ref{thm:rsi-oracle} give the result.\n\\end{proof}\n\n\\subsection{Sequential decisions and validation cost}\n\nWithin a round, the controller may accept as soon as $L_{t,n}\\geq\\gamma_t$\nand reject as soon as $U_{t,n}<\\gamma_t$. While neither condition holds,\nit can collect another paired task, subject to its resource cap. A\nrejection at a finite cap also preserves Theorem~\\ref{thm:rsi-oracle}\nprovided the candidate is not accepted and $L_{t,n}<\\gamma_t$.\n\n\\begin{proposition}[A finite sample bound away from the decision boundary]\n\\label{prop:rsi-sequential-cost}\nConsider exact recorded observations and fix round $t$. Suppose the\ntrue improvement satisfies $|\\Delta_t-\\gamma_t|\\geq g$ for some\n$0<g\\leq1$. Put\n\\begin{equation}\n K=\\frac{2t(t+1)}{\\delta},\\qquad\n H=\\log\\frac{16K}{g^2},\\qquad\n N=\\left\\lceil\\frac{64H}{g^2}\\right\\rceil.\n \\label{eq:rsi-sequential-N}\n\\end{equation}\nOn $\\mathcal E$, the sequential rule reaches the correct side of the\nthreshold by sample $N$, if the resource cap permits those samples.\nThus a sufficient validation count is\n\\begin{equation}\n O\\!\\left(g^{-2}\\left[\\log\\frac{t}{\\delta}\n +\\log\\frac1g\\right]\\right).\n \\label{eq:rsi-sequential-order}\n\\end{equation}\n\\end{proposition}\n\\begin{proof}\nSince $K>4$ and $g\\leq1$, one has $H>\\log64>4$.\nThe definition of $N$ gives $N\\geq64H/g^2$,\n$N\\leq65H/g^2$, and $N+1\\leq66H/g^2$. Therefore\n\\[\n \\log(KN(N+1))\n \\leq\\log K+\\log4290+2\\log H+4\\log(1/g)<6H.\n\\]\nIndeed, the first and last terms together are at most $2H$,\n$2\\log H\\leq H$ for $H>4$, and $\\log4290<3H$.\nConsequently,\n\\[\n r_{t,N}^2=\\frac2N\\log(KN(N+1))\n <\\frac{3g^2}{16}<\\frac{g^2}{4}.\n\\]\nOn $\\mathcal E$, the lower confidence bound is at least\n$\\Delta_t-2r_{t,N}$ and the upper confidence bound is at most\n$\\Delta_t+2r_{t,N}$. If $\\Delta_t\\geq\\gamma_t+g$, the lower bound\nstrictly exceeds $\\gamma_t$ by time $N$. If\n$\\Delta_t\\leq\\gamma_t-g$, the upper bound is strictly below it.\nThe stated order follows by expanding $\\log K$ and $H$.\n\\end{proof}\n\nThe boundary case $\\Delta_t=\\gamma_t$ has no finite stopping guarantee.\nThis is the expected information limitation of a threshold decision.\nThe positive consequence is a conditional form of data efficiency:\ncandidates far from the acceptance boundary can be decided with fewer\nevaluation samples than those close to it. These samples, including\nthose used for early rejection, are charged by the ledger below.\n\n\\begin{example}[Why ordinary benchmark reuse changes the theorem]\nDraw a finite validation set from the uniform distribution on $[0,1]$.\nLet an incumbent have constant loss $1/2$. After seeing the validation\npoints, choose a candidate with loss zero on precisely those points\nand loss one elsewhere. Its observed improvement is $1/2$, but its\ntrue risk is one because a fresh continuous draw hits the finite set\nwith probability zero. The candidate is worse by $1/2$. This\nconstruction violates the fresh-evaluation conditional mean in\nAssumption~\\ref{ass:rsi-evaluation}; it is not a counterexample to the\ntheorem. It explains why changing proposal systems and reusing exposed\nevaluation data are different operations.\n\\end{example}\n\n\\section{Charging every attempted computation}\n\nA performance gate alone is not a compute-efficiency theorem. A\ncontroller may spend arbitrarily much on proposals it rejects. To\ninclude these expenditures, let $c_t\\geq0$ be the actual cost of round\n$t$, including proposal construction, training, certification, and\nvalidation. Fix an exchange factor $\\lambda>0$ converting those cost\nunits to the units of \\eqref{eq:rsi-risk}. Let $B_0\\geq0$ be an\ninitial exploration credit.\n\nBefore starting a round, the controller declares a cost cap\n$\\overline c_t$ and requires\n\\begin{equation}\n \\lambda\\overline c_t\\leq B_{t-1}.\n \\label{eq:rsi-reserve}\n\\end{equation}\nIt must stop the round before exceeding that cap. After the round,\nwith $c_t\\leq\\overline c_t$, update the certified credit by\n\\begin{equation}\n B_t=B_{t-1}-\\lambda c_t+I_tL_t.\n \\label{eq:rsi-ledger}\n\\end{equation}\nFor a failed or canceled attempt with no accepted candidate, the last\nterm is zero and the cost is still charged. Optional validation\nstopping is permitted within the predeclared total cap.\n\n\\begin{theorem}[A complete resource ledger]\n\\label{thm:rsi-ledger}\nUnder \\eqref{eq:rsi-reserve}--\\eqref{eq:rsi-ledger} and the acceptance\nrule, $B_t\\geq0$ for every completed round. On $\\mathcal E$,\n\\begin{equation}\n \\lambda\\sum_{t=1}^T c_t\n \\leq B_0+\\sum_{t=1}^T I_tL_t\n \\leq B_0+R(s_0)-R(s_T)\n \\leq B_0+1\n \\label{eq:rsi-total-cost}\n\\end{equation}\nfor every $T$. If every attempted round costs at least $c_{\\min}>0$,\nthen the number of completed rounds is at most\n$(B_0+1)/(\\lambda c_{\\min})$ on that event.\n\\end{theorem}\n\\begin{proof}\nBefore adding any accepted improvement,\n$B_{t-1}-\\lambda c_t\\geq B_{t-1}-\\lambda\\overline c_t\\geq0$.\nAcceptance implies $L_t\\geq\\gamma_t\\geq0$, so induction proves\n$B_t\\geq0$. Summing \\eqref{eq:rsi-ledger} gives\n\\[\n \\lambda\\sum_{t=1}^T c_t\n =B_0+\\sum_{t=1}^TI_tL_t-B_T\n \\leq B_0+\\sum_{t=1}^TI_tL_t.\n\\]\nOn $\\mathcal E$, $L_t\\leq\\Delta_t$, and the actual accepted risk\ndifferences telescope exactly:\n$\\sum_{t=1}^TI_t\\Delta_t=R(s_0)-R(s_T)$. This proves the middle\nbound. The risk range $[0,1]$ proves the last one. Finally,\n$\\sum_{t=1}^T c_t\\geq Tc_{\\min}$ gives the round-count bound.\n\\end{proof}\n\nThe ledger is an accounting guarantee, not a source of free compute.\nWith zero initial credit, no positive-cost experiment can start under\n\\eqref{eq:rsi-reserve}. With fixed positive per-round costs and fixed\ncredit, the number of experiments is finite. Accepted improvements can\njustify additional expenditures under the chosen exchange factor;\nthey do not make expenditures disappear.\n\nThis matters for evaluating an RSI proposal. A reported gain per\naccepted modification omits the cost of the unsuccessful search unless\nthose attempts are explicitly charged. Equation~\\eqref{eq:rsi-total-cost}\nuses all attempts and exposes the role of the initial exploration\nbudget.\n\n\\section{Using trajectory certificates to allocate the next job}\n\nThe optimization-trajectory results can inform which proposed\nexperiment to run. Suppose a finite set of prospective jobs has\ncertified benefit intervals $[\\ell_j,u_j]$ for a common declared\nobjective and costs $c_j$. Define their net intervals by\n\\begin{equation}\n [\\underline V_j,\\overline V_j]\n =[\\ell_j-\\lambda c_j,u_j-\\lambda c_j].\n \\label{eq:rsi-net-intervals}\n\\end{equation}\nInclude a stop action $j=0$ with\n$\\underline V_0=\\overline V_0=0$.\n\n\\begin{proposition}[Robust selection from certified intervals]\n\\label{prop:rsi-allocation}\nIf the actual net utilities $V_j$ lie in\n$[\\underline V_j,\\overline V_j]$, then:\n\\begin{enumerate}\n \\item stopping is optimal for every compatible realization if\n $\\overline V_j\\leq0$ for all jobs;\n \\item a job $j$ is optimal for every compatible realization if\n $\\underline V_j\\geq\\max_{i\\ne j}\\overline V_i$;\n \\item choosing job $j$ has regret at most\n $\\max\\{0,\\max_{i\\ne j}(\\overline V_i-\\underline V_j)\\}$.\n\\end{enumerate}\nIf all combinations of interval endpoints are jointly feasible, the\nthird bound is exact.\n\\end{proposition}\n\\begin{proof}\nThe first claim follows from $V_j\\leq\\overline V_j\\leq0=V_0$.\nFor the second,\n$V_j\\geq\\underline V_j\\geq\\overline V_i\\geq V_i$ for every other action.\nFor the third,\neach competing excess satisfies\n$V_i-V_j\\leq\\overline V_i-\\underline V_j$, while regret\nis nonnegative and comparison with the selected action itself adds\nzero. If all endpoint combinations are feasible, set $V_j=\\underline V_j$\nand an action attaining the largest competing endpoint to its\n$\\overline V_i$ to\nattain the bound whenever it is positive.\n\\end{proof}\n\nThese rules exploit provable domination without requiring a point\nforecast. When intervals overlap, the theorem does not certify a\nunique best action; the controller can acquire information, use a\ndeclared minimax criterion, or spend exploration credit. If the\nintervals describe different training losses, they must first be\nrelated to a common objective before this comparison is meaningful.\nValidation of the resulting candidate remains separate from a\ntraining-curve certificate.\n\n\\section{Conditional convergence to the best reachable candidate}\n\nFinite comparison guarantees do not require that the proposal\nmechanism eventually finds a good candidate. A convergence theorem\nmust state that additional search assumption. Let $\\mathcal A$ be a\ncountable collection of candidate states containing the initial state\nand every state that may be accepted. Assume that every\n$a\\in\\mathcal A$ is proposed infinitely often along the run. This is\na fairness condition on the actual candidate schedule, not a claim\nthat a practical generator satisfies it.\n\n\\begin{theorem}[Search consistency under fair proposals]\n\\label{thm:rsi-convergence}\nSuppose the run has infinitely many completed rounds, the proposal\nschedule is fair for $\\mathcal A$, and the acceptance rule uses\n$\\gamma_t\\to0$ and valid radii $r_t\\to0$. Then, on $\\mathcal E$,\n\\begin{equation}\n R(s_t)\\longrightarrow \\inf_{a\\in\\mathcal A}R(a).\n \\label{eq:rsi-convergence}\n\\end{equation}\n\\end{theorem}\n\\begin{proof}\nBy Theorem~\\ref{thm:rsi-oracle}, the incumbent risks form a\nnonincreasing sequence bounded below by zero. Let its limit be $r_*$.\nSince every accepted state belongs to $\\mathcal A$, one has\n$r_*\\geq\\inf_{a\\in\\mathcal A}R(a)$.\n\nIf this inequality were strict, choose $a\\in\\mathcal A$ and\n$\\eps>0$ such that $R(a)\\leq r_*-\\eps$. Fairness proposes this same\ncandidate at arbitrarily large rounds. At each such round,\n$\\Delta_t=R(s_{t-1})-R(a)\\geq\\eps$. On $\\mathcal E$,\n$L_t\\geq\\Delta_t-2r_t\\geq\\eps-2r_t$. For a sufficiently large one\nof these rounds, $2r_t+\\gamma_t<\\eps$, so the candidate is accepted.\nThen $R(s_t)=R(a)\\leq r_*-\\eps$, contradicting that $r_*$ is a lower\nbound for the entire nonincreasing incumbent sequence. Equality of\nthe two limits follows.\n\\end{proof}\n\nFor the elementary confidence radius, choosing $n_t=t^2$ makes\n$r_{t,n_t}(\\delta)\\to0$. This is an existence schedule, not a\ncompute-optimal prescription. In a finite candidate set, the finite\ntournament bound often answers the practical question more directly.\nThe theorem also does not combine an infinite positive-cost run with\na fixed finite ledger credit: Theorem~\\ref{thm:rsi-ledger} rules out\nthat combination. Asymptotic consistency requires a separate resource\nregime, for example increasing externally supplied budgets.\n\n\\section{Changing objectives with an explicit drift budget}\n\nSometimes the common evaluation task itself changes. Let the risk in\nround $t$ be $R_t:\\mathcal S\\to[0,1]$, fixed during that round, and\nsuppose a known sequence $d_t\\geq0$ satisfies\n\\begin{equation}\n \\sup_{s\\in\\mathcal S}|R_{t+1}(s)-R_t(s)|\\leq d_t.\n \\label{eq:rsi-drift}\n\\end{equation}\nFresh validation now estimates\n$\\Delta_t=R_t(s_{t-1})-R_t(a_t)$. The conditional concentration proof\nis unchanged, because it applies within each round. A drift bound is\nadditional information; it is not inferred from the fact that two\nsuccessive benchmark scores happen to be close.\n\n\\begin{theorem}[Cumulative improvement and consistency under summable drift]\n\\label{thm:rsi-drift}\nOn the simultaneous validation event, the acceptance rule satisfies\n\\begin{equation}\n R_T(s_T)\\leq R_1(s_0)-\\sum_{t=1}^T I_t\\gamma_t\n +\\sum_{t=1}^{T-1}d_t.\n \\label{eq:rsi-drift-cumulative}\n\\end{equation}\nIf $\\sum_td_t<\\infty$, then $R_t$ converges uniformly to a limit\nfunction $R_\\infty$, and $R_t(s_t)$ has a limit. If, in addition,\nthe fairness and vanishing-radius conditions of\nTheorem~\\ref{thm:rsi-convergence} hold, then\n\\begin{equation}\n R_t(s_t)\\longrightarrow\\inf_{a\\in\\mathcal A}R_\\infty(a).\n \\label{eq:rsi-drift-limit}\n\\end{equation}\n\\end{theorem}\n\n\\begin{proof}\nWithin round $t$, acceptance gives\n$R_t(s_t)\\leq R_t(s_{t-1})-I_t\\gamma_t$. Between rounds,\n$R_{t+1}(s_t)\\leq R_t(s_t)+d_t$. Alternately applying these two\ninequalities proves \\eqref{eq:rsi-drift-cumulative}.\n\nFor $u>t$, \\eqref{eq:rsi-drift} gives\n$\\sup_s|R_u(s)-R_t(s)|\\leq\\sum_{j=t}^{u-1}d_j$. Summability makes\nthe functions uniformly Cauchy, so they converge uniformly to a\nbounded function $R_\\infty$. Let $D_t=\\sum_{j=t}^{\\infty}d_j$.\nThe adjusted incumbent sequence\n\\[\n V_t=R_t(s_t)+D_t\n\\]\nis nonincreasing: for $t\\geq2$,\n\\[\n V_t\\leq R_{t-1}(s_{t-1})+d_{t-1}-I_t\\gamma_t+D_t\n =V_{t-1}-I_t\\gamma_t.\n\\]\nIt is bounded below by zero and therefore converges. Because\n$D_t\\to0$, the unadjusted $R_t(s_t)$ converges to the same limit,\nsay $r_*$. Uniform convergence and $s_t\\in\\mathcal A$ imply\n$r_*\\geq\\inf_{a\\in\\mathcal A}R_\\infty(a)$.\n\nIf the inequality were strict, choose $a\\in\\mathcal A$ with\n$R_\\infty(a)<r_*-3\\eps$ for some $\\eps>0$. At sufficiently large\nrounds, uniform convergence gives $R_t(a)<r_*-2\\eps$. Moreover,\n$R_t(s_{t-1})$ tends to $r_*$, since its distance from\n$R_{t-1}(s_{t-1})$ is at most $d_{t-1}\\to0$. Thus at every\nsufficiently late proposal of $a$, the true improvement exceeds\n$\\eps$. Fairness supplies such proposals, and $r_t,\\gamma_t\\to0$\nforces acceptance. The accepted risk would then be below\n$r_*-2\\eps$ at an arbitrarily late round, contradicting convergence\nto $r_*$. This proves \\eqref{eq:rsi-drift-limit}.\n\\end{proof}\n\nWithout a bound such as \\eqref{eq:rsi-drift}, changing the objective\ncan make an arbitrary sequence of systems appear to improve simply\nby changing what is scored. The theorem exposes the exact extra term\nneeded to compare genuine progress under a changing evaluation task.\nIt does not turn different metrics into interchangeable quantities.\n\n\\section{An auditable controller and the resulting RSI benefit}\n\nThe supplied \\texttt{recursive\\_controller.py} implements the\ncomparison gate and credit ledger using rational input values. Its\nlogarithm is bounded above by a convergent rational series, and its\nsquare root is rounded upward on an explicit rational grid. Thus a\ncomputed confidence radius is an upper bound on\n\\eqref{eq:rsi-radius}, rather than an unexamined floating-point\napproximation. The code accepts separate measurement-error bounds and\ncharges costs for rejected attempts. Its tests audit these arithmetic\nand accounting identities. The probability theorem and the validity\nof a real evaluation setup remain mathematical and modeling\nobligations, respectively.\n\nThe contribution to recursive learning is a complete conditional\ninterface. Proposals can be generated by changing optimizers,\narchitectures, representations, or proposal policies. Finite\ntrajectory certificates can reject impossible forecasts or eliminate\nprovably dominated computations. Fresh paired evaluation can certify\nimprovement in a common risk. The acceptance theorem carries those\ncertificates through arbitrary adaptive future proposal choices,\nwhile the ledger accounts for the full search cost. Under explicit\nfairness and sufficient evaluation resources, the loop converges to\nthe best risk in its reachable candidate set, including the stated\nsummably drifting extension.\n\nThese benefits are mathematical: valid decisions, controlled\nuncertainty, cumulative performance guarantees, and explicit compute\naccounting. A claim of faster practical self-improvement would\nadditionally require a specified candidate generator, task\ndistribution, comparison policy, and measured total cost. None of\nthose empirical quantities is supplied by a theorem about curve\ncurvature alone.\n"}
{"format": "latex", "id": "10_verification", "sha256": "ef3cbb234a5b4cdbccd8d38b177eac1f49676ddcd9a7cba58c409119fa041640", "source_path": "research/chapters/10_verification.tex", "text": "\\chapter{Verification, scientific scope, and use}\n\\label{ch:verification}\n\n\\section{A claim is accompanied by its kind of evidence}\n\nThe central results are mathematical statements with quantified proofs.\nThe executable material serves three complementary purposes: checking\nalgebra that is tedious to audit by eye, verifying fully specified\nrational witnesses, and making the proposed decision interfaces\ninspectable. These purposes should remain distinct. A thousand exact\nfinite cases do not prove a theorem for every positive integer; a\nfloating-point optimum does not prove a universal inequality; and a\nstatistical acceptance rule does not establish that its sampling\nassumptions hold in an application.\n\nThe following ledger locates the principal assertions. Exact audit\nmeans integer or rational arithmetic unless a display approximation is\nexplicitly marked.\n\n\\begin{center}\n\\small\n\\begin{tabularx}{\\textwidth}{@{}>{\\raggedright\\arraybackslash}p{.25\\textwidth}>{\\raggedright\\arraybackslash}p{.35\\textwidth}>{\\raggedright\\arraybackslash}X@{}}\n\\toprule\nAssertion & Mathematical evidence & Executable evidence\\\\\n\\midrule\nCompensation threshold &\nInterpolation identity, square completion, strict interpolable\ncounterexamples, analytic transfer &\nCoefficient identities, rational strict slacks, horizon identities\\\\\nAdjacent threshold &\nAlgebraic root isolation, positive matrix, simultaneous negative\ncurvatures above the root &\nExact polynomial coefficients, rational interval bounds, repaired\nrational transcript\\\\\nArbitrary long runs &\nChamber induction, rational curvature identity, Gaussian tail\ncontrol, global contraction &\nExact branch decisions and 1,026 negative curvatures across the\ndeclared finite suite\\\\\nComplete finite prediction &\nConstructive interpolation, Gram realization, compactness, separation &\nAmbient/Gram equality checks and exact residual-certificate checks\\\\\nInformation barriers &\nExplicit indistinguishable objectives and finite-jet perturbations &\nExact common prefixes, continuation values, and vanishing derivatives\\\\\nRecursive controller &\nConditional concentration, a simultaneous event, telescoping,\nresource accounting &\nRational confidence allowances and accept/reject/reservation checks\\\\\n\\bottomrule\n\\end{tabularx}\n\\end{center}\n\nEvery row has a written proof; none obtains its quantified conclusion by\nextrapolating the last column. The package contains no assertion that\nthese proofs have been verified by a proof-assistant kernel. Internal\ncross-checking is also distinguished from external replication.\n\n\\section{The exact arithmetic suite}\n\nThe entry point is \\texttt{run\\_verification.py}. It runs the seven\nproof-audit programs, retains their structured reports, and writes an\naggregate summary under \\texttt{verification/}. Its dependency is the\nPython standard library. Plotting and semidefinite exploration are\nseparate optional workflows.\n\n\\subsection{Long-run geometry and algebra}\n\nThe default long-run audit uses\n\\[\n m\\in\\{1,2,3,4,8,16,32,64,128,256,512\\}.\n\\]\nIt constructs the actual squared-distance objective values along each\nrational orbit, verifies the active projection chambers, checks the\nstated gradient-descent updates, and compares every queried second\ndifference with its rational upper bound. The total number of checked\nnegative second differences is\n\\[\n 1+2+3+4+8+16+32+64+128+256+512=1026.\n\\]\nThe script also checks finite algebraic ingredients of the chamber\nanalysis and recurrence. Values converted to decimal for a JSON display\nor a graph are not used to decide a strict inequality.\n\nThe analytic objective is defined by an exact Gaussian convolution.\nIts analyticity, globally optimal curvature constants, and uniform\nGaussian error bounds are established in Chapter~\\ref{ch:analytic}.\nThe verifier does not replace those arguments with quadrature. This\ndistinction is especially useful when a certified negative curvature\nmargin is very small: subtracting three nearly equal decimal losses\nwould be a poor way to verify its sign.\n\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=\\textwidth]{figures/analytic_scales.pdf}\n \\caption{Two exact quantitative conclusions of the analytic long-run\n theorem, plotted using decimal arithmetic. The smoothing width decays\n as $1/[M\\sqrt{\\log M}]$, while the optimal condition number is exactly\n $128M$. The curves illustrate formulas; they are not fitted\n measurements or estimates from an optimizer.}\n \\label{fig:analytic-scales}\n\\end{figure}\n\n\\subsection{Sharp boundaries}\n\nThe compensation audit checks a concrete rational counterexample at\n$h=93/50$, all twelve ordered interpolation inequalities among its four\nobserved states, and a positive admissible strong-convexity constant.\nIt separately verifies 100 finite-horizon scalar identities. The\nall-parameter construction and the distinction between scalar\nsharpness and gradient-descent realizability are in\nChapter~\\ref{ch:universal}.\n\nThe adjacent-boundary audit uses coefficientwise polynomial arithmetic.\nIt checks the discriminant factorization, the repeated-root identities,\nthe endpoint signs and derivative bounds isolating $h_*$, and the\nparameter enclosures used to prove matrix positivity. A second program\nchecks the negative construction and a completely rational repaired\ntranscript at $h=233/125$.\n\nFor that repaired transcript the two curvatures are approximately\n$-5.69\\cdot10^{-17}$ and $-4.83\\cdot10^{-17}$. Their signs are decided by\nexact rational numerators. These deliberately conservative margins\ncertify existence close to the boundary; they are not predictions of a\nmeasurable speedup. The written perturbation and smoothing arguments\nthen establish analytic strong-convexity realizations.\n\nThe coefficient degrees are part of the audit. When a polynomial\nidentity is checked on a tensor grid, the grid contains more distinct\npoints in each variable than the corresponding degree. Consequently\nzero values on that grid imply the polynomial is identically zero.\nThis is exact polynomial interpolation, not a numerical sample test\nwithout an accompanying degree argument.\n\n\\subsection{Decision certificates and information limits}\n\nThe finite-model audit compares 200 ambient-coordinate interpolation\ncalculations with their Gram-matrix forms. It checks positive matrix\ncertificates at five declared steps, rejects a certificate outside its\nvalid step range, and tests semidefinite corner cases including a zero\ndiagonal with an incompatible nonzero off-diagonal entry. Twelve\ninexact quadratic trajectories check the displacement/error interface.\nThe three-step measurement example reproduces the exact allowance\n\\[\n \\frac{4750081}{125000000000}=0.000038000648.\n\\]\n\nThe information audit includes 18 common-history cases and 756 exact\nfuture values, together with 18 analytic finite-jet cases and 378\nvanishing-jet equalities. It also checks the finite curvature-expenditure\nidentity for 64 declared lengths. These are witnesses and algebraic\naudits of theorems whose arbitrary-history and arbitrary-finite-order\nquantifiers are proved in Chapter~\\ref{ch:information}.\n\n\\begin{figure}[htbp]\n \\centering\n \\includegraphics[width=\\textwidth]{figures/information_barrier.pdf}\n \\caption{An information barrier in one dimension. The two objectives\n agree on the observed half-line, including values and gradients, but\n have different curvature in the unobserved region. Starting at the\n same endpoint with $v=1$, $\\mu=1/4$, $L=1$, and $\\eta=1/2$, their\n additional gains converge to $2$ and $1/2$. The displayed functions\n are an illustrative piecewise-quadratic member of the construction;\n the separate analytic theorem preserves finite jets rather than an\n entire half-line.}\n \\label{fig:information-barrier}\n\\end{figure}\n\nThe recursive-controller audit comprises 39 assertion groups. Its\nconfidence-radius implementation bounds the logarithm by an exact\nrational series remainder and rounds the square root upward on a\ndeclared rational grid. Thus a rounding artifact cannot silently turn\na lower confidence bound into an overstatement. The ledger exercises\ncompleted, rejected, and canceled attempts, including a reservation\nthat must be refused because it exceeds available credit.\n\n\\section{Numerical exploration and exact certification}\n\nThe optional semidefinite program illustrates how to search a feasible\ntranscript model for extreme future queries. Its output records the\nsolver status, interpolation residuals, and the smallest Gram\neigenvalue. The supplied exploratory run includes the status\n\\texttt{optimal\\_inaccurate}. It is retained as an explicitly numerical\nrecord and is not counted as a proved prediction interval.\n\nA usable certificate can be extracted from such a search without\ntrusting its final floating-point declaration. One records candidate\nnonnegative interpolation weights and a candidate positive matrix,\nrounds them to rationals, and computes the residual identity exactly.\nThe allowance in Chapter~\\ref{ch:transcripts} accounts for imperfect\ncoefficient cancellation and a quantified matrix defect. The resulting\nrational lower bound is valid if the exact audit succeeds. If it does\nnot succeed, the numerical output remains a proposal for further work.\n\nThe distinction is an architectural advantage for automated research.\nA large or heuristic search can generate useful candidates while a\nsmall independent checker determines which claims may enter the\ntrusted decision state. The search algorithm and the checker need not\nhave comparable complexity. This manuscript supplies instances of that\nseparation; it does not claim that every difficult certificate can be\nfound efficiently.\n\n\\section{Relationship to established and concurrent work}\n\nThe classical sharp $7/4$ boundary and the question of consecutive\nconcavity come from Barzilai, Shamir and Zamani\n\\cite{barzilai2025}. Smooth strong interpolation and its application to\nexact finite performance estimation are established tools\n\\cite{taylor2017,drori2014}. Their proofs and notation are included here\nto make the constructions reproducible; their inclusion does not turn\nthem into new discoveries. Hieu studies related robustness, locality,\nquadratic, and algorithmic questions \\cite{hieu2025}.\n\nThe public \\texttt{optcurves} project already reports exact\nconsecutive-violation examples and threshold brackets for lengths two\nthrough six \\cite{optcurves2026}. In particular it brackets the\ntwo-violation threshold between $1.8634$ and $1.8635$, and gives\ntwo-dimensional examples. It also develops a condition-dependent exact\none-violation threshold and selected Lean certificates.\nThe present algebraic value $h_*$ resolves that two-violation bracket.\nThe arbitrary-length theorem adds an explicit two-dimensional analytic\nfamily with exact conditioning and quantitative margins. Bare\nconsecutive existence and the use of interpolation certificates are\ntherefore not claimed as exclusive contributions here. The cited\nrepository's formal artifacts have not been rerun as part of this\npackage.\n\nFor recursive validation, bounded-variable concentration and the need\nto protect evaluation against adaptive reuse are well established\n\\cite{hoeffding1963,howard2021,dwork2015}. The controller specializes\nthose ideas into an explicit contract with candidate comparisons,\nnumerical allowances, an oracle inequality, and a ledger charging every\nattempt. Its theorem is not evidence of a new concentration inequality.\n\nThe literature review was targeted: primary papers and the closely\nrelated public repository were inspected for direct overlap. This\nestablishes the comparisons stated above. It does not establish\nexhaustive historical priority for every lemma or guarantee future\npublication impact. Those questions are separate from whether the\nproofs establish their stated mathematical conclusions.\n\n\\section{A concrete integration contract for recursive learning}\n\nA system using these results needs to make the following decisions\nexplicit. Each item has a mathematical role, rather than being an\nadministrative label.\n\n\\begin{enumerate}\n \\item \\textbf{Declare the training model.} Specify the objective class,\n smoothness and strong-convexity bounds, update rule, observations, and\n error budget. Convexity assumptions cannot be inferred from a\n visually smooth loss plot.\n \\item \\textbf{State the proposed decision.} A query such as at least\n $\\tau$ further decrease in $N$ steps becomes an affine loss\n functional. The complete model provides its attainable interval;\n a cheap universal law may already answer the query.\n \\item \\textbf{Check the certificate.} Accept a bound only after the\n interpolation weights, matrix positivity, and residual allowance have\n been checked. If the interval crosses the decision threshold, the\n available model leaves the decision unresolved.\n \\item \\textbf{Generate and freeze a candidate.} A proposed change may\n include the model, optimizer, or proposal mechanism itself. Its\n evaluation behavior must remain fixed during the paired comparison.\n \\item \\textbf{Validate on the common task.} Use fresh observations\n satisfying the conditional sampling contract. Include measurement\n and numerical allowances in the lower confidence bound. A training\n improvement alone need not imply an improvement in the task risk.\n \\item \\textbf{Reserve and reconcile resources.} Reserve a cost cap\n before the attempt. Charge all used compute and release unused\n reservations. Credit only the certified gain of an accepted change.\n\\end{enumerate}\n\nOn the simultaneous event of Chapter~\\ref{ch:recursive}, this procedure\nhas an auditable cumulative improvement guarantee and a finite\ncomparison with every evaluated candidate. It also has a finite\nresource bound in the chosen risk units. The benefit is that recursive\nchanges can carry precise, checkable evidence across iterations.\n\nThere is a real tradeoff. More stringent acceptance thresholds reduce\nthe set of certifiable changes and may require more validation samples.\nSmall or uncertain improvements may be rejected even when they are\nbeneficial. A fresh evaluation stream costs resources. Poor proposal\ngeneration can make the procedure stagnate. These facts are consistent\nwith the theorems: certification controls what may be asserted about\naccepted changes; it does not manufacture favorable candidates.\n\n\\section{Remaining mathematical and empirical questions}\n\nLet $\\theta_m$ be the infimum normalized step at which some convex\nsmooth gradient-descent trajectory has $m$ consecutive strictly negative\nsecond differences, allowing arbitrary finite dimension. The results\ngive\n\\[\n \\theta_1=\\frac74,\\qquad \\theta_2=h_*,\n \\qquad\n h_*\\leq\\theta_m\\leq 2-\\frac1{128(m+1)}\n \\quad(m\\geq2).\n\\]\nThe lower bound follows because a longer run contains an adjacent\npair. The upper bound is the explicit analytic construction.\nFurthermore $\\theta_m$ is nondecreasing in $m$, by containment of the\nevents. The upper bound tending to $2$ does \\emph{not} prove that\n$\\theta_m$ tends to $2$.\n\nSeveral sharply stated problems remain.\nExact algebraic thresholds for $m\\geq3$ would extend the local phase\ndiagram. Uniform lower bounds on the condition number necessary for an\n$m$-step run would determine how close the linear construction is to\noptimal. Fixed-dimension or fixed-condition-number thresholds impose\nadditional restrictions not settled by the unrestricted interpolation\nmodel. Reusable certificates for broader adaptive methods would require\ncareful treatment of coefficients chosen from the observed history.\nEfficient algorithms for complete prediction would need complexity\nguarantees in addition to the existence of exact finite models.\n\nThe application questions are different. Whether the controller improves\nneural-training efficiency, how to estimate useful model bounds, and\nwhich candidate generators produce favorable improvements require\nexperiments on declared tasks. A credible evaluation would report all\nattempt costs, validation costs, accepted and rejected proposals,\nuncertainty coverage, and comparison with a fixed-budget baseline.\nNone of those empirical results is asserted in this release.\n\nThe broad research program is a theory of finite learning evidence:\nwhich trajectories are possible, which future claims follow from an\nobserved history, which certificates can be checked independently, and\nhow those claims compose under recursive changes. The exact boundaries,\nanalytic constructions, information barriers, and validation ledger\nprovide concrete mathematical components of that program.\n\n"}
{"format": "latex", "id": "11_appendices", "sha256": "d3a47dd397c222457ceeb025fed8efd7fc82993c4fa5efb85e4acd31f0337c71", "source_path": "research/chapters/11_appendices.tex", "text": "\\chapter{Notation and an independent audit guide}\n\\label{ch:audit-guide}\n\nThis appendix gives a short route through the assumptions, arithmetic,\nand logical implications behind the main results. It is intended for a\nreader reproducing a certificate without rerunning the numerical search\nthat suggested it. The main chapters contain the proofs; the checks\nbelow identify exactly which parts can be verified by finite arithmetic\nand which require the accompanying mathematical argument.\n\n\\section{Notation and assumption ledger}\n\n\\begin{longtable}{@{}p{0.22\\textwidth}p{0.73\\textwidth}@{}}\n\\toprule\nNotation & Meaning and convention\\\\\n\\midrule\n\\endhead\n$\\Fclass_{\\mu,L}$ &\nDifferentiable, globally $\\mu$-strongly convex, globally $L$-smooth\nfunctions on a finite-dimensional Euclidean space; $0\\leq\\mu<L$.\nThe parameters are certified bounds unless optimality is stated.\\\\[3pt]\n$\\eta,\\ h$ &\nThe physical gradient step and its normalized value $h=\\eta L$.\nIn an $L=1$ proof, the step is $h$.\\\\[3pt]\n$x_k,\\ g_k,\\ f_k$ &\nIterate, exact gradient $\\nabla f(x_k)$, and value $f(x_k)$ of one\nfixed objective, unless a perturbation or change is explicitly modeled.\\\\[3pt]\n$D_k,\\ \\kappa_k$ &\n$D_k=f_k-f_{k+1}$ and\n$\\kappa_k=f_k-2f_{k+1}+f_{k+2}=D_k-D_{k+1}$.\nConcavity means $\\kappa_k<0$.\\\\[3pt]\n$d_k$ &\nGradient difference $g_{k+1}-g_k$, with this orientation used\nin every curvature certificate.\\\\[3pt]\n$I_{ij}$ &\nOrdered interpolation slack\n$f_i-f_j-\\ip{g_j}{x_i-x_j}-\\norm{g_i-g_j}^2/(2L)$.\nReversing the indices changes the slack.\\\\[3pt]\n$G$ &\nGram matrix of the gradients, $G_{ij}=\\ip{g_i}{g_j}$.\nIt is symmetric and positive semidefinite.\\\\[3pt]\n$\\mathcal K$ &\nThe feasible set of finite transcripts after fixing the method,\nnormalization, function class, and observations. Predictions are\nrelative to these declared assumptions.\\\\[3pt]\n$q_-,q_+$ &\nMinimum and maximum of a stated affine query on a nonempty compact\ntranscript set. Every intermediate value is attainable.\\\\[3pt]\n$\\hiso$ &\n$(19-\\sqrt{17})/8$, the sharp universal endpoint for the specific\nlaw $2\\kappa_k+\\kappa_{k+1}\\geq0$.\\\\[3pt]\n$h_*,r_*$ &\nThe isolated algebraic step and positive weight of\nChapter~\\ref{ch:adjacency}; $r_*\\kappa_k+\\kappa_{k+1}\\geq0$\nholds for every $0<h\\leq h_*$.\\\\[3pt]\n$R(s),\\Delta_t$ &\nCommon evaluation risk of a candidate system and true improvement\nin proposal round $t$, as defined in Chapter~\\ref{ch:recursive}.\\\\\n\\bottomrule\n\\end{longtable}\n\nNormalizing $L$ to one means replacing $f$ by $f/L$.\nThe gradients and loss differences are divided by $L$, the strong\nconvexity parameter becomes $\\mu/L$, and the same iterates use step\n$h=\\eta L$. Therefore the sign of every curvature is preserved.\nTranslations of the coordinates and a common translation of function\nvalues also preserve the statements. The $L=1$ convention alone does\nnot assert that the smallest possible smoothness constant equals one.\n\nThe universal curvature laws assume exact constant-step gradient descent\non one fixed globally smooth convex objective. Their proofs do not\nrequire a minimizer. A negative witness additionally checks the whole\nfinite orbit and all ordered interpolation constraints. Strong\nconvexity and analyticity are supplied by separate realization arguments.\nThe scalar envelope in Theorem~\\ref{thm:scalar-horizon-envelope} is\nsharp over nonnegative decrement sequences satisfying its stated\nconstraint; that statement does not assert that every extremal\nsequence is realized by smooth gradient descent.\n\nThe forecasting results assume a nonempty bounded model set with the\nnormalization specified in Chapter~\\ref{ch:transcripts}. The recursive\nevaluation results use bounded paired observations with the conditional\nmean in Assumption~\\ref{ass:rsi-evaluation}; a changing candidate or\nreused evaluation data must still satisfy that contract. These are\ndifferent assumptions from convexity of a training objective.\n\n\\section{Logical dependencies of the proofs}\n\nThe implications below provide an audit order. A numerical optimizer\nis not a premise of any theorem in this list.\n\n\\begin{enumerate}[leftmargin=*,label=\\arabic*.]\n\\item \\textbf{Global realizability from finite data.}\nThe smooth convex interpolation theorem\n(Theorem~\\ref{thm:convex-interpolation}) is the classical\ninterpolation result of Taylor, Hendrickx, and Glineur\n\\cite{taylor2017}. Its constructive proof and the transformation\n$f\\mapsto f-\\mu\\norm{x}^2/2$ give\nTheorem~\\ref{thm:strong-interpolation}.\nGram factorization then gives Theorem~\\ref{thm:finite-transcript}.\nThe performance-estimation use of finite Gram models follows the\nframework of Drori and Teboulle \\cite{drori2014}.\n\n\\item \\textbf{Universal laws.}\nInterpolation necessity, exact coefficient identities, and\npositive semidefiniteness prove the curvature inequalities.\nThe sharp $7/4$ result of Barzilai, Shamir, and Zamani\n\\cite{barzilai2025} is presented with its proof in\nTheorem~\\ref{thm:classical-seven-fourths}.\nTheorems~\\ref{thm:compensation} and \\ref{thm:sharp-adjacency}\nuse their displayed certificates. Summing already proved inequalities\ngives the filtered and horizon consequences.\n\n\\item \\textbf{Sharpness and regularity.}\nA negative value of a proposed quadratic certificate does not\nestablish a counterexample. The negative directions first construct\ndata obeying the update equations and all interpolation inequalities.\nThe interpolation theorem realizes the data globally. Quantitative\nsmoothing bounds then preserve strict curvature margins along the\nperturbed orbit. The all-length witnesses in\nChapters~\\ref{ch:witness}--\\ref{ch:analytic} additionally use the\nexplicit orbit recurrence and its chamber margins.\n\n\\item \\textbf{Exact prediction and finite certificates.}\nTheorem~\\ref{thm:complete-prediction} uses realizability, compactness,\nand convexity of the transcript model. Strict certificate completeness,\nTheorem~\\ref{thm:certificate-completeness}, uses finite-dimensional\nconvex separation followed by rational approximation with a positive\nmargin. It does not assume that a semidefinite dual attains an exact\nzero-margin boundary certificate.\n\n\\item \\textbf{Perturbations and recursive decisions.}\nNonexpansiveness or contraction plus induction gives the trajectory\ntube in Theorem~\\ref{thm:trajectory-tube}. Smoothness converts that\ndistance bound into a loss error and hence a transferred linear law.\nThe recursive validation rule instead uses the bounded-variable\nexponential inequality of Hoeffding \\cite{hoeffding1963} and a\nsummable allocation of failure probabilities. Confidence-sequence\nmethods provide broader context \\cite{howard2021}; the particular\nguarantees here follow from the stated evaluation contract and the\ndirect proof in Chapter~\\ref{ch:recursive}.\n\\end{enumerate}\n\n\\section{A hand-checkable rational certificate}\n\nSet $L=1$ and $h=9/5$. For any four consecutive gradient-descent\nsamples, put\n\\[\n S=\\frac23f_0-f_1+\\frac13f_3\n =\\frac13(2\\kappa_0+\\kappa_1).\n\\]\nThe five interpolation multipliers, indexed by\n$(01,02,20,31,32)$, are respectively\n\\[\n \\lambda=\\left(\\frac79,\\frac29,\\frac13,\\frac29,\\frac19\\right).\n\\]\nAll are nonnegative. Expansion of the interpolation slacks gives\n$S-\\sum\\lambda_{ij}I_{ij}=d^TQd$, where\n\\[\n Q=\n \\begin{pmatrix}\n 1/15&7/90&0\\\\\n 7/90&17/90&1/9\\\\\n 0&1/9&1/6\n \\end{pmatrix}.\n\\]\nHere $d^TQd=\\sum_{i,j=0}^2Q_{ij}\\ip{d_i}{d_j}$.\nAn exact positive factorization, requiring no eigenvalue calculation, is\n\\begin{equation}\n d^TQd=\n \\frac1{15}\\left\\|d_0+\\frac76d_1\\right\\|^2\n +\\frac16\\left\\|d_2+\\frac23d_1\\right\\|^2\n +\\frac{13}{540}\\norm{d_1}^2.\n \\label{eq:appendix-rational-sos}\n\\end{equation}\nFor example, the middle diagonal coefficient is checked by\n\\[\n \\frac1{15}\\left(\\frac76\\right)^2+\n \\frac16\\left(\\frac23\\right)^2+\\frac{13}{540}\n =\\frac{49+40+13}{540}=\\frac{17}{90}.\n\\]\nThe two cross coefficients are $7/45$ and $2/9$.\nThus the multipliers, identity, and matrix positivity are independently\nvisible, and $2\\kappa_0+\\kappa_1\\geq0$ follows. A matrix with\nnonnegative diagonal entries alone would not suffice.\n\nThe same procedure accommodates a rational residual. Suppose exact\ncoefficient comparison for a different finite model gives\n\\[\n Q_{\\rm query}-b-\\sum_j\\lambda_j\\ell_j\n =\\tr\\!\\left(\n \\begin{pmatrix}-1&0\\\\0&2\\end{pmatrix}G\\right)\n +\\frac1{10}f_0-\\frac15f_1+3,\n\\]\nwith $\\lambda_j\\geq0$, $\\tr G\\leq4$, $|f_0|\\leq1$, and\n$|f_1|\\leq2$. Adding the identity matrix makes the residual matrix\n$\\diag(0,3)\\succeq0$. Theorem~\\ref{thm:residual-cert} therefore gives\n\\[\n Q_{\\rm query}\\geq b+3-4-\\frac1{10}-\\frac25\n =b-\\frac32.\n\\]\nThis small example illustrates why an indefinite rounded residual must\nbe charged to the trace bound. Its negative eigenvalue cannot simply\nbe discarded. Any normalization constraints that fix a value exactly\ncan instead be eliminated before forming the residual.\n\n\\section{Independently checking the algebraic endpoint}\n\nThe endpoint audit can be performed using integer and rational\narithmetic, followed by the exact mathematical implications indicated\nbelow.\n\n\\begin{enumerate}[leftmargin=*,label=\\arabic*.]\n\\item \\textbf{Isolate the specified root.}\nUse the coefficients of $P$ in \\eqref{eq:adj-polynomial} and the\nrational endpoints $\\ell,u$ in \\eqref{eq:adj-root-interval}.\nHorner evaluation verifies $P(\\ell)<0<P(u)$.\nLet $c_0=(\\ell+u)/2$, $w=(u-\\ell)/2$, and expand\n$P(c_0+z)=\\sum_{j=0}^7q_jz^j$ exactly. Check\n\\[\n 933<q_1<934,\\qquad\n \\sum_{j=2}^7j|q_j|w^{j-1}<\\frac1{100000}.\n\\]\nThis proves that $P'$ is positive throughout the interval, so the\nsign change identifies exactly one root there. No assertion about\nother real roots is needed.\n\n\\item \\textbf{Check the algebraic elimination.}\nForm the cubic $D(r,h)$ from \\eqref{eq:adj-D} and verify\n$D=-288\\det M$ by coefficient comparison. With its coefficients\n$a,b,c,d$, set $H=b^2-3ac$ and compute the discriminant $\\Delta$.\nVerify $\\Delta=432(h-2)P(h)$, then clear denominators in\n\\eqref{eq:adj-cubic-derivative}--\\eqref{eq:adj-cubic-value}.\nAt the isolated root, $H>0$ makes these identities legitimate and\ngives $D(r_*,h_*)=\\partial_rD(r_*,h_*)=0$.\n\n\\item \\textbf{Check all sign conditions.}\nUse rational interval bounds for $r_*$ and $h_*$. Verify every\ninterpolation multiplier is nonnegative and the first and last\ndiagonal entries of $M(r_*,h_*)$ are positive.\nThe zero determinant then makes the remaining Schur complement zero.\nCheck the positive semidefinite increment matrix in\n\\eqref{eq:adj-matrix-monotone}. This last step extends the endpoint\ncertificate to the whole interval $0<h\\leq h_*$.\n\n\\item \\textbf{Audit the opposite direction separately.}\nThe rational transcript at $h=233/125$ is a finite arithmetic check:\nverify the updates, both strict curvature signs, and every ordered\nstrong interpolation inequality. The family construction in\nChapter~\\ref{ch:adjacency} is additionally needed for every step\narbitrarily close above $h_*$. Its explicit repair and smoothing\nmargins prove realizability and analyticity. A single rational\nwitness, or failure of a chosen matrix to be positive semidefinite,\ndoes not establish that full sharpness statement.\n\\end{enumerate}\n\nThe boundary and negative-direction verifiers are separate programs so\nthat the positive certificate and the realizable obstruction can be\naudited independently. Decimal fields in the result files are\nconvenient displays; the assertions use exact fractions. The\nconservative negative witness has a very small strict margin, which\nestablishes existence rather than a practically large effect.\n\n\\section{Reproducing the supplied record}\n\nFrom the directory containing \\texttt{manuscript.tex}, the complete\nverification and typesetting entry points are\n\\begin{lstlisting}\npython run_verification.py\npython build.py\n\\end{lstlisting}\nThe exact certificate programs use Python's standard-library rational\narithmetic. The build requires a complete XeLaTeX installation with\nthe fonts named in \\texttt{preamble.tex}. To regenerate the\nmathematical figures as part of the build, use\n\\begin{lstlisting}\npython build.py --figures\n\\end{lstlisting}\nFigure generation has the plotting dependencies listed with the\nreproduction material. Figures visualize the formulas; they are not\ninputs to the exact verifiers.\n\nFor a focused endpoint audit, run\n\\begin{lstlisting}\npython code/verify_adjacent_boundary.py\npython code/verify_adjacent_negative.py\npython code/verify_sharp_compensation.py\n\\end{lstlisting}\nThe accompanying JSON records in \\texttt{verification/} retain exact\nrational quantities and named assertions. The remaining exact scripts\ncheck the general certificate implementation, finite witness instances,\nand information bounds. The controller demonstration tests the declared\nacceptance and accounting rules. Numerical transcript optimization is\nreported separately, including the solver's actual status.\n\nA reproduction report should distinguish three outcomes. Successful\nexact checks establish the reported finite identities and inequalities.\nThe written arguments extend those facts to all dimensions, all\nadmissible objectives, or all horizons when the theorem says so.\nSuccessful numerical optimization supplies a candidate or a numerical\ncomparison. The supplied Python checks are not a proof-assistant\nformalization, and a finite sweep of horizon lengths is not a substitute\nfor the all-length witness proof.\n"}