% Manuscript revision dated 19 November 2026. % Based on bkt_diamond_updated.tex and MANUSCRIPT_RECOMMENDATIONS.md; numbering preserved. \documentclass[a4paper,11pt]{article} \usepackage[utf8]{inputenc} \usepackage[T1]{fontenc} \usepackage[english]{babel} \usepackage{mathpazo} \usepackage{amsmath,amsthm,amssymb,mathtools,graphicx} \usepackage[top=2.5cm,bottom=2.5cm,left=2cm,right=2cm]{geometry} \usepackage[unicode,colorlinks=true,linkcolor=blue,citecolor=red, urlcolor=blue,pdfencoding=auto,psdextra]{hyperref} \linespread{1.1} \usepackage[numbers,sort&compress]{natbib} \usepackage{enumitem} \setlist[itemize]{leftmargin=2em,itemsep=0.25em,topsep=0.35em} \setlist[enumerate]{leftmargin=2.25em,itemsep=0.25em,topsep=0.35em} \numberwithin{equation}{section} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \renewcommand{\P}{\mathbb P} \newcommand{\E}{\mathbb E} \newcommand{\Var}{\operatorname{Var}} \newcommand{\dd}{\,\mathrm d} \numberwithin{equation}{section} \title{BKT asymptotics for the quenched free energy of hierarchical pinning} \author{ChatGPT, directed by Igor Kortchemski \thanks{CNRS \& DMA, \'Ecole normale sup\'erieure, PSL University, 75005 Paris, France, \texttt{igor.kortchemski@math.cnrs.fr}}} \date{November 2026} \begin{document} \maketitle \begin{abstract} We study the disordered pinning model of Derrida, Hakim and Vannimenus on the diamond hierarchical lattice. Its partition functions satisfy $R_{n+1}\overset d=(R_n^{(1)}R_n^{(2)}+B-1)/B$, where $B>2$, the two variables on the right are independent copies of $R_n$, and $R_0=e^hW$ with $W>0$ and $\E W=1$. Let $F$ be the quenched free energy and $h_c$ its critical point. If $h_c>\log(B-1)$, we prove $F(h_c+u)=\exp[-(K+o(1))/\sqrt u]$ as $u\downarrow0$, for a finite positive constant $K$ characterized by the response at criticality to a change of field. We identify a limiting profile of the rescaled tilted measures and extend their convergence in exponential weights to every time before its first pole. This yields the free-energy upper bound; a comparison with a max-type recursion and a second-moment estimate for pairs of leaves give the matching lower bound. A Lean 4 formalization of the main results accompanies the paper. \end{abstract} \section{Introduction} \label{sec:introduction} Pinning models describe a path interacting with a defect on which the contact rewards may be random; see \citet{Giacomin07,Giacomin11} for background. We study the diamond hierarchical model of \citet[Section~4.2, (4.10)]{DHV92}, whose partition function satisfies an exact distributional recursion. Our aim is to determine how its quenched free energy vanishes as the contact reward $h$ decreases to the critical point $h_c$. The predicted Berezinskii--Kosterlitz--Thouless (BKT) singularity is an exponential decay in the inverse square root of the distance to criticality \citep{TC01,DR14}. We prove that $\log F(h_c+u)\sim-K/\sqrt u$ as $u\downarrow0$, with $K\in(0,\infty)$, whenever the quenched and annealed critical points differ. We also characterize $K$ through the response of the critical recursion to a change of field. \subsection{The model and main result} Let $B>2$, let $W>0$ satisfy $\E W=1$, and set \begin{equation}\label{eq:model} R_0(h)=e^hW,\qquad R_{n+1}(h)\overset d=\frac{R_n^{(1)}(h)R_n^{(2)}(h)+c}{B}, \qquad c=B-1, \end{equation} where the two variables on the right are independent copies of $R_n(h)$. For integer $B$, this is the normalized partition function on a diamond lattice of longitudinal size $2^n$, with random rewards on a distinguished path; the recursion also makes sense for every real $B>2$ \citep[Section~1.2]{GLT10}. The field $h$ controls the contact reward, while the law of $W$ is fixed. The quenched free energy and its critical point are \[ F(h)=\lim_{n\to\infty}2^{-n}\E\log R_n(h),\qquad h_c=\sup\{h:F(h)=0\}. \] The annealed model follows the deterministic recursion for $\E R_n$ and has critical point $h_a=\log c$ \citep[Theorem~1.2]{GLT10}. The standard existence and finite-volume bounds for the quenched free energy \citep[Theorem~1.1 and Appendix~A.1]{GLT10} are recalled in Lemma~\ref{lem:basic-bounds}, with the elementary adaptation needed for the present finite-mean assumption. In particular, $F\ge0$ and $h_a\le h_c<\infty$. Our assumption $h_c>h_a$ means that disorder shifts the critical point. To state the constant, write $a=\log B$ and $\theta=\log2/a\in(0,1)$. Unless a field is specified, all moments below are evaluated at $h_c$. Set $A_n=\E R_n^\theta$, $\Pi_n=\prod_{jh_a$. The limits \begin{equation}\label{eq:constants} \ell_3=\lim_{n\to\infty}\frac{J_n}{\Pi_n},\qquad \kappa=\lim_{n\to\infty}\frac{\partial_hH_n(h_c)}{\Pi_n} \end{equation} exist and lie in $(0,\infty)$. Moreover, $\Pi_n\sim 3a^3n^2/(2\ell_3)$. With \begin{equation}\label{eq:gamma} \gamma=\frac{3a^2\kappa}{2\ell_3} =\lim_{n\to\infty}\frac{\partial_hH_n(h_c)}{an^2}, \end{equation} one has \begin{equation}\label{eq:bkt} \lim_{u\downarrow0}\sqrt u\,\log F(h_c+u) =-\frac{\pi\log2}{\sqrt\gamma}. \end{equation} \end{theorem} Thus the constant in the BKT law is $K=\pi\log2/\sqrt\gamma$. The assumption $\E W=1$ normalizes a finite positive mean. It supplies every fixed tilted logarithmic moment $\E[R_n^\theta|\log R_n|^q]$ with $q\ge0$, including at $n=0$, as shown in Lemma~\ref{lem:basic-bounds}; no additional logarithmic moment assumption is imposed. For $W=e^{\beta\omega-\beta^2/2}$ with $\omega$ standard normal and $\beta>0$, the strict inequality $h_c>h_a$ holds whenever $B>2+\sqrt2$ \citep[Theorem~1.4(2)]{GLT10}. The case $h_c=h_a$ is outside the theorem. \subsection{Relation to earlier work} The rigorous framework for \eqref{eq:model} was developed by \citet{GLT10}. Related results include the hierarchical smoothing bound for Gaussian disorder of \citet[Theorem~3.1 of the author version]{LT09} and Gaussian marginal relevance at $B=2+\sqrt2$, in the present normalization, proved by \citet[Remark~2.1 and Theorem~2.3]{GLT10m}. \citet{TC01} predicted a BKT transition through rare-contact renormalization, and \citet[Section~4]{DR14} supplied numerical evidence for \eqref{eq:model}. \citet[Sections~II--III]{MG08} investigated a competing scenario with a broad critical law and a finite correlation-length exponent. The link with max-type recursions is seen in logarithmic coordinates. Writing $X_n=\log R_n/a$, equation~\eqref{eq:model} becomes \begin{equation}\label{eq:log-model} X_{n+1}\overset d=\frac1a\log\!\left( B^{X_n^{(1)}+X_n^{(2)}-1}+1-B^{-1}\right). \end{equation} This is a smooth analogue of the \emph{hard recursion} $Y_{n+1}=(Y_n^{(1)}+Y_n^{(2)}-1)^+$. We refer to \eqref{eq:model} as the \emph{soft recursion}. The hard recursion was studied in the spin-glass setting by \citet{CEGM84a,CEGM84b} and subsequently by \citet{DR14}; see also \citet[Section~1]{DS20}. Unlike $Y_n$, the logarithmic variable $X_n$ need not be nonnegative; the rigorous comparison in Section~\ref{escape:section} uses a deterministic lower bound for $R_n$. The identity $R_n^\theta=2^{X_n}$ gives $H_n=a\E[2^{X_n}(X_n-1)]$ and $J_n=a^3\E[2^{X_n}(X_n-1)^3]$. For nonnegative integer-valued hard laws with $\P(Y_0\ge2)>0$ and $\E[Y_0 2^{Y_0}]<\infty$, the condition $\E[(Y_0-1)2^{Y_0}]=0$ describes the critical manifold \citep{CEGM84b}. This explains the tilted moments used here; the soft recursion introduces correction terms in their evolution. For the integer-valued hard recursion, \citet[Theorem~1.1]{CDDHLS21} prove the exponent $1/2$ as $p\downarrow p_c$ along the family $(1-p)\delta_0+p\mathcal L(Y_0^\star)$, where $Y_0^\star$ is positive and integer-valued, $\P(Y_0^\star\ge2)>0$ and $\E[(Y_0^\star)^3 2^{Y_0^\star}]<\infty$, without identifying a limiting constant. Other integrability regimes can lead to different exponents; see \citet[Theorem~1.2]{CDDHLS21} for positive critical parameter, \citet{HS18} for regimes with zero critical parameter, and the survey of \citet[Sections~6.2 and~8]{DS20}. In the present logarithmic coordinates, the finite-mean assumption is $\E B^{X_0}<\infty$, with $B>2$, so it already supplies an exponential moment beyond the tilt $2^{X_0}$. Precise BKT constants were obtained by \citet[Theorem~1.10(i)--(ii)]{HMP20} for an exactly solvable continuous-time family. For discrete-time models with geometric offspring, \citet[Theorems~1.1 and~1.3]{LZ25} study geometric-type invariant laws, and \citet[Theorems~1.3, 2.6(b) and 2.12(b)]{AHM26} establish precise constants for solvable families. These results concern recursions different from \eqref{eq:model}. Theorem~\ref{th:bkt} treats the diamond recursion itself, for arbitrary positive finite-mean disorder with a strict critical-point shift. \subsection{Main ideas and organization} The proof relates the response at criticality to the number of generations over which a small field perturbation changes the recursion. For the binary hard recursion, critical product bounds and the cubic-moment method go back to \citet[Appendix~III of the 1983 archival version]{CEGM84b}; see \citet[Section~2]{CDHLS19} for the general-branching treatment. Here estimates on the Mellin transforms $G_n(q)=\E R_n^q$, including the soft correction terms, give the critical product scale $\Pi_n\asymp n^2$. Moments of the Gibbs contact count, the number of rewarded leaves in the partition-function expansion, control changes of field in the window $h-h_c=\lambda/N^2$. At rescaled time $t=n/N$, these estimates give compactness of the rescaled tilted measures, which weight the law of $\log R_n/(aN)$ by $N(R_n^\theta-1)^+$; see \eqref{scaling:measure}. They also determine the measures' behavior as $t\downarrow0$, which we call the \emph{entrance data}. The field response fixes the normalization $\gamma$. The limiting coalescence equation and its exponential profiles were identified by \citet{DR14} and \citet{CDDS20}. We select the profile from the entrance data by an entire-function argument. For $\lambda>0$, put $v=\gamma\lambda$; its first pole occurs at rescaled time $T_v=\pi/\sqrt v$. Two points require additional control for the soft recursion: the additive correction in \eqref{eq:model} must be estimated uniformly along the near-critical orbit, and the unweighted mass of the limiting profile becomes infinite before the pole. We therefore extend the convergence in exponential weights. The weighted convergence gives the free-energy upper bound. For the lower bound, a comparison with the hard recursion shows that sufficient weighted mass in one bounded interval forces growth within a controlled number of generations. The pair-counting argument follows the two-leaf amplification method of \citet[Section~4.1]{HS18}. For $u=\lambda/N^2$, the pole corresponds to the energy scale \[ NT_v=\frac{\pi}{\sqrt{\gamma u}},\qquad 2^{-NT_v}=\exp\!\left(-\frac{\pi\log2}{\sqrt{\gamma u}}\right). \] The matching bounds turn this scale into the limit in \eqref{eq:bkt}. Sections~\ref{sec:critical}--\ref{sec:response} establish the critical and field estimates. The proof of the scaling limit proceeds through compactness (Section~\ref{scaling:section}), identification of the profile (Section~\ref{sec:profile-identification}), and continuation before the pole (Section~\ref{cont:section}). Section~\ref{escape:section} proves both free-energy bounds and concludes the proof of Theorem~\ref{th:bkt}. \subsection{AI use} This work was developed through extensive, iterative interaction with OpenAI’s ChatGPT, including the use of multiple AI agents. ChatGPT proposed proof strategies and generated and revised substantial portions of the mathematical arguments and manuscript text. This contribution included the central argument for the Berezinskii–Kosterlitz–Thouless asymptotics of the free energy in the disordered hierarchical pinning model on a diamond lattice. The author selected the research problem and directed the investigation and successive revisions, including the simplification of arguments, organization of the manuscript and attribution of prior work. ChatGPT also assisted with literature searches and reference checks. Claude was used for editorial and bibliographic audits. All numbered results of the paper (the Gaussian example of the introduction excepted) were then formalized in Lean~4 with the Mathlib library, entirely by AI agents, in about $43\,000$ lines of Lean. The formalization is available at \url{https://github.com/igor-kortchemski/BKT-hierarchical-pinning}. The formalization was coordinated by Anthropic’s Claude Code running Claude Opus~5.5, which delegated statements and proofs to agents running Claude Opus~5.5 and Claude Sonnet~5.5, OpenAI’s GPT-6.2 Sol, GPT-6 Luna and GPT-6.2 Astra (through Codex), and Moonshot AI’s Kimi K3.1 Pro (through Open Math OS). The author fixed the scope and the conventions of the formal statements, and delegated the approval of the Lean statements to GPT-6.2 Astra. Each proof was re-audited by a second model, usually of another family, for fidelity to the paper. The main theorem was checked with Comparator against a self-contained restatement depending only on Mathlib, by Lean’s kernel and the independent checkers NanoDa, Lean4Lean, con-leche and con-ron, with no axioms beyond \texttt{propext}, \texttt{Classical.choice} and \texttt{Quot.sound}. No human has reviewed the Lean statements or proofs line by line. The formalization took about twelve hours of wall-clock time on 18–19 November 2026, largely unattended. It used about $1.5$ billion tokens: Claude Opus~5.5 $0.73$ billion, Kimi K3.1 Pro $0.36$ billion, GPT-6.2 Sol $0.21$ billion, GPT-6 Luna $0.16$ billion, Claude Sonnet~5.5 $15$ million and GPT-6.2 Astra $14$ million. Almost all of these are input tokens, the agents' context re-read at each step and mostly served from cache. Output tokens, reasoning included, number about $6$ million. % Begin critical.tex \section{Basic bounds and critical moments} \label{sec:critical} \subsection{Floor and finite-volume bounds} Here $h\in\mathbb R$ is arbitrary and $R_n$ denotes $R_n(h)$. We recall the finite-volume argument of \citet[Theorem~1.1 and Appendix~A.1]{GLT10}, keeping the deterministic floor explicit under the present assumption $\E W=1$. Let $\phi_B(x)=(x^2+c)/B$, $b_0=0$ and $b_{n+1}=\phi_B(b_n)$. We write $\rho=2/B\in(0,1)$. \begin{lemma}\label{lem:basic-bounds} We have $R_n\ge b_n\uparrow1$ and $1-b_n=O(\rho^n)$. For $n\ge0$, $\E R_n<\infty$ and $\E[R_n^\theta|\log R_n|^q]<\infty$ for every fixed $q\ge0$. The free energy $F$ exists, is nonnegative and satisfies \begin{equation}\label{eq:finitesize} 2^{-n}(\E\log R_n-a)\le F(h) \le2^{-n}(\E\log R_n-\log b_n),\qquad n\ge1. \end{equation} Moreover, $0\le F(h+u)-F(h)\le u$ for $u\ge0$, $h_a\le h_c<\infty$ and $F(h_c)=0$. \end{lemma} \begin{proof} Monotonicity of \eqref{eq:model} gives $R_n\ge b_n\uparrow1$. Since $1-b_{n+1}=B^{-1}(1-b_n)(1+b_n)\le\rho(1-b_n)$, we obtain the stated rate. The recursion preserves a finite first moment. The tilted logarithmic moments are finite because $r^\theta|\log r|^q$ is bounded near zero and is $O(r)$ as $r\to\infty$; this also covers $n=0$. The two inequalities $\log R_{n+1}\ge\log R_n^{(1)}+\log R_n^{(2)}-a$ and $R_{n+1}/b_{n+1}\le(R_n^{(1)}/b_n)(R_n^{(2)}/b_n)$ show that the left-hand side of \eqref{eq:finitesize} increases with $n$ and the right-hand side decreases. Their difference tends to zero, proving existence of $F$ and \eqref{eq:finitesize}. The floor gives $F\ge0$. The partition function is multiaffine in its $2^n$ leaf weights, with nonnegative coefficients. Hence $R_n(h)\le R_n(h+u)\le e^{2^nu}R_n(h)$ for $u\ge0$, which gives the asserted Lipschitz bound. At $h=h_a$, $\E R_n=c$ for every $n$, so Jensen's inequality gives $F(h_a)=0$. As $h\to\infty$, $\log R_1(h)$ increases to infinity and is bounded below by $\log b_1$. Thus $\E\log R_1(h)\to\infty$, and \eqref{eq:finitesize} yields $h_c<\infty$. Continuity now gives $F(h_c)=0$. \end{proof} For $S_n=R_n^{(1)}R_n^{(2)}$ and $n\ge1$, set \[ C_n(z)=\E[(S_n+c)^z-S_n^z],\qquad \Re z<1. \] This correction term records the contribution of the additive constant in \eqref{eq:model}. \begin{lemma}\label{lem:mellin-correction} The functions $C_n$, $n\ge1$, are holomorphic and uniformly bounded on compact subsets of $\{\Re z<1\}$. They satisfy \begin{equation}\label{eq:C} C_n(z)=z\int_0^c\E(S_n+t)^{z-1}\dd t. \end{equation} \end{lemma} \begin{proof} Integrate the derivative of $(S_n+t)^z$ over $t\in[0,c]$ to obtain \eqref{eq:C}. Since $S_n+t\ge b_1^2>0$, the integrand and its $z$-derivatives are uniformly bounded on each compact subset of $\{\Re z<1\}$. Differentiation under the integral and expectation proves the assertion. \end{proof} \subsection{Critical moment bounds} Write $G_n(q)=\E R_n^q$. We first control the fractional moment at $\theta$ and its logarithmic derivative when $F(h)=0$. \begin{proposition}\label{prop:criticalmoments} If $F(h)=0$, then, for every $n\ge1$, \begin{equation}\label{eq:criticalmoments} G_n(\theta)\le c^\theta,\qquad H_n:=\E[R_n^\theta(\log R_n-a)]\le C_B\rho^n. \end{equation} Moreover, $\sup_n\E[R_n^\theta|\log R_n-a|]<\infty$. \end{proposition} \begin{proof} \emph{Fractional moments.} We first obtain a tail bound below $\theta$ and then use the moment recursion to reach $\theta$. Fix $n$, $t\ge b_n$ and $p=\P(R_n\ge t)>0$. In a tree of $2^k$ independent height-$n$ blocks, retain the favorable blocks and their ancestors. In units of $\log(R/b_n)$, each retained block contributes at least $\log(t/b_n)$ and each ancestor costs at most $\log(B/b_n)$. The expected number of retained ancestors at height $j$ above the blocks is at most $2^{k-j}\min(1,2^jp)$. Dividing by $2^k$ and letting $k\to\infty$, we obtain, since $F=0$, \[ p\log(t/b_n) \le\log(B/b_n)\sum_{j\ge1}\min(2^{-j},p) \le p\log(B/b_n)\bigl(\log_2(1/p)+2\bigr). \] Thus, with $\theta_n=\log2/\log(B/b_n)\uparrow\theta$, \begin{equation}\label{eq:tail} \P(R_n\ge t)\le\min\{1,4(b_n/t)^{\theta_n}\}. \end{equation} This is a floor-normalized adaptation of the favorable-block argument for the hierarchical model in \citet[proof of Theorem~3.1, (3.7)--(3.10)]{LT09}. The earlier rare-stretch argument appears in \citet[Section~3.1]{GT06}, and the rare-contact mechanism in \citet{TC01}. In particular, $\sup_nG_n(q)<\infty$ for each $0c$, this recursion would force divergence, contrary to the tail bound. Thus $G_n(q)\le c^q$ for $q<\theta$. Letting $q\uparrow\theta$ proves the first assertion. \emph{The endpoint logarithmic moment.} Set $H_n(q)=\E[R_n^q(\log R_n-a)]$ and $\psi_q(r)=(r+c)^q(\log(r+c)-2a)-r^q(\log r-2a)$. The exact identity is \begin{equation}\label{eq:H} H_{n+1}(q)=2B^{-q}G_n(q)H_n(q) +B^{-q}\E\psi_q(R_n^{(1)}R_n^{(2)}). \end{equation} For $01$, the difference $\psi_q$ increases on $[1,r_0]$ from $\psi_q(1)=a(2-B^q)\ge0$; beyond $r_0$ the claim is immediate. Using $R_n\ge b_n$ and $1-b_n=O(\rho^n)$, we obtain, uniformly for $q\in[\theta/2,\theta]$, $H_{n+1}(q)\ge d_qG_n(q)H_n(q)-C_B\rho^n$, where $d_q=2B^{-q}$. For fixed $q<\theta$, $H_j(q)$ is bounded by~\eqref{eq:tail}, whereas \[ \prod_{i=n}^{j-1}d_qG_i(q)\ge c_Bd_q^{j-n},\qquad d_q>1. \] We divide by these products and sum to infinity. The error is at most $C_B\rho^n\sum_{k\ge0}(\rho/d_q)^k$, uniformly as $q\uparrow\theta$. Hence $H_n(q)\le C_B'\rho^n$. Monotone convergence on $R_n\ge B$ and bounded convergence for the negative part give the endpoint bound. The bounded negative part also yields the asserted absolute moment. \end{proof} \subsection{The critical barrier and returns to one} From now on, we assume $h=h_c>h_a$. The following barrier is a floor-centered version of the fractional-moment argument in \citet[Proposition~4.1, (4.1)--(4.3)]{GLT10}. \begin{lemma}[Critical barrier]\label{lem:barrier} For $\thetah_c$, a contradiction. \end{proof} \begin{theorem}\label{th:return} If $h_c>h_a$, then \begin{equation}\label{eq:return} \liminf_{n\to\infty}G_n(\theta)=1. \end{equation} \end{theorem} \begin{proof} We first rule out uniform moments above $\theta$, then apply the same continuation argument at $\theta$. \emph{A uniform continuation criterion.} For a related continuation argument for fixed laws of the hard recursion, see \citet[Appendix~B]{DR14}. The family $C(z)=(C_n(z))_{n\ge1}$ defined in~\eqref{eq:C} is holomorphic with values in $\ell^\infty$ on $\Re z<1$ and bounded on compact subsets. Each $G_n$ is holomorphic there as well, because $\E R_n<\infty$, and \begin{equation}\label{eq:inverse} G_n(z)^2=B^zG_{n+1}(z)-C_n(z). \end{equation} Suppose $g=(G_n(q_0))_{n\ge N}$ is bounded, bounded away from zero, and \begin{equation}\label{eq:inversecontraction} \sup_{n\ge N}\frac{B^{q_0}}{2G_n(q_0)}<1. \end{equation} The map $(\mathcal K_zv)_n=\sqrt{B^zv_{n+1}-C_n(z)}$, with the positive branch at $(q_0,g)$, is a uniform contraction in a small $\ell^\infty$ ball: its radicands stay away from zero. Its fixed point is holomorphic in $z$ and locally unique. If $G(q)\to g$ in $\ell^\infty$ as $q\uparrow q_0$, it agrees with this fixed point on the left. The coordinatewise identity theorem then gives a uniform moment bound to the right of $q_0$. \emph{No uniform moment above $\theta$.} Suppose such a moment exists and set \[ q_* =\sup\{q\in(0,1):\sup_{n\ge1}G_n(q)<\infty\}>\theta. \] For $q2$, the neighborhood and $\varepsilon$ can be chosen so that \[ \frac{B^{q_2}}{G_n(q_1)+G_n(q_2)}\le\eta<1 \qquad(q_1c$. \emph{Returns to one.} If $G_n(\theta)\ge1+\delta$ for all large $n$, \eqref{eq:inversecontraction} holds at $q_0=\theta$, since $B^\theta=2$. Proposition~\ref{prop:criticalmoments} gives $\|G(q)-G(\theta)\|_\infty\le C(\theta-q)$ near $\theta$. The criterion would produce a uniform moment above $\theta$. Together with $G_n(\theta)\ge b_n^\theta\to1$, this proves~\eqref{eq:return}. \end{proof} % End critical.tex % Begin product.tex \section{The critical product} \label{sec:product} For the binary hard recursion, the critical product bounds and cubic-moment method already appear in \citet[Appendix~III of the 1983 archival version]{CEGM84b}. The general-branching treatment and further critical estimates are developed in \citet[Section~2, Theorem~3 and Propositions~1--2]{CDHLS19}. Here the moment recurrences retain the soft correction kernels. For the diamond model, \citet[Theorem~1.6]{GLT10} prove tightness at criticality and convergence to one strictly below it; the argument below treats convergence at $h_c>h_a$. At $h=h_c>h_a$, set \[ A_n=\E R_n^\theta,\quad \Pi_n=\prod_{j0$, $\Pi_n$ tends either to a positive finite limit or to infinity. Write $\alpha_n=A_n-1$, $d(x)=1-(x+c)^\theta+x^\theta$ and $D_n=\E d(U_n)$. On $x\ge1$, $d(x)\ge0$ and $0\le\omega_1(x)\le C_Bd(x)$. The sign follows from the proof of Proposition~\ref{prop:criticalmoments}. For the comparison, we differentiate near one and use $d(x)\to1$, $\omega_1(x)\to0$ at infinity. The floor yields \begin{equation}\label{eq:prodsource} D_n\ge-C_B\rho^n,\qquad |\E\omega_1(U_n)|\le C_BD_n+C_B'\rho^n. \end{equation} Since $\alpha_{n+1}=\alpha_n+\alpha_n^2/2-D_n/2$, telescoping gives \begin{equation}\label{eq:prodtele} \sum_{j=n}^m\frac{\Pi_nD_j}{\Pi_{j+1}} =2\alpha_n-\sum_{j=n}^m\frac{\Pi_n\alpha_j^2}{\Pi_{j+1}} -\frac{2\Pi_n\alpha_{m+1}}{\Pi_{m+1}}. \end{equation} \emph{The product diverges.} Suppose $\Pi_n\to\Pi_\infty<\infty$. Its negative logarithmic increments are summable, so $\sum\alpha_n^+<\infty$ and $R_n\to1$ in probability. By~\eqref{eq:prodsource}, the bounded sequence $H_n/\Pi_n$ has summable negative increments. Thus $H_n\to h\in[-a,0]$; the lower bound follows from $r^\theta(\log r-a)\ge-a$ on $r\ge1$ and the floor. If $h>-a$, \eqref{eq:prodK} would make the increments of $K_n/\Pi_n$ tend to $(h^2-a^2)/\Pi_\infty<0$, contradicting $K_n\ge0$. Consequently $M_n:=\E[R_n^\theta\log R_n]=H_n+aA_n\to0$. For small $t\ge0$, the nonnegative Taylor remainder $E_n(t)=G_n(\theta+t)-A_n-tM_n$ satisfies \[ E_{n+1}(t)\le(A_n+Ct)E_n(t)+CE_n(t)^2+Ct^2. \] Indeed, we expand~\eqref{eq:inverse}; the first two derivatives of $C_n$ are uniformly bounded near $\theta$. After division by the bounded product, $e_{n+1}\le(1+Ct)e_n+Ce_n^2+Ct^2$ and $e_1=O(t^2)$. Choosing a large constant and then a small $\tau$, we obtain by induction \begin{equation}\label{eq:prodtaylor} E_n(t)\le C(n+1)t^2\qquad((n+1)t\le\tau). \end{equation} The induction costs $O(\tau t^2)$ from $te_n$ and $O(\tau^2t^2)$ from $e_n^2$ at each step. The summable sequence $d_n=\alpha_n^++\rho^{n/2}$ has a subsequence along which $nd_n\to0$. Fix $k>1/(2a)$ and put $t_n=kd_n$. Along this subsequence, \[ G_n(\theta+t_n)\le1+d_n+t_nM_n+C(n+1)t_n^2 =1+d_n+o(d_n), \] whereas~\eqref{eq:barrier} gives $G_n(\theta+t_n)\ge1+2akd_n-C_B\rho^n$. This contradiction proves $\Pi_n\to\infty$. We can now let $m\to\infty$ in~\eqref{eq:prodtele}. All ratios $\Pi_{j+1}/\Pi_n$ are uniformly bounded below, so \eqref{eq:prodsource} gives \[ \sum_{j\ge n}\frac{\Pi_n|\E\omega_1(U_j)|}{\Pi_{j+1}} \le C_B(\alpha_n^++\rho^n). \] Summing~\eqref{eq:prodH} backwards, with $H_n/\Pi_n\to0$, proves~\eqref{eq:Hsmall}. \emph{A summable inverse product.} To use an increasing product, we remove the floor by setting \[ 1+z_n=A_n/b_n^\theta,\qquad P_n=\prod_{j=1}^{n-1}(1+z_j)\asymp\Pi_n, \qquad T_n=1+\sum_{j=1}^{n-1}P_j^{-1}. \] The sequence $P_n$ is increasing, $P_n\ge1$, and its ratios are bounded. Equations~\eqref{eq:prodK}--\eqref{eq:prodJ} imply $K_n/\Pi_n\le CT_n$ and $J_n/\Pi_n\le CT_n$: the only additional positive term is $O(\rho^nT_n)$, summable since $T_n\le n$. Cauchy--Schwarz and the bounded negative part of $Z_n$ give \begin{equation}\label{eq:prodKT} K_n^2\le C(J_n+C_B)\le CP_nT_n. \end{equation} For small $z_n$ and large $n$, \eqref{eq:Hsmall} gives $\E(V\log(R_n/b_n))\ge a/2$, where $V=(R_n/b_n)^\theta$. Since $\E(V-1)=z_n$ and $\E\log(R_n/b_n)\le z_n/\theta$, Cauchy--Schwarz under the positive measure $(V-1)\dd\P$ yields \[ (a/4)^2\le z_n\E[(V-1)\log^2(R_n/b_n)] \le Cz_n(K_n+1). \] For larger $z_n$ the same lower bound is immediate. Hence, from a fixed index onwards, \begin{equation}\label{eq:prodNash} z_n\ge\frac{c_0}{K_n+1}\ge\frac{c_1}{\sqrt{P_nT_n}}. \end{equation} For $x_n=\sqrt{P_n}$ this gives \[ x_{n+1}-x_n\ge cT_n^{-1/2},\qquad T_{n+1}-T_n=x_n^{-2},\qquad x_n^{-1}-x_{n+1}^{-1} \ge c\frac{T_{n+1}-T_n}{\sqrt{T_n}}. \] Summing the last inequality bounds $\sqrt{T_n}$, hence $T_n$. It follows that $x_n\ge cn$, $\sum\Pi_n^{-1}<\infty$ and $K_n^2\le C\Pi_n$. Also $J_n\le C\Pi_n$. \emph{The cubic residue is positive.} The positive increments of $J_n/\Pi_n$ are summable by \eqref{eq:prodJ}, since $H_n^+\le C\rho^n$ and $\sum\Pi_n^{-1}<\infty$. As $J_n\ge-C_B$, this sequence converges to $\ell_3\ge0$. Suppose $\ell_3=0$. Summing backwards gives \begin{equation}\label{eq:prodretroJ} J_n\le CP_n\sum_{j\ge n} \frac{(z_j+\rho^j)K_j+1}{P_{j+1}}. \end{equation} Assume $K_n\le MP_n^g$ for $n\ge N$, where $M\ge1$ and $00$. \emph{The matching upper bound.} Since $J_n^+\le C\Pi_n$, $J_n^-\le C_B$ and $H_n^+\le C\rho^n$, we have $H_nJ_n\le C\rho^n\Pi_n+C_B$. Dividing~\eqref{eq:prodL} by $\Pi_{n+1}$ and using $K_n^2\le C\Pi_n$, we obtain $L_n\le C(n+1)\Pi_n$. The positive cubic residue and Cauchy--Schwarz now give \[ c\Pi_n^2\le J_n^2\le K_nL_n \le C(n+1)\Pi_n^{3/2}. \] Thus $\Pi_n\le C(n+1)^2$, proving~\eqref{eq:product} and \eqref{eq:critical-fourth}. \emph{Rates.} Subtracting the floor recurrence, as in~\eqref{eq:contraction}, gives \[ z_{n+1}\le\frac{b_n^{2\theta}}{b_{n+1}^\theta} (z_n+z_n^2/2)\le z_n+z_n^2/2. \] Reading backwards gives $z_j\ge z_n/[1+(n-j)z_n/2]$. On the other hand, bounded $z_j$ and $P_n\asymp n^2$ imply \[ \sum_{j=\lfloor n/2\rfloor}^nz_j \le C\log\frac{P_{n+1}}{P_{\lfloor n/2\rfloor}}\le C'. \] Summing the lower bound yields $2\log(1+nz_n/4)\le C''$. Hence $z_n=O(n^{-1})$, $A_n-1=O(n^{-1})$ and, by \eqref{eq:Hsmall}, $H_n=O(n^{-1})$. The floor and $\E(R_n^\theta-b_n^\theta)\to0$ imply $R_n\to1$. Finally, $K_n^2\le C\Pi_n\le Cn^2$ gives $K_n\le Cn$, while \eqref{eq:prodNash} and $z_n\le C/n$ give $K_n+1\ge c/z_n\ge cn$. Thus $K_n\asymp n$, and \eqref{eq:prodNash} also gives $z_n\ge c/n$. Since $1-b_n^\theta=O(\rho^n)$, this proves~\eqref{eq:criticalrates}. \end{proof} % End product.tex % Begin response.tex \section{Field response and contact moments} \label{sec:response} Unless a field is specified, all quantities are evaluated at $h_c>h_a$. Set \[ Q_n=\left.\partial_h\log R_n(h)\right|_{h_c},\qquad \mathcal D_n=\left.\partial_h\E R_n(h)^\theta\right|_{h_c},\qquad \mathcal E_n=\left.\partial_hH_n(h)\right|_{h_c}. \] We will use the critical estimates of Theorem~\ref{th:product}: \begin{equation}\label{response:inputs} \begin{gathered} \Pi_n\asymp(n+1)^2,\qquad K_n\asymp n+1,\qquad |H_n|\le C(n+1)^{-1},\qquad H_n^+\le C\rho^n,\\ \E[R_n^\theta|\log R_n-a|^4]\le C(n+1)^3. \end{gathered} \end{equation} \subsection{The spine and a passage estimate} The distinguished-leaf change of measure uses the classical spine viewpoint \citep[Section~2]{LPP95}, already applied to the hard recursion by \citet[Section~4, (4.1)--(4.3)]{HS18}. It represents the field derivatives in terms of a path with independent increments. The following passage estimate will control its minimum. \begin{lemma}\label{response:ballot} Let $Y_j$ be independent integrable random variables and $S_n=\sum_{j0$ for all sufficiently large $n$. Then \[ \P\left(\min_{0\le k\le n}S_k\ge-z\right) \le\frac{C(z+1)}{n+1},\qquad z\ge0. \] \end{lemma} \begin{proof} For $\tau=\inf\{k\ge1:S_k<-z\}$, independence gives $\E S_{\tau\wedge n}=\sum_{jj)\le C$. The overshoot below $-z$ is at most $b$, except at the first step, whose expected cost is bounded by $\E Y_0^-$. Hence \[ \E[(S_n+z+b)\mathbf1_{\{\tau>n\}}]\le C(z+1). \] Both $\{\tau>n\}$ and $\{S_n\ge\eta n\}$ are increasing events in the independent increments. Harris's inequality \citep[Lemma~4.1]{Harris60}, in its standard product-measure form, bounds the left-hand side below by $\eta n\delta\,\P(\tau>n)$. Adjusting $C$ covers the remaining $n$. \end{proof} \begin{lemma}[Spine identities]\label{response:spine-identities} Under $\widehat\P$, let $R_0^*,U_0,U_1,\ldots$ be independent, with respective laws $r^\theta\P(R_0\in\dd r)/A_0$ and $r^\theta\P(R_j\in\dd r)/A_j$. Set $S_k=\sum_{j0$. Then \begin{align} \frac{\mathcal D_n}{\Pi_n} &=\theta A_0\widehat\E w_n,\qquad w_n=\left(1+\frac c{BR_0^*}\sum_{k=1}^ne^{-S_k}\right)^{-p}, \label{response:spine}\\ \frac{\mathcal E_n}{\Pi_n} &=\theta A_0\widehat\E[w_n(\log T_n+d_0)],\qquad T_n=R_0^*e^{S_n}w_n^{-1/p}. \label{response:mixed-spine} \end{align} \end{lemma} \begin{proof} We first justify the field differentiations used in this section. At fixed height $n$, on the event that all leaves are positive, $R_n(h)=\sum_{k=0}^{2^n}c_{n,k}e^{kh}$ with nonnegative coefficients. For $x(h)=\log R_n(h)$, the probability weights proportional to $c_{n,k}e^{kh}$ give $0\le x'(h)\le2^n$ and $0\le x''(h)\le4^n$. Thus $|x(h)-x(h_0)|\le2^n$ on $|h-h_0|\le1$. For each integer $j\ge0$, the first two field derivatives of $e^{\theta x(h)}(x(h)-a)^j$ are bounded on this interval by \[ C_{n,j}R_n(h_0)^\theta(1+|\log R_n(h_0)-a|)^j. \] The same bound with exponent $7/2$ applies to $e^{\theta x(h)}(1+(x(h)-a)^2)^{7/4}$. These envelopes are integrable by Lemma~\ref{lem:basic-bounds}, including when $n=0$. Dominated convergence therefore permits the first two field derivatives to pass under expectation at every $h_0$. Under the original law, distinguish one leaf and let $U_j$ be the mass of its sibling subtree at height $j$. Along the distinguished path, $T_0=R_0$ and $T_{j+1}=(T_jU_j+c)/B$. Iterating gives \[ T_n=e^{S_n}\left(R_0+\frac cB\sum_{k=1}^ne^{-S_k}\right), \qquad R_0\partial_{R_0}T_n=R_0e^{S_n}. \] By multiaffinity and symmetry of the $2^n$ leaves, $\mathcal D_n=\theta2^n\E[T_n^{\theta-1}R_0e^{S_n}]$. Differentiating $H_n$ inserts the factor $\log T_n+d_0$ in this expectation, with differentiation justified by the preceding envelopes. Changing measure and using $B^\theta=2$ gives both identities. The leaf factor $A_0$ is separate from the $U_0$ factor in $\Pi_n$: $R_0^*$ and $U_0$ are independent tilted copies. \end{proof} \begin{lemma}\label{response:linear} For all $n$, $\mathcal D_n\le C(n+1)$. \end{lemma} \begin{proof} Under the tilted law, $Y_j=\log U_j-a$ satisfies \[ \widehat\E Y_j=H_j/A_j,\qquad \widehat\E Y_j^2=K_j/A_j,\qquad \widehat\E Y_j^4\le C(j+1)^3. \] The floor bounds $Y_j$ below for $j\ge1$, and \eqref{response:inputs} yields \[ -C\log(n+1)\le\widehat\E S_n\le C,\qquad \Var(S_n)\asymp n^2,\qquad \widehat\E(S_n-\widehat\E S_n)^4\le Cn^4. \] For the centered variable $Z=S_n-\widehat\E S_n$, $\widehat\E|Z|\ge(\widehat\E Z^2)^{3/2}/(\widehat\E Z^4)^{1/2}$. Thus $\widehat\E Z^+=\widehat\E|Z|/2\ge cn$, and Paley--Zygmund applied to $Z^+$ gives $\widehat\P(S_n\ge\eta n)\ge\delta>0$. Lemma~\ref{response:ballot} therefore applies. With $M_n=-\min_{0\le k\le n}S_k\ge0$, \[ \widehat\E e^{-pM_n} =p\int_0^\infty e^{-pz}\widehat\P(M_n\le z)\,\dd z \le\frac C{n+1},\qquad w_n\le\max\{1,(BR_0^*/c)^p\}\,e^{-pM_n}. \] The maximum also covers a minimum attained at time zero. Its initial factor is independent of $M_n$ and integrable: $\widehat\E(R_0^*)^p=\E R_0/A_0<\infty$. Now use \eqref{response:spine} and $\Pi_n\le C(n+1)^2$. \end{proof} \subsection{Strict transversality} The relation between field response, escape time and the leading constant also appears in the solvable systems of \citet[Section~5]{AHM26}. We establish the critical response directly for \eqref{eq:model}. \begin{theorem}\label{th:response} There is a constant $\kappa\in(0,\infty)$ such that \begin{equation}\label{response:transversality} \frac{\mathcal E_n}{\Pi_n}\longrightarrow\kappa. \end{equation} In particular, $\mathcal E_n\asymp n^2$. \end{theorem} \begin{proof} \emph{Positive lower bound.} Fix $\ell>1$ and consider $V_k=z+S_k-\ell\log(k+1)$, stopped at $\tau=\inf\{k\ge1:V_k<0\}$. The logarithmic correction is $o(n)$, so the modified increments still satisfy Lemma~\ref{response:ballot}. Their means are bounded below by $-C/(k+1)$. Hence \[ \widehat\P(\tau>k)\le\min\{1,C(z+1)/(k+1)\}. \] Since $V_\tau<0$, stopping in the other direction gives \[ \widehat\E[V_n\mathbf1_{\{\tau>n\}}] \ge z-C\sum_{j\ge0}\frac{\min\{1,C(z+1)/(j+1)\}}{j+1} \ge z-C'\log(z+2)-C'. \] Fix $z$ large enough that this is positive. On $\{\tau>n\}$, $\sum_{k\le n}e^{-S_k}\le e^z\sum_{k\ge1}(k+1)^{-\ell}$. Choose $r_0>0$ with $\widehat\P(R_0^*\ge r_0)>0$. On $\{\tau>n,R_0^*\ge r_0\}$, $w_n$ is bounded below, and $\log T_n+d_0\ge V_n$ for large $n$. The integrand in \eqref{response:mixed-spine} is nonnegative everywhere for such $n$, since $T_n\ge b_n\uparrow1$. Independence therefore gives \begin{equation}\label{response:positive-flux} \liminf_n\mathcal E_n/\Pi_n>0. \end{equation} \emph{Convergence.} For $n\ge1$, let $S=R_n^{(1)}R_n^{(2)}$. Differentiation gives \[ \mathcal D_{n+1}=A_n\mathcal D_n-\mathcal L_n,\qquad \mathcal L_n=\theta\E\left[ S^\theta Q_n^{(1)} \left(1-\left(\frac S{S+c}\right)^p\right)\right]\ge0. \] Thus $\sum_{n\ge1}\mathcal L_n/\Pi_{n+1}<\infty$. For $\psi(s)=(s+c)^\theta\log((s+c)/B^2)-s^\theta\log(s/B^2)$, differentiation with $z=s/(s+c)$ and $z^p\log(1/z)\le(1-z^p)/p$ give \[ s\psi'(s)\le C s^\theta \left(1-\left(\frac s{s+c}\right)^p\right), \qquad s\ge b_1^2. \] The functions $\psi(s)$ and $s\psi'(s)$ are bounded on $[b_1^2,\infty)$, while $0\le\partial_h\log S\le2^{n+1}$. The floor holds at every field, so dominated convergence also justifies differentiating $\E\psi(S)$ locally in $h$. Differentiating $H_{n+1}=A_nH_n+\frac12\E\psi(S)$ consequently yields \[ \mathcal E_{n+1} \le A_n\mathcal E_n+H_n\mathcal D_n+C\mathcal L_n. \] Since $\sum |H_n|\mathcal D_n/\Pi_{n+1}<\infty$, the positive variations of $\mathcal E_n/\Pi_n$ are summable. This sequence is eventually nonnegative, so it has a finite limit, which is positive by \eqref{response:positive-flux}. Since $\Pi_n\asymp n^2$, the last assertion follows. \end{proof} \subsection{Contact moments} We use the contact representation of \citet[Section~1.2 and Appendix~A, (A.11)]{GLT10}. At a fixed environment, the Gibbs measure chooses the product term $S=R_n^{(1)}R_n^{(2)}$ with probability $S/(S+c)$, adding the contact counts of the two independent subtrees; otherwise its count is zero. The initial count is $\mathcal C_0=1$. We write $\langle\cdot\rangle_n$ for the critical Gibbs expectation. Weighted open-path moments \citep[Lemma~3.2]{CHS22} and uniform factorial inductions \citep[Theorems~1.2 and~2.3]{CS21} give hard-model precedents for the estimates below. Our observable is the Gibbs contact count. \begin{proposition}\label{prop:contact-moments} The contact representation satisfies \begin{equation}\label{contacts:change-field} Q_n=\langle\mathcal C_n\rangle_n,\qquad R_n(h_c+u)=R_n(h_c)\langle e^{u\mathcal C_n}\rangle_n. \end{equation} There are $M,L<\infty$ such that \begin{equation}\label{contacts:all-moments} M_{k,n}:=\E[R_n^\theta\langle\mathcal C_n^k\rangle_n] \le k!ML^{k-1}(n+1)^{2k-1},\qquad k\ge1,\ n\ge0. \end{equation} Consequently, if $u\ge0$ and $Lu(n+1)^2<1$, \begin{equation}\label{contacts:exponential} \E[R_n^\theta\langle e^{u\mathcal C_n}-1\rangle_n] \le\frac{Mu(n+1)}{1-Lu(n+1)^2}. \end{equation} The same bound holds for $\E R_n(h_c+u)^\theta-A_n$. \end{proposition} \begin{proof} Induction over the height, using the root choice, gives the second identity in \eqref{contacts:change-field}. Differentiating it at $u=0$ gives the first. The case $k=1$ is Lemma~\ref{response:linear}. The root choice and $B^{-\theta}S(S+c)^{\theta-1}\le S^\theta/2$ give \[ M_{k,n+1}\le A_nM_{k,n} +\frac12\sum_{\ell=1}^{k-1}\binom k\ell M_{\ell,n}M_{k-\ell,n}. \] Divide by $k!\Pi_{n+1}$ and sum. Using $\Pi_n/\Pi_{j+1}\le C_\Pi(n+1)^2/(j+1)^2$, induction in $k$ yields \[ \frac{M_{k,n}}{k!} \le(n+1)^{2k-1} \left(\frac{C_0}{k!} +\frac{C_\Pi(k-1)}{2(2k-3)}M^2L^{k-2}\right),\qquad k\ge2, \] because $\sum_{j0$ such that, for $n\ge1$ and $0\le u(n+1)^2\le\eta$, \begin{align} A_n(h_c+u)&=A_n+O(u(n+1)),\label{response:windowA}\\ H_n(h_c+u)&=H_n+u\mathcal E_n+O(u^2(n+1)^4),\label{response:windowH}\\ K_n(h_c+u)&=K_n+O(u(n+1)^3),\label{response:windowK}\\ J_n(h_c+u)&=J_n+O(u(n+1)^4),\label{response:windowJ}\\ \E[R_n(h_c+u)^\theta|\log R_n(h_c+u)-a|^{7/2}] &\le C(n+1)^{5/2}.\label{response:windowSevenHalves} \end{align} Also, $\Pi_n(h_c+u)/\Pi_n(h_c)=1+O(u(n+1)^2)$. \end{proposition} \begin{proof} We first combine contact and logarithmic moments. Write $\dd\mu_n=R_n(h_c)^\theta\dd\P$ and $Z_n=\log R_n(h_c)-a$. For $s\in\{0,1,2,3,7/2\}$ and integers $r\ge1$, \begin{equation}\label{contacts:mixed} \int(1+|Z_n|)^s\langle\mathcal C_n^r\rangle_n\,\dd\mu_n \le C_s^r r!(n+1)^{2r+s-1}. \end{equation} For $s=0$, use \eqref{contacts:all-moments}. For $s=1,2$, apply Cauchy under the joint environment--Gibbs measure, using logarithmic moments of order two and four. For $s=3,7/2$, use Hölder with exponents $(4,4/3)$ and $(8,8/7)$, respectively, and the fourth logarithmic moment. The factorials are controlled by $((pr)!)^{1/p}\le p^r r!$. Expanding the exponential therefore gives, for $d=1,2$, \begin{equation}\label{contacts:mixed-exponential} \int(1+|Z_n|)^s \langle\mathcal C_n^d e^{v\mathcal C_n}\rangle_n\,\dd\mu_n \le C(n+1)^{2d+s-1} \end{equation} whenever $v(n+1)^2\le\eta$; here $\sum_{k\ge0}(k+d)!x^k/k!=d!/(1-x)^{d+1}$. We now transfer these estimates to the perturbed field. Put $T(v)=\langle e^{v\mathcal C_n}\rangle_n\ge1$. The change of Gibbs measure gives \[ R_n(h_c+v)^\theta\langle\mathcal C_n^d\rangle_{n,h_c+v} =R_n(h_c)^\theta T(v)^{\theta-1} \langle\mathcal C_n^d e^{v\mathcal C_n}\rangle_n. \] Since $t^{\theta-1}(1+\log t)^s$ is bounded on $[1,\infty)$, \eqref{contacts:mixed-exponential} also bounds this expression with the additional factor $(1+|\log R_n(h_c+v)-a|)^s$. Now $\partial_h\log R_n=\langle\mathcal C_n\rangle$ and $\partial_h^2\log R_n=\Var(\mathcal C_n)$ under Gibbs. The fixed-height envelopes in the proof of Lemma~\ref{response:spine-identities} justify differentiating the expectations. Differentiating $e^{\theta x}(x-a)^j$ then gives, throughout the window, \[ |A_n'|\le C(n+1),\quad |H_n''|\le C(n+1)^4,\quad |K_n'|\le C(n+1)^3,\quad |J_n'|\le C(n+1)^4. \] For the second derivative, use $(\partial_h\log R_n)^2+\partial_h^2\log R_n =\langle\mathcal C_n^2\rangle$ and $(d,s)=(2,1)$. Integration proves \eqref{response:windowA}--\eqref{response:windowJ}. The same calculation for $e^{\theta x}(1+(x-a)^2)^{7/4}$ gives a field derivative $O((n+1)^{9/2})$. Its critical expectation is $O((n+1)^{5/2})$ by interpolation between the absolute logarithmic moments of orders three and four. This proves \eqref{response:windowSevenHalves}. Finally, integrate $\partial_h\log\Pi_n(h)=\sum_{j0$ may be real; generation indices are always integers. For $n=\lfloor Nt\rfloor$, define \begin{gather} \nu_N(t,\dd x) =N\E\!\left[(R_n^\theta-1)\mathbf1_{\{R_n\ge1\}} \delta_{\log R_n/(aN)}(\dd x)\right], \label{scaling:measure}\\ q_N(s)=\theta-\frac{s}{aN},\qquad L_{n,N}(s)=N\bigl(G_n(q_N(s))-1\bigr),\quad 0\le s\le N\log2. \label{scaling:unshifted} \end{gather} We write $L_N(t,s)=L_{\lfloor Nt\rfloor,N}(s)$. Statements on a fixed compact set of nonnegative weights are understood for $N$ large enough that the set lies in this range. Put $v=\gamma\lambda$ and $T_v=\pi/\sqrt v$ for $v>0$, with $T_0=\infty$. The exponential family and its sine parametrization are due to \citet[Section~3.2]{DR14}; see also \citet[Section~2, (24)--(29)]{CDDS20}. Write \[ u_v(t)= \begin{cases}\sin(\sqrt v\,t)/\sqrt v,&v>0,\\t,&v=0,\end{cases} \] and, for $00$, write \begin{equation}\label{cont:laplaceprofile} F_v(t,s)=\frac{a_v(t)}{s+b_v(t)}. \end{equation} Exponential weights allow us to follow the profile after its total mass becomes infinite. \begin{theorem}\label{cont:continuation} Suppose $\lambda>0$. Let $I\subset(0,T_v)$ be compact and let $s>0$ satisfy $s+\inf_{t\in I}b_v(t)>0$. Then \begin{equation}\label{cont:transformconvergence} L_N(t,s)\longrightarrow F_v(t,s) \end{equation} uniformly for $t\in I$. At each such time, $e^{-sx}\nu_N(t,\dd x)$ converges weakly to $e^{-sx}f_v(t,x)\,\dd x$. \end{theorem} \subsection{Compactness and entrance data} For the initial compactness argument, take $T>0$ with $\lambda T^2$ sufficiently small for Proposition~\ref{prop:window-moments} to apply. For $01$, $d_q(r)>0$, whereas $d_q(1)=0$ and $d_q(r)\to B^q-1>0$. Its derivative in $r$ at $1$ is bounded away from zero on compact $q$-intervals. Hence $|F(r)-F(1)|\le C_Fd_{q_0}(r)$ for $r\ge1$. For $n\ge1$, the floor $r_n\ge b_n^2\ge b_1^2$ controls $r<1$ within the domain on which $F$ is bounded. Applying the same comparison to $\partial_qd_q$, which is bounded and vanishes at $1$, gives \eqref{cont:defectdifference}. Substituting $G_{n+1}(q)=B^{-q}(G_n(q)^2+B^q-1-D_n(q))$ gives \eqref{cont:recurrence}. Expanding yields \eqref{cont:onesided}; summing the recurrence, and using the exponentially small negative part of $D_n$, gives \eqref{scaling:defectsum} and the $BV$ bound. Fix $0\le s_0\le s_1<\infty$ and take $N$ large enough that $s_1\le N\log2$. Then, for $s\in[s_0,s_1]$, $0\le q_N(s_1)\le q_N(s)\le q_N(s_0)\le\theta<1$. On $b_n\le R_n<1$, both $1-R_n^{q_N(s)}$ and $R_n^{q_N(s)}-R_n^{q_N(s_0)}$ lie in $[0,1-b_n]$; on $R_n\ge1$, the powers are monotone in their exponent. Hence \[ -N(1-b_n)\le L_{n,N}(s) \le L_{n,N}(s_0)+N(1-b_n),\qquad s\in[s_0,s_1]. \] Since $1-b_n\le\rho^n$, the errors are $O(N\rho^n)$ uniformly on this compact weight interval. We now subtract the recurrences at $s$ and $s_0$. By \eqref{cont:defectdifference}, the remaining pointwise error is $O_s(N^{-2}+|D_n(q_N(s_0))|+\rho^n)$, whose sum is $O_s(N^{-1})$. \end{proof} \begin{lemma}[One-sided compactness]\label{lem:one-sided} Let $f_N$ be uniformly bounded on a compact interval and suppose \[ f_N(t)-f_N(r)\le C(t-r)+\varepsilon_N \quad(r\le t),\qquad \varepsilon_N\to0. \] If $f_N\to f$ in $L^1$ and $f$ is continuous, then convergence is uniform on every interior compact interval. \end{lemma} \begin{proof} An upward excursion of fixed height persists on an interval of fixed length to its left; a downward excursion persists to its right. Choose that length using $C$ and the modulus of continuity of $f$. Either excursion contradicts $L^1$ convergence. \end{proof} \begin{proof}[Proof of Theorem~\ref{scaling:compactness}] \emph{Bounds and time compactness.} Fix $r>0$. The critical estimates and Proposition~\ref{prop:window-moments} give \eqref{scaling:entrancebounds} for the approximations uniformly on $[r,T]$, with comparison constants independent of $r$ and errors vanishing as $N\to\infty$ for each fixed $r$. All generations used then satisfy $n=\lfloor Nt\rfloor\ge1$ for $N$ large enough. We first pass to the limit on each such interval and only afterwards let $t\downarrow0$. The lower bounds follow from monotonicity of $A_n$ and, for large $n$, of $r^\theta(\log r-a)^3$. Increasing $r$ can decrease $r^\theta(\log r-a)^2$ by only a bounded amount. The moment estimate \eqref{response:windowSevenHalves} gives \begin{equation}\label{scaling:perturbedmoments} \int x^{7/2}\nu_N(t,\dd x)\le C(t+N^{-1})^{5/2}. \end{equation} Apply Lemma~\ref{scaling:mellin-calculus} at $s_0=0$. Helly's theorem yields the asserted $L^1$ convergence of $S_N$; choose a Borel $BV$ representative $S$ and the subsequence so that pointwise convergence also holds at every continuity point of $S$. The exact recurrences \eqref{eq:prodH}--\eqref{eq:prodJ} hold at every field. Their bounded kernels satisfy $\omega_1(1)=0$ and $\omega_2(1)=-2a^2$. Equations \eqref{scaling:kernel} and \eqref{scaling:defectsum} make their normalized cumulative errors vanish. After removing these errors, the linear interpolants are uniformly Lipschitz. This gives compactness of $d_N,\mathcal K_N,\mathcal P_N$ and \eqref{scaling:odes}. Since $L_N(t,s)=S_N(t)+\mathcal H_N(t,s)-s$, \eqref{scaling:transform-difference} at $s_0=0$ gives \begin{equation}\label{scaling:transformincrement} \Delta\mathcal H_N =N^{-1}\left(S_N\mathcal H_N+ \tfrac12\mathcal H_N^2-\tfrac12s^2\right) +e_{n,N}(s),\qquad \sum_n|e_{n,N}(s)|=O_s(N^{-1}). \end{equation} The first tilted logarithmic moment bounds the derivative in $s$. Arzelà--Ascoli and a diagonal extraction therefore give joint compactness and \eqref{scaling:riccati}. Products pass to the limit using $L^1$ convergence of $S_N$ and uniform convergence of the other factors. \emph{Positive measures and the boundary.} The mass of $\nu_N$ differs from $S_N$ by $O(N\rho^n)$. Together with \eqref{scaling:perturbedmoments}, this gives tightness and uniform integrability through the third moment. Let $E$ be the full-measure Borel set of continuity points of $S$. At each $t\in E$, any measure limit satisfies \eqref{scaling:laplace}--\eqref{scaling:momentslimit}, which determine it uniquely. The transform discrepancy tends to zero because \[ \E(\log R_n)^+ \le C(A_n-1+\rho^n)=O(N^{-1}). \] Denote this unique limit by $\nu_t$ for $t\in E$ and set $\nu_t=0$ otherwise. For every integer $j\ge0$, \[ t\longmapsto\int e^{-jx}\nu_t(\dd x) =\mathbf1_E(t)\bigl(S(t)+\mathcal H(t,j)-j\bigr) \] is Borel. The algebra spanned by $x\mapsto e^{-jx}$, $j\ge0$, contains the constants and generates the Borel sigma-field of $[0,\infty)$. The functional monotone-class theorem therefore shows that $t\mapsto\nu_t(A)$ is Borel for every Borel set $A$. Thus $(\nu_t)$ is a measurable finite-measure kernel, with locally bounded mass. For $t\in E$, the positive representation extends $\mathcal H(t,\cdot)$ holomorphically to $\Re s>0$ and continuously to $\Re s\ge0$. The kernel property makes the corresponding time integrals measurable. On compact time intervals, the bound $|\mathcal H(t,s)|\le |s|+2S(t)$ gives domination on compact subsets of the closed right half-plane. The integral equation consequently extends by analytic continuation for endpoints in $E$. Fixing one such endpoint then defines the continuous extension in $t$ at the remaining times. For $t\in E$ and real $s\ge0$, write $b(t,s)=s-\mathcal H(t,s)=\int(1-e^{-sx})\nu_t(\dd x)$. Then $0\le b\le S$ and \[ \partial_t\mathcal H=s(S-b)-Sb+\tfrac12b^2. \] To exclude an atom at zero, we integrate between continuity points of $S$, divide by $s$, and let $s\to\infty$. The leading terms $s$ cancel at the endpoints, while dominated convergence on the right gives $\int\nu_t(\{0\})\,\dd t=0$. This proves \eqref{scaling:noatom} without assuming $S$ continuous. \emph{Entrance data.} Taylor's formula and \eqref{scaling:perturbedmoments} give \eqref{scaling:taylor}; its remainder is $o(P(t))$ for fixed $s$. The uniform field expansions are \[ H_n(h_c+u)=H_n(h_c)+u\,\partial_hH_n(h_c)+O(u^2n^4), \qquad J_n(h_c+u)=J_n(h_c)+O(un^4). \] For each fixed $t>0$, insert $u=\lambda/N^2$ and $n=\lfloor Nt\rfloor$ and first let $N\to\infty$ along the subsequence. Using $\partial_hH_n(h_c)/\Pi_n(h_c)\to\kappa$ and $J_n(h_c)/\Pi_n(h_c)\to\ell_3$, we obtain \[ d(t)-\frac{\lambda\kappa a^2}{\ell_3}P(t) =O(\lambda^2t^4). \] Now let $t\downarrow0$. Since $P(t)\asymp t^2$, this gives \eqref{scaling:matching}. The common factor $\Pi_n$ cancels; its precise asymptotic constant is not needed here. \end{proof} % End scaling.tex % Begin profile.tex \section{Identification of the limiting profile} \label{sec:profile-identification} We identify the profile \eqref{eq:profile} from the entrance data by showing that a regularized Laplace transform is a polynomial of degree at most one. \begin{theorem}\label{th:profile-identification} Every subsequential limit in Theorem~\ref{scaling:compactness} agrees, on a common initial interval, with $f_v(t,x)\,\dd x$, where $f_v$ is defined in \eqref{eq:profile}. At criticality, \begin{equation}\label{eq:exact-critical} n(A_n-1)\to2,\qquad \frac{K_n}{a^2n}\to1,\qquad \frac{J_n}{a^3n^2}\to\frac32,\qquad \frac{\Pi_n}{n^2}\to\frac{3a^3}{2\ell_3}. \end{equation} \end{theorem} The constants $2,1,3/2$ in \eqref{eq:exact-critical} agree with the critical moment predictions for the binary hard recursion; see Appendix~III of the 1983 archival version of \citet{CEGM84b} and \citet[Section~5.1]{DS20}. These predictions go beyond the earlier proved quadratic product bounds. The limits asserted here concern the soft diamond recursion. The general inverse correspondence for canonical systems is due to de Branges \citep{deBranges61,deBranges68}; a modern statement is \citet[Theorem~2.2]{EKT18}. We prove the rigidity needed for the present entrance law directly. For the first assertion of the following lemma, compare the Krein--de Branges exponential-type formula in \citet[Section~6, Theorem~11]{Romanov14}. The second assertion is the classical Phragm\'en--Lindel\"of principle at minimal exponential type \citep[Section~II, no.~5]{PL08}. We include short proofs. \begin{lemma}[Growth bound and Phragm\'en--Lindel\"of principle] \label{lem:entire-bounds} Fix a measurable $w:(0,T)\to(0,\infty)$ such that $w,w^{-1}$ are bounded. The fundamental matrix of $Y'=-z\left(\begin{smallmatrix}0&w^{-1}\\w&0\end{smallmatrix}\right)Y$ normalized by $M(0,z)=I$ satisfies \[ \|M(T,z)\|\le C_{\varepsilon,T,w} e^{T|\Re z|+\varepsilon|z|}\qquad(\varepsilon>0). \] A holomorphic function of minimal exponential type in a half-plane, continuous and bounded by one on its boundary, is bounded by one throughout that half-plane. \end{lemma} \begin{proof} By density in $L^1(0,T)$, we choose a step function $w_0\in[w_-,w_+]$, where $00$, define \[ y=\exp\!\left(-\tfrac12\int_0^t\mathcal H(v,s)\dd v\right), \qquad U=\frac{2\mathcal H y}{s^2Q}. \] Direct differentiation gives \begin{equation}\label{eq:regular-system} U'=(U-y)/Q,\qquad y'=-s^2QU/4,\qquad (wU')'=s^2wU/4. \end{equation} Moreover, \begin{equation}\label{eq:regular-traces} wU'=\frac yR(2\mathcal H/s^2-Q),\qquad U(0,s)=1,\qquad (wU')(0,s)=-s/3-2\alpha/s. \end{equation} The traces follow from \eqref{eq:entrance-expansion}; no trace of $w$ is assumed. \emph{Entire-function rigidity.} The value and flux of $sU$ at zero are the polynomials $s$ and $-s^2/3-2\alpha$. The integral form of \eqref{eq:regular-system} therefore makes $\mathcal B_t(s)=se^{st/2}U(t,s)$ entire. For $\Re s\ge0$, positivity of the limiting measure gives \[ |\mathcal H+sd|\le|s|^2K/2,\qquad |\mathcal H|\le|s|(2+d),\qquad 0\le\Re(s-\mathcal H)\le C\min(t^{-1},|s|). \] For $s\ne0$, split the time integral at $\min(t,|s|^{-1})$ to obtain \[ \exp\!\left(\tfrac12\int_0^t \Re(s-\mathcal H(v,s))\dd v\right) \le C(1+t|s|)^C. \] Since $\mathcal B_t(s)=2\mathcal H(t,s) \exp(\tfrac12\int_0^t(s-\mathcal H(v,s))\dd v)/(sQ(t))$, this gives polynomial growth in the right half-plane, with degree bound independent of $t$; the value at $s=0$ follows by continuity. In the left half-plane with $|s|\ge1$, apply Lemma~\ref{lem:entire-bounds} to $(U,-2wU'/s)$ with $z=s/2$. Its entrance value is $(1,2/3+4\alpha/s^2)$, which is bounded in this range. For each fixed $t$, the factor $e^{st/2}$ cancels $e^{t|\Re s|/2}$ and gives $|\mathcal B_t(s)|\le C_{\varepsilon,t,w,\alpha}|s|e^{\varepsilon|s|}$. For a fixed sufficiently large integer $k$, $\mathcal B_t(s)/(1-s)^k$ is bounded on the imaginary axis and has minimal exponential type in the left half-plane. The second part of the lemma and Cauchy's estimates show that $\mathcal B_t$ is a polynomial of degree at most $k$. Write $\mathcal B_t(s)=\sum_{j=0}^k b_j(t)s^j$. For each fixed $s$, the integral system makes the value and flux of $sU$ locally absolutely continuous. Since $w$ is locally absolutely continuous, the same holds for $\mathcal B_t(s)$ and $w\partial_t\mathcal B_t(s)$. Interpolation at $k+1$ distinct positive weights therefore shows that both $b_j$ and $wb_j'$ are locally absolutely continuous. Direct differentiation gives \begin{equation}\label{eq:polynomial-equation} (w\partial_t\mathcal B_t)'= s\bigl(w\partial_t\mathcal B_t+\tfrac12w'\mathcal B_t\bigr) \end{equation} for almost every $t$ at each fixed $s$. Choose $k+2$ distinct positive weights and a common full-measure set of times where the derivatives exist and these identities hold. The difference of the two sides of \eqref{eq:polynomial-equation} has degree at most $k+1$ in $s$ and vanishes at these weights, so all its coefficients vanish almost everywhere. The entrance values at any $k+1$ of these weights give $b_j(t)\to\mathbf1_{\{j=1\}}$ as $t\downarrow0$. If $k\ge2$, the coefficient of $s^{k+1}$ gives $(\sqrt w\,b_k)'=0$. Boundedness of $w$ and the zero entrance trace force $b_k=0$. Repeating this argument leaves degree at most one. \emph{A scalar Riccati equation.} Put $\ell=d/Q$ and $I=\int_0^t d$. Expanding at $s=0$, and using $K=Q+dI$, yields \begin{equation}\label{eq:polynomial-final} \mathcal B_t(s)=s(1-t\ell)-2\ell. \end{equation} Equation \eqref{eq:polynomial-equation} gives $w(1-t\ell)^2=w_0>0$, while $\ell'=\alpha/w$. Since $\ell=O(t)$, $w\to w_0$, and \[ wU'=e^{-st/2} \left[-\tfrac s2w(1-t\ell)-\alpha t-\frac{2\alpha}s\right]. \] The flux in \eqref{eq:regular-traces} forces $w_0=2/3$. Consequently $r=\ell/(1-t\ell)$ solves \[ r'=v+r^2,\qquad r(0)=0,\qquad v=3\alpha/2=\gamma\lambda. \] Hence $r=vu_v/u_v'$ initially. The identities $Q=w\int_0^t w^{-1}$, $S=w'/w+2/Q$ and $R=Q^2/w$ recover the moments. Uniqueness in \eqref{eq:regular-system} recovers $\mathcal H$, and uniqueness of Laplace transforms gives \eqref{eq:profile}. \emph{Critical asymptotics.} For $\lambda=0$, the moments are $S=2/t$, $K=t$, $P=3t^2/2$. Lemma~\ref{lem:one-sided} upgrades the $L^1$ convergence of $S_N$ to uniform convergence on interior compact intervals. Fix such a time $t_0$ and take $N=n/t_0$ to obtain the first three limits in \eqref{eq:exact-critical}. The last follows from $J_n/\Pi_n\to\ell_3$. Together with $\partial_hH_n(h_c)/\Pi_n\to\kappa$, it also proves the second identity in \eqref{eq:gamma}. \end{proof} % End profile.tex % Begin continuation.tex \section{Continuation up to the pole} \label{cont:section} Fix $\lambda>0$ and use the notation of Section~\ref{scaling:section}. We now complete the proof of the weighted convergence theorem by continuing the profile identified in Section~\ref{sec:profile-identification}. \begin{proof}[Proof of Theorem~\ref{cont:continuation}] Integrating \eqref{eq:profile} gives $F_v(t,s)=\int_0^\infty e^{-sx}f_v(t,x)\,\dd x$ whenever $s+b_v(t)>0$; its unweighted mass is infinite when $b_v(t)\le0$. We first extend the convergence from a time where the profile is identified. At the smallest weight used in this extension, possible loss of mass at infinity requires a separate argument. \emph{Local bounds and compactness.} On interior compact subsets of the initial interval of identification, Lemma~\ref{lem:one-sided} and the convergence of the transform differences give uniform convergence at every fixed nonnegative weight. Choose a time $r>0$ in an interval where this convergence is known and a base weight $s_0>0$ with $s_0+b_v(r)>0$. We obtain a local bound by a first-exit argument. Choose $A>1+F_v(r,s_0)$ and put $M=A+1$. For large $N$, $L_N(r,s_0)\le A$, while the floor gives $L_{n,N}(s_0)\ge-CN\rho^n>-1$ for $n\ge\lfloor Nr\rfloor$. The one-step estimate \eqref{cont:onesided} uses only the bound at the current index. Before, and at, the first crossing of $M$, summing it therefore gives \[ L_{n,N}(s_0)\le A+C\delta+o_N(1),\qquad n\le\lfloor N(r+\delta)\rfloor, \] where $C=1+\max_{-1\le z\le M}\Psi_{s_0}(z)$; the accumulated $N^{-2}$ and $N\rho^n$ errors tend to zero. Choose $C\delta<1/2$. A first crossing is then impossible for large $N$. Thus $|L_N(t,s_0)|\le M$ on $[r,r+\delta]$, with $\delta$ depending only on $s_0$ and an upper bound for $F_v(r,s_0)$. Shrink $\delta$ if necessary so that the profile is finite at this weight throughout the interval. All further finite-$N$ estimates in the weight variable are used on fixed compact sets before taking $N\to\infty$. Lemma~\ref{scaling:mellin-calculus} gives $BV$ compactness at $s_0$, the integrated defect estimate \eqref{scaling:defectsum}, and equicontinuity of the differences. After a diagonal extraction, $L_N(\cdot,s_0)$ converges in $L^1$ and $L_N(\cdot,s)-L_N(\cdot,s_0)$ converges uniformly for fixed $s\ge s_0$. Their limits satisfy \begin{equation}\label{cont:differenceequation} \frac{\dd}{\dd t}\bigl(L(t,s)-L(t,s_0)\bigr) =sL(t,s)+\tfrac12L(t,s)^2 -s_0L(t,s_0)-\tfrac12L(t,s_0)^2 \end{equation} in integral form. \emph{The limiting measure and its trace.} Uniformly on the local interval, \[ L_N(t,s)-\int e^{-sx}\nu_N(t,\dd x)\longrightarrow0, \qquad s\ge s_0. \] Indeed, the discrepancy is bounded by $C_s\E(\log R_n)^++CN\rho^n$, and \[ \E(\log R_n)^+ \le q_N(s_0)^{-1} \bigl(G_n(q_N(s_0))-1+C\rho^n\bigr)=O(N^{-1}). \] A bound at $s_0$ need not give tightness at $s_0$. At every $\sigma>s_0$, however, \[ \int_{x>R}e^{-\sigma x}\nu_N(t,\dd x) \le e^{-(\sigma-s_0)R} \int e^{-s_0x}\nu_N(t,\dd x). \] Choose the Helly subsequence so that the base transforms converge at every continuity point of a Borel $BV$ representative of their limit. On this common full-measure Borel set, weighted tightness and uniqueness of Laplace transforms give a positive measure $\mu_t$ with $\int e^{-s_0x}\mu_t(\dd x)\le M$ and \[ L(t,s)=\int e^{-sx}\mu_t(\dd x),\qquad s>s_0. \] Set $\mu_t=0$ on the exceptional set. For fixed $\sigma>s_0$, the integrals of $e^{-kx}$, $k\in\mathbb N_0$, against $e^{-\sigma x}\mu_t(\dd x)$ are Borel functions of $t$. The algebra generated by these functions of $x$ contains the constants and generates the Borel sets of $[0,\infty)$. The functional monotone-class argument used in the proof of Theorem~\ref{scaling:compactness} therefore shows that $e^{-\sigma x}\mu_t(\dd x)$, and hence $\mu_t(\dd x)$, are measurable measure kernels. Their convolutions are measurable kernels as well, by applying the same argument to rectangles and then integrating against the product kernels. Integrating \eqref{cont:differenceequation}, dividing by $s$, and letting $s\to\infty$ gives $\int\mu_t(\{0\})\,\dd t=0$ on every interior time interval: the endpoint terms divided by $s$ vanish, and dominated convergence applies since $0\le L(t,s)\le M$ almost everywhere. Thus there is no atom at zero almost everywhere; enlarge the exceptional Borel set accordingly and set the kernel to zero there. To identify the trace at $r$, we use the continuity of the transform differences up to $r$. If $\Delta=L(r+,s_0)-F_v(r,s_0)$, their right traces are therefore $F_v(r,s)+\Delta$ for $s>s_0$. The one-sided bound gives $\Delta\le0$; positivity and $F_v(r,s)\to0$ as $s\to\infty$ give $\Delta\ge0$. Hence $\Delta=0$. The tail bound and uniqueness of Laplace transforms now show that, along the good times, $e^{-\sigma x}\mu_t(\dd x)$ converges weakly as $t\downarrow r$ to $e^{-\sigma x}f_v(r,x)\,\dd x$ for every $\sigma>s_0$. We denote this measure trace by $\mu_r=f_v(r,x)\,\dd x$. \emph{Transport and uniqueness.} Linear combinations of \eqref{cont:differenceequation} whose coefficients sum to zero yield \begin{equation}\label{cont:weakpde} \frac{\dd}{\dd t}\langle\varphi,\mu_t\rangle =-\langle\varphi',\mu_t\rangle +\tfrac12\langle\varphi,\mu_t*\mu_t\rangle, \qquad \varphi\in C_c^1((0,\infty)). \end{equation} The identity is understood in integral form. To justify the passage from exponentials to test functions, write $\varphi(x)=e^{-\sigma x}g(e^{-x})$, with $\sigma>s_0$ and $g$ zero near $0$ and $1$. Approximate $g'$ uniformly by polynomials and integrate from $1$. This approximates both $g$ and $g'$ uniformly and preserves $g(1)=0$. The resulting exponential combinations vanish at zero, and the errors are controlled in the weighted norms of $\mu_t$ and $\mu_t*\mu_t$. In particular, each pairing $\langle\varphi,\mu_t\rangle$ has an absolutely continuous representative, whose value at $r$ is $\langle\varphi,\mu_r\rangle$. Taking a countable family dense in $C^1$ on each spatial compact and using the local mass bounds gives a common full-measure set of times where all these representatives equal the measure pairings. We also need time-dependent tests. Let $\psi(\tau,x)$ be $C^1$ with its spatial support in one compact subset of $(0,\infty)$. Its piecewise linear interpolants in $\tau$ are finite sums of fixed spatial tests with piecewise linear coefficients. The product rule for the preceding absolutely continuous pairings applies to these interpolants. As the mesh tends to zero, the interpolants and their spatial derivatives converge uniformly; their time derivatives converge uniformly away from the mesh points, in the essential-supremum sense. The local mass bounds for $\mu_\tau$ and $\mu_\tau*\mu_\tau$ therefore give, at every endpoint $t$ in the common full-measure set, \[ \langle\psi(t,\cdot),\mu_t\rangle -\langle\psi(r,\cdot),\mu_r\rangle =\int_r^t\left( \langle\partial_\tau\psi-\partial_x\psi,\mu_\tau\rangle +\tfrac12\langle\psi,\mu_\tau*\mu_\tau\rangle \right)\dd\tau. \] Let $\mathsf T_\tau$ be left translation with absorption at zero: \[ \langle\varphi,\mathsf T_\tau\mu\rangle =\int_{x>\tau}\varphi(x-\tau)\mu(\dd x),\qquad \|\mu\|_\sigma=\int e^{-\sigma x}|\mu|(\dd x). \] Then \[ \|\mathsf T_\tau\mu\|_\sigma\le e^{\sigma\tau}\|\mu\|_\sigma, \qquad \|\mu*\eta\|_\sigma\le\|\mu\|_\sigma\|\eta\|_\sigma. \] Apply the time-dependent identity to $\psi(\tau,x)=\varphi(x-t+\tau)$, where $\varphi\in C_c^1((0,\infty))$ is extended by zero to the real line. For $r\le\tau\le t$, these tests have a common compact support inside $(0,\infty)$ and satisfy $\partial_\tau\psi=\partial_x\psi$. This gives the mild equation \begin{equation}\label{cont:mild} \mu_t=\mathsf T_{t-r}\mu_r+ \tfrac12\int_r^t \mathsf T_{t-\tau}(\mu_\tau*\mu_\tau)\,\dd\tau. \end{equation} The integral is defined as a measure integral of the positive kernels: on a Borel set $A\subset(0,\infty)$ its integrand is $\int_{x>t-\tau}\mathbf1_A(x-t+\tau) (\mu_\tau*\mu_\tau)(\dd x)$. It has finite weighted mass, bounded by $\delta e^{\sigma\delta}\sup_\tau\|\mu_\tau\|_\sigma^2$. Equality of the compactly supported test integrals identifies the measures almost everywhere in time. The right-hand side of \eqref{cont:mild} then defines a measurable mild representative at every time, with bounded weighted norm; changing the source kernel on a null set does not change this formula. For uniqueness, let two such solutions have the same trace and weighted norms bounded by $M_\sigma$ on $[r,r+\delta]$. Put $C_\sigma=M_\sigma e^{\sigma\delta}$ and define deterministic functions by \[ E_0(t)=2M_\sigma,\qquad E_{j+1}(t)=C_\sigma\int_r^t E_j(\tau)\,\dd\tau. \] If $\|\mu_\tau-\widetilde\mu_\tau\|_\sigma\le E_j(\tau)$ almost everywhere, the convolution bound gives $\|\mu_\tau*\mu_\tau- \widetilde\mu_\tau*\widetilde\mu_\tau\|_\sigma \le2M_\sigma E_j(\tau)$ there. Test the difference of the mild equations with any Borel function bounded in absolute value by $e^{-\sigma x}$, integrate this deterministic bound, and then take the supremum over the test functions. It follows that \[ \|\mu_t-\widetilde\mu_t\|_\sigma\le E_{j+1}(t),\qquad E_j(t)=2M_\sigma\frac{(C_\sigma(t-r))^j}{j!}. \] Letting $j\to\infty$ proves uniqueness. This argument requires neither measurability of the total-variation distance nor strong continuity of translation in total variation. Since $a_v'=-b_va_v$ and $b_v'=-a_v/2$, the density \eqref{eq:profile} solves $\partial_tf_v=\partial_xf_v+\tfrac12f_v*f_v$. Its weighted norms are bounded on the local interval, and the same time-dependent testing argument shows that it satisfies \eqref{cont:mild} with trace $\mu_r$. It is therefore the unique limit at every weight above $s_0$. \emph{The base weight and continuation.} It remains to exclude loss of mass at infinity at the base weight $s_0$. Put $e(t)=L(t,s_0)-F_v(t,s_0)$. Since the transform at $s>s_0$ has been identified with a smooth function, the integral form of \eqref{cont:differenceequation} first shows that $e$ is absolutely continuous. Subtracting the profile equation then gives \begin{equation}\label{cont:infinitydefect} e'=(s_0+F_v(t,s_0))e+\tfrac12e^2,\qquad e(r)=0. \end{equation} Scalar uniqueness gives $e=0$. All subsequential limits at the base weight are identified. Lemma~\ref{lem:one-sided} upgrades $L^1$ convergence to uniform convergence on interior compact intervals. The uniform convergence of the transform differences gives the same conclusion at every fixed larger weight. Now fix $I$ and $s$ as in the theorem. Since $b_v'=-a_v/2<0$, we can choose $00$. Choose $t_0<\min I$ in the initial interval of identification. The profile at weight $s_0$ is bounded on $[t_0,t_2]$, so the local comparison interval can be chosen with a uniform length $\delta$, further reduced to be smaller than $t_2-\max I$. Starting at $t_0$, take successive initial times separated by $\delta/2$, strictly inside the preceding intervals. Finitely many such intervals have interiors covering $I$, and the final interval ends strictly after $\max I$. The interior uniform upgrade therefore covers both endpoints of $I$ and proves \eqref{cont:transformconvergence}. At each time in $I$, the bound at $s_00$, set $v=\gamma\lambda$ and $T_v=\pi/\sqrt v$, and use the rescaled quantities of Section~\ref{scaling:section}, evaluated at $h_c+\lambda/N^2$. \subsection{Upper bound} \begin{corollary}\label{cont:freeenergyupper} The free energy satisfies \begin{equation}\label{cont:upperconstant} \limsup_{u\downarrow0}\sqrt u\,\log F(h_c+u) \le-\frac{\pi\log2}{\sqrt\gamma}. \end{equation} \end{corollary} \begin{proof} For $00$, \[ \E\log R_{\lfloor Nt\rfloor} \le q_N(s)^{-1}\log G_{\lfloor Nt\rfloor}(q_N(s)) =O(N^{-1}). \] The finite-size bound yields $F(h_c+\lambda/N^2)\le C2^{-\lfloor Nt\rfloor}$. Letting $N\to\infty$ and then $t\uparrow T_v$ gives \[ \limsup_{N\to\infty}N^{-1}\log F(h_c+\lambda/N^2) \le-\frac{\pi\log2}{\sqrt{\gamma\lambda}}. \] Take $N=\sqrt{\lambda/u}$ to obtain \eqref{cont:upperconstant}. \end{proof} \subsection{Pair counting and the band criterion} Two-leaf witnesses, common-ancestor second moments and iterative tail amplification are used by \citet[Section~4.1, Lemma~4.1]{HS18}. We give a finite-band version: its hypothesis concerns one bounded interval, so their global-tail lemma does not apply directly. Let $Y_0$ be nonnegative and integer-valued. In this subsection, expectations of nonnegative random variables are taken in $[0,\infty]$. We use the recursion \[ Y_{j+1}=(Y_j^{(1)}+Y_j^{(2)}-1)^+,\qquad F_{\mathrm d}=\lim_{j\to\infty}2^{-j}\E Y_j\in[0,\infty], \] where the two copies are independent. \begin{lemma}[Pair counting]\label{escape:pairs} Let $L\ge2$ be an integer, $I\subset\{\lceil3L/4\rceil,\ldots,L\}$, and $W=\E[2^{Y_0}\mathbf1_{\{Y_0\in I\}}]$. Then \begin{equation}\label{escape:pair-tail} \P(Y_L\ge k)\ge2^{-k}\,2\left(\frac W{W+2}\right)^2, \qquad 1\le k\le\lfloor L/2\rfloor. \end{equation} \end{lemma} \begin{proof} Two leaves of masses $m,\ell\in I$ produce root mass $k$ when their most recent common ancestor has height $d=m+\ell-L-k+1$ above the leaves and all other leaves are zero. Indeed, $1\le d\le\min(m,\ell)$ and the resulting root mass is $m+\ell-L-d+1=k$. Restoring the other leaves can only increase it. Let $Z$ count unordered pairs with this property. There are $2^{L+d-2}$ pairs with common-ancestor height $d$. Assigning $m$ and $\ell$ to the left and right leaves of each pair, respectively, gives \begin{equation}\label{escape:first-pair-moment} \mu:=\E Z =\sum_{m,\ell\in I}2^{m+\ell-k-1}\P(Y_0=m)\P(Y_0=\ell) =2^{-k-1}W^2. \end{equation} Disjoint pairs have independent indicators. Given a leaf of mass $m\in I$, its expected number of partners is \[ g(m)=\sum_{\ell\in I}2^{d(m,\ell)-1}\P(Y_0=\ell) =2^{m-L-k}W\le2^{-k}W, \] and we set $g(m)=0$ outside $I$. Distinct partners are conditionally independent, so overlapping ordered pairs contribute at most $\sum_{\text{leaves }z}\E g(Y_z)^2\le2^{1-k}W\mu$. Consequently, \[ \E Z^2\le\mu+\mu^2+2^{1-k}W\mu,\qquad \P(Y_L\ge k)\ge\frac{\mu}{1+\mu+2^{1-k}W}. \] Substituting \eqref{escape:first-pair-moment} and using $k\ge1$ proves \eqref{escape:pair-tail}. \end{proof} \begin{proposition}[Uniform band criterion]\label{escape:band-exit} There are absolute constants $D_*,L_*<\infty$ such that \begin{equation}\label{escape:band-condition} L\ge L_*,\qquad L\,\E[2^{Y_0}\mathbf1_{\{\lceil3L/4\rceil\le Y_0\le L\}}]\ge D_* \end{equation} imply $\E Y_T\ge3/2$ for some integer $T\le2L$. In particular, \begin{equation}\label{escape:hard-energy} F_{\mathrm d}\ge\tfrac12\,2^{-2L}. \end{equation} One may take $D_*=L_*=10^4$. \end{proposition} \begin{proof} Let $\mathcal G(\alpha,L)$, $0\le\alpha\le2$, be the distribution supported on $\{0,\ldots,L\}$ with $\P(Y\ge k)=\alpha2^{-k}$ for $1\le k\le L$. Its upper-quarter weighted mass is $\alpha+\frac\alpha2(L-\lceil3L/4\rceil)\ge\alpha L/8$. After $L$ generations, Lemma~\ref{escape:pairs} gives the stochastic lower law $\mathcal G(\alpha',L')$, where \begin{equation}\label{escape:amplification} L'=\lfloor L/2\rfloor,\quad \alpha'=2\left(\frac{\alpha L}{\alpha L+16}\right)^2. \end{equation} Set $M=128$ and $A_*=9(M+16)^2$. Suppose $\alpha L^2\ge A_*$. As long as $w=\alpha L300$. Since $L$ is halved at each step, eventually $w\ge128$. The next step has $\alpha'\ge128/81$ and $L'\ge150$, hence mean $\alpha'(1-2^{-L'})>3/2$. The elapsed time is at most $L+\lfloor L/2\rfloor+\cdots\le2L$. For the general initial law in \eqref{escape:band-condition}, write $W$ for its band moment. After $L$ generations, Lemma~\ref{escape:pairs} gives $\mathcal G(\alpha_1,L_1)$ with $\alpha_1=2(W/(W+2))^2$ and $L_1=\lfloor L/2\rfloor$. For $L\ge3$, \[ \alpha_1L_1^2\ge\frac2{81}\min\{(WL)^2,L^2\}\ge A_* \] with the stated constants. The preceding amplification takes at most $2L_1$ further generations, so the total is at most $L+2L_1\le2L$. Finally, $Y_{j+1}\le Y_j^{(1)}+Y_j^{(2)}$ makes $2^{-j}\E Y_j$ nonincreasing in $[0,\infty]$, so its limit exists. If $\E Y_0=\infty$, retaining one leaf gives $Y_j\ge(Y_0^{\rm leaf}-j)^+$; hence every mean and $F_{\mathrm d}$ are infinite. If $\E Y_0<\infty$, iteration of $\E Y_{j+1}\ge2\E Y_j-1$ gives the standard bound \citep[(1.6)]{HS18} \begin{equation}\label{escape:finite-hard} F_{\mathrm d}\ge2^{-j}(\E Y_j-1). \end{equation} This bound also holds in the infinite-mean case, with $\infty-1=\infty$. Apply this at the generation $T\le2L$ obtained above to deduce \eqref{escape:hard-energy}. \end{proof} \subsection{Exact coupling at the deterministic floor} Fix a field $h$ and a generation $n\ge1$ of the diamond model, and set \begin{equation}\label{escape:normalization} a_n=\log(B/b_n),\qquad \theta_n=\frac{\log2}{a_n},\qquad Y_0=\left\lfloor\frac{\log(R_n/b_n)}{a_n}\right\rfloor. \end{equation} The next comparison transfers the band criterion to the diamond model. \begin{lemma}\label{escape:comparison} For the initial law in \eqref{escape:normalization}, \begin{equation}\label{escape:energy-transfer} F(h)\ge2^{-n}a_nF_{\mathrm d}. \end{equation} If this law, or any stochastically smaller integer law, satisfies \eqref{escape:band-condition}, then \begin{equation}\label{escape:diamond-exit} F(h)\ge\frac{a_n}{2}\,2^{-n-2L}. \end{equation} \end{lemma} \begin{proof} Here $n\ge1$ and $0\le Y_0\le\log(R_n/b_n)/a_n$, so $\E Y_0<\infty$ by Lemma~\ref{lem:basic-bounds}. This also holds for any stochastically smaller nonnegative integer law. We keep the same floor $b_n$ at all subsequent generations. Since $(b_n^2+c)/B\ge b_n$, \[ \frac{R'}{b_n}\ge \max\left\{1,\frac{b_n}{B} \frac{R_1}{b_n}\frac{R_2}{b_n}\right\}. \] We can therefore couple the integer recursion below the diamond so that $R_{n+j}\ge b_ne^{a_nY_j}$. Taking free-energy limits proves \eqref{escape:energy-transfer}. The same comparison holds for a stochastically smaller initial law, and Proposition~\ref{escape:band-exit} gives \eqref{escape:diamond-exit}. \end{proof} \subsection{The matching lower bound} Use the profile $f_v$ and its pole time $T_v$ from Theorem~\ref{cont:continuation}. The link between this pole and the BKT scale was identified in \citet[Section~3.2]{DR14} and \citet[(27)--(29)]{CDDS20}. The band criterion provides the lower bound for the present recursion. \begin{theorem}\label{escape:constant} For every fixed $\lambda>0$, \begin{equation}\label{escape:liminf} \liminf_{N\to\infty}\frac1N\log F(h_c+\lambda/N^2) \ge-T_v\log2. \end{equation} \end{theorem} \begin{proof} Fix $\varepsilon>0$, let $L=\lfloor\varepsilon N\rfloor$, and put $I_\varepsilon=[4\varepsilon/5,9\varepsilon/10]$. At $n=\lfloor Nt\rfloor$, $00$. Since $N>0$ is real, we may set $N=\sqrt{\lambda/u}$ in \eqref{escape:liminf}. Combining the resulting lower bound with Corollary~\ref{cont:freeenergyupper}, we obtain \begin{equation}\label{escape:energy-constant} \lim_{u\downarrow0}\sqrt u\,\log F(h_c+u) =-\frac{\pi\log2}{\sqrt\gamma}. \end{equation} \end{proof} % End escape.tex \setlength{\bibsep}{2pt plus 1pt minus 1pt} \begin{thebibliography}{99} \raggedright \bibitem[Alsmeyer, Hu and Mallein(2026)]{AHM26} G.~Alsmeyer, Y.~Hu and B.~Mallein. \newblock The Derrida--Retaux model on a geometric Galton--Watson tree. \newblock \href{https://arxiv.org/abs/2502.02991v2}{arXiv:2502.02991v2} (2026). \bibitem[Chen, Dagard, Derrida, Hu, Lifshits and Shi(2021)]{CDDHLS21} X.~Chen, V.~Dagard, B.~Derrida, Y.~Hu, M.~Lifshits and Z.~Shi. \newblock The Derrida--Retaux conjecture on recursive models. \newblock \emph{Ann. Probab.} \textbf{49} (2021), 637--670. \newblock \href{https://doi.org/10.1214/20-AOP1457}{doi:10.1214/20-AOP1457}. \bibitem[Chen, Dagard, Derrida and Shi(2020)]{CDDS20} X.~Chen, V.~Dagard, B.~Derrida and Z.~Shi. \newblock The critical behaviors and the scaling functions of a coalescence equation. \newblock \emph{J. Phys. A: Math. Theor.} \textbf{53} (2020), article 195202. \newblock \href{https://doi.org/10.1088/1751-8121/ab8134}{doi:10.1088/1751-8121/ab8134}. \bibitem[Chen, Derrida, Hu, Lifshits and Shi(2019)]{CDHLS19} X.~Chen, B.~Derrida, Y.~Hu, M.~Lifshits and Z.~Shi. \newblock A max-type recursive model: some properties and open questions. \newblock In V.~Sidoravicius (ed.), \emph{Sojourns in Probability Theory and Statistical Physics--III}, Springer Proceedings in Mathematics \& Statistics \textbf{300}, Springer, Singapore, 2019, pp.~166--186. \newblock \href{https://doi.org/10.1007/978-981-15-0302-3_6}{doi:10.1007/978-981-15-0302-3\_6}. \bibitem[Chen, Hu and Shi(2022)]{CHS22} X.~Chen, Y.~Hu and Z.~Shi. \newblock The sustainability probability for the critical Derrida--Retaux model. \newblock \emph{Probab. Theory Related Fields} \textbf{182} (2022), 641--684. \newblock \href{https://doi.org/10.1007/s00440-021-01091-z}{doi:10.1007/s00440-021-01091-z}. \bibitem[Chen and Shi(2021)]{CS21} X.~Chen and Z.~Shi. \newblock The stable Derrida--Retaux system at criticality. \newblock In M.~E.~Vares, R.~Fern\'andez, L.~R.~Fontes and C.~M.~Newman (eds.), \emph{In and Out of Equilibrium 3: Celebrating Vladas Sidoravicius}, Progress in Probability \textbf{77}, Birkh\"auser, Cham, 2021, pp.~239--264. \newblock \href{https://doi.org/10.1007/978-3-030-60754-8_12}{doi:10.1007/978-3-030-60754-8\_12}. \bibitem[Collet, Eckmann, Glaser and Martin(1984a)]{CEGM84a} P.~Collet, J.-P.~Eckmann, V.~Glaser and A.~Martin. \newblock A spin glass with random couplings. \newblock \emph{J. Stat. Phys.} \textbf{36} (1984), 89--106. \newblock \href{https://doi.org/10.1007/BF01015728}{doi:10.1007/BF01015728}. \bibitem[Collet, Eckmann, Glaser and Martin(1984b)]{CEGM84b} P.~Collet, J.-P.~Eckmann, V.~Glaser and A.~Martin. \newblock Study of the iterations of a mapping associated to a spin glass model. \newblock \emph{Comm. Math. Phys.} \textbf{94} (1984), 353--370. \newblock \href{https://doi.org/10.1007/BF01224830}{doi:10.1007/BF01224830}. \newblock Archival version: \emph{Les rencontres physiciens-math\'ematiciens de Strasbourg--RCP25} \textbf{33} (1983), expos\'e~6, 117--142. \href{https://www.numdam.org/item/RCP25_1983__33__117_0/}{Numdam}. \bibitem[de Branges(1961)]{deBranges61} L.~de Branges. \newblock Some Hilbert spaces of entire functions. II. \newblock \emph{Trans. Amer. Math. Soc.} \textbf{99} (1961), 118--152. \newblock \href{https://doi.org/10.1090/S0002-9947-1961-0133456-2}{doi:10.1090/S0002-9947-1961-0133456-2}. \bibitem[de Branges(1968)]{deBranges68} L.~de Branges. \newblock \emph{Hilbert Spaces of Entire Functions}. \newblock Prentice-Hall, Englewood Cliffs, 1968. \bibitem[Derrida, Hakim and Vannimenus(1992)]{DHV92} B.~Derrida, V.~Hakim and J.~Vannimenus. \newblock Effect of disorder on two-dimensional wetting. \newblock \emph{J. Stat. Phys.} \textbf{66} (1992), 1189--1213. \newblock \href{https://doi.org/10.1007/BF01054419}{doi:10.1007/BF01054419}. \bibitem[Derrida and Retaux(2014)]{DR14} B.~Derrida and M.~Retaux. \newblock The depinning transition in presence of disorder: a toy model. \newblock \emph{J. Stat. Phys.} \textbf{156} (2014), 268--290. \newblock \href{https://doi.org/10.1007/s10955-014-1006-y}{doi:10.1007/s10955-014-1006-y}. Author version: \href{https://arxiv.org/abs/1401.6919v1}{arXiv:1401.6919v1}. \bibitem[Derrida and Shi(2020)]{DS20} B.~Derrida and Z.~Shi. \newblock Results and conjectures on a toy model of depinning. \newblock \emph{Mosc. Math. J.} \textbf{20} (2020), 695--709. \newblock \href{https://doi.org/10.17323/1609-4514-2020-20-4-695-709}{doi:10.17323/1609-4514-2020-20-4-695-709}. Author version: \href{https://arxiv.org/abs/2005.10208v1}{arXiv:2005.10208v1}. \bibitem[Eckhardt, Kostenko and Teschl(2018)]{EKT18} J.~Eckhardt, A.~Kostenko and G.~Teschl. \newblock Spectral asymptotics for canonical systems. \newblock \emph{J. Reine Angew. Math.} \textbf{736} (2018), 285--315. \newblock \href{https://doi.org/10.1515/crelle-2015-0034}{doi:10.1515/crelle-2015-0034}. \bibitem[Giacomin(2007)]{Giacomin07} G.~Giacomin. \newblock \emph{Random Polymer Models}. \newblock Imperial College Press, London, 2007. \newblock \href{https://doi.org/10.1142/p504}{doi:10.1142/p504}. \bibitem[Giacomin(2011)]{Giacomin11} G.~Giacomin. \newblock \emph{Disorder and Critical Phenomena Through Basic Probability Models: \'Ecole d'\'Et\'e de Probabilit\'es de Saint-Flour XL--2010}. \newblock Lecture Notes in Mathematics \textbf{2025}, Springer, Heidelberg, 2011. \newblock \href{https://doi.org/10.1007/978-3-642-21156-0}{doi:10.1007/978-3-642-21156-0}. \bibitem[Giacomin, Lacoin and Toninelli(2010a)]{GLT10} G.~Giacomin, H.~Lacoin and F.~L.~Toninelli. \newblock Hierarchical pinning models, quadratic maps and quenched disorder. \newblock \emph{Probab. Theory Related Fields} \textbf{147} (2010), 185--216. \newblock \href{https://doi.org/10.1007/s00440-009-0205-y}{doi:10.1007/s00440-009-0205-y}. \bibitem[Giacomin, Lacoin and Toninelli(2010b)]{GLT10m} G.~Giacomin, H.~Lacoin and F.~L.~Toninelli. \newblock Marginal relevance of disorder for pinning models. \newblock \emph{Comm. Pure Appl. Math.} \textbf{63} (2010), 233--265. \newblock \href{https://doi.org/10.1002/cpa.20301}{doi:10.1002/cpa.20301}. \bibitem[Giacomin and Toninelli(2006)]{GT06} G.~Giacomin and F.~L.~Toninelli. \newblock Smoothing effect of quenched disorder on polymer depinning transitions. \newblock \emph{Comm. Math. Phys.} \textbf{266} (2006), 1--16. \newblock \href{https://doi.org/10.1007/s00220-006-0008-2}{doi:10.1007/s00220-006-0008-2}. \bibitem[Harris(1960)]{Harris60} T.~E.~Harris. \newblock A lower bound for the critical probability in a certain percolation process. \newblock \emph{Proc. Cambridge Philos. Soc.} \textbf{56} (1960), 13--20. \newblock \href{https://doi.org/10.1017/S0305004100034241}{doi:10.1017/S0305004100034241}. \bibitem[Hu, Mallein and Pain(2020)]{HMP20} Y.~Hu, B.~Mallein and M.~Pain. \newblock An exactly solvable continuous-time Derrida--Retaux model. \newblock \emph{Comm. Math. Phys.} \textbf{375} (2020), 605--651. \newblock \href{https://doi.org/10.1007/s00220-019-03465-w}{doi:10.1007/s00220-019-03465-w}. \bibitem[Hu and Shi(2018)]{HS18} Y.~Hu and Z.~Shi. \newblock The free energy in the Derrida--Retaux recursive model. \newblock \emph{J. Stat. Phys.} \textbf{172} (2018), 718--741. \newblock \href{https://doi.org/10.1007/s10955-018-2066-1}{doi:10.1007/s10955-018-2066-1}. \bibitem[Lacoin and Toninelli(2009)]{LT09} H.~Lacoin and F.~L.~Toninelli. \newblock A smoothing inequality for hierarchical pinning models. \newblock In A.~Boutet de Monvel and A.~Bovier (eds.), \emph{Spin Glasses: Statics and Dynamics}, Progress in Probability \textbf{62}, Birkh\"auser, Basel, 2009, pp.~271--278. \newblock \href{https://doi.org/10.1007/978-3-7643-9891-0_12}{doi:10.1007/978-3-7643-9891-0\_12}. Author version dated 16 May 2008. \bibitem[Li and Zhang(2025)]{LZ25} Z.~Li and R.~Zhang. \newblock Asymptotic behavior of the generalized Derrida--Retaux recursive model. \newblock \emph{J. Stat. Phys.} \textbf{192} (2025), article 98. \newblock \href{https://doi.org/10.1007/s10955-025-03480-3}{doi:10.1007/s10955-025-03480-3}. Corrected author version: \href{https://arxiv.org/abs/2411.13068v3}{arXiv:2411.13068v3} (18 August 2025). \bibitem[Lyons, Pemantle and Peres(1995)]{LPP95} R.~Lyons, R.~Pemantle and Y.~Peres. \newblock Conceptual proofs of $L\log L$ criteria for mean behavior of branching processes. \newblock \emph{Ann. Probab.} \textbf{23} (1995), 1125--1138. \newblock \href{https://doi.org/10.1214/aop/1176988176}{doi:10.1214/aop/1176988176}. \bibitem[Monthus and Garel(2008)]{MG08} C.~Monthus and T.~Garel. \newblock Critical points of quadratic renormalizations of random variables and phase transitions of disordered polymer models on diamond lattices. \newblock \emph{Phys. Rev. E} \textbf{77} (2008), article 021132. \newblock \href{https://doi.org/10.1103/PhysRevE.77.021132}{doi:10.1103/PhysRevE.77.021132}. \bibitem[Phragm\'en and Lindel\"of(1908)]{PL08} E.~Phragm\'en and E.~Lindel\"of. \newblock Sur une extension d'un principe classique de l'analyse et sur quelques propri\'et\'es des fonctions monog\`enes dans le voisinage d'un point singulier. \newblock \emph{Acta Math.} \textbf{31} (1908), 381--406. \newblock \href{https://doi.org/10.1007/BF02415450}{doi:10.1007/BF02415450}. \bibitem[Romanov(2014)]{Romanov14} R.~Romanov. \newblock Canonical systems and de Branges spaces. \newblock \href{https://arxiv.org/abs/1408.6022}{arXiv:1408.6022} (2014). \bibitem[Tang and Chat\'e(2001)]{TC01} L.-H.~Tang and H.~Chat\'e. \newblock Rare-event induced binding transition of heteropolymers. \newblock \emph{Phys. Rev. Lett.} \textbf{86} (2001), 830--833. \newblock \href{https://doi.org/10.1103/PhysRevLett.86.830}{doi:10.1103/PhysRevLett.86.830}. \end{thebibliography} \end{document}