\begin{partabstract} We study the Parisi variational problem of the negative spherical perceptron, a mean-field model of jamming, at constraint densities $\alpha=2+e$ with $e>0$ small. We prove that replica symmetry is lost exactly at the de Almeida--Thouless margin, through a continuous transition to full replica symmetry breaking: in a window of width of order $e^2$ ending at this margin the replica-symmetric order parameter is the unique minimizer, and in a window of the same width beyond it every minimizer has a Parisi measure supported on one interval, with density of order $e^{-2}$ and atoms at both endpoints. In the second window the no-overlap-gap hypothesis of El Alaoui and Sellke holds. The de Almeida--Thouless margin lies $\theta_{\mathrm d}e^2(1+o(1))$ below Gardner's prediction $\kRS$ for the capacity, for an explicit $\theta_{\mathrm d}>0$, and $\kRS$ exceeds the critical margin of the functional by exact order $e^3$. With the identification of the capacity in Part~I, Gardner's formula therefore overestimates the capacity by exact order $(\alpha-2)^3$. The proof rests on a compactness theorem for minimizers and on a universal flat limit. In the compactness theorem, a balance law implied by the optimality conditions bounds the depth of the support. A jamming identity for the functional and a bound on the jump at the top of the support give the remaining bounds. \end{partabstract} \part{Full replica symmetry breaking in the negative spherical perceptron} \chapter{Introduction}\label{II:sec:intro} Let $N\ge1$, and let $\mu_N$ be the uniform probability measure on the sphere $S_N=\{\sigma\in\R^N:\norm\sigma^2=N\}$. Fix a constraint density $\alpha>0$ and a margin $\kappa\in\R$, let $M=M_N$ be integers with $M/N\to\alpha$, and let $g_1,\dots,g_M$ be independent standard Gaussian vectors in $\R^N$. The spherical perceptron asks for points of the sphere that satisfy the $M$ random linear constraints \begin{equation}\label{II:eq:constraints} \frac{g_a\cdot\sigma}{\sqrt N}\ge\kappa,\qquad a=1,\dots,M. \end{equation} Let $S_N(\kappa)$ be the set of $\sigma\in S_N$ that satisfy \eqref{II:eq:constraints}, let $V_N(\kappa)=\mu_N(S_N(\kappa))$ be its volume, and let $\kappa_N=\max_{\sigma\in S_N}\min_a g_a\cdot\sigma/\sqrt N$ be the largest achievable margin. The basic questions are the exponential rate $N^{-1}\log V_N(\kappa)$, called the free energy, and the limit of $\kappa_N$, called the capacity. For $\kappa\ge0$ each constraint confines $\sigma$ to a spherical cap, and the feasible set is an intersection of geodesically convex sets. For $\kappa<0$ each constraint removes a spherical cap, and the feasible set need be neither convex nor connected. Franz and Parisi proposed this negative spherical perceptron as the simplest mean-field model of jamming \cite{FranzParisi2016}: its capacity is a satisfiability (SAT--UNSAT) transition, whose critical behavior was argued to be that of the jamming of hard spheres \cite{FPSUZ2017}. Gardner and Derrida computed the free energy by the replica method, under the replica-symmetric ansatz \cite{Gardner1988,GardnerDerrida1988}. Their formula predicts the capacity $\kRS(\alpha)$, the solution of \[ \alpha\,\E(\kRS-G)_+^2=1 , \] where $G$ is a standard Gaussian variable; $\kRS(\alpha)$ is negative exactly when $\alpha>2$. For $\kappa\ge0$ the replica-symmetric formula was proved by Shcherbina and Tirozzi \cite{ShcherbinaTirozzi2003} and by Talagrand \cite[Chapter~8]{TalagrandVolII}; for this and other rigorous work on perceptrons see Part~I, \Cref{I:sec:intro}. For $\kappa<0$ the physical picture is richer \cite{FranzParisi2016,FPSUZ2017}. At small constraint densities the replica-symmetric solution is predicted to be correct. As the constraints become more demanding, it loses stability on a de Almeida--Thouless line \cite{AlmeidaThouless1978}, inside the satisfiable phase. For margins between $\kappa_{\rm 1RSB}\approx-2.05$ and zero \cite{FranzParisi2016}, this instability is predicted to be a continuous transition to full replica symmetry breaking in the sense of Parisi \cite{Parisi1979,Parisi1980Sequence}. In this phase the order parameter, a probability measure that describes the overlaps of typical solutions, has an interval in its support. The jamming line then lies inside the broken phase, and Gardner's formula for the capacity should be an upper bound that is not exact. Annesi, Malatesta, and Zamponi computed the satisfiability threshold numerically from the full replica-symmetry-breaking equations, and located the region of the phase diagram in which the overlap distribution of typical solutions has connected support \cite{AnnesiMalatestaZamponi2025}. Rigorous results for negative margins have been partial. Stojnic proved, using Gordon's comparison inequality \cite{GordonY1985}, that Gardner's prediction $\kRS$ bounds the capacity from above for every margin \cite{Stojnic2013Gardner}. The upper bound of Montanari, Zhong, and Zhou \cite[Theorem~4.2]{MontanariZhongZhou2024} shows that Gardner's formula overestimates the capacity when the margin is sufficiently negative, that is, when $\alpha$ is large. Near $\alpha=2$, where $\kRS(\alpha)$ is close to zero, these bounds do not quantify the separation. El Alaoui and Sellke stated the Parisi variational problem for this model in the form studied below, and conjectured, following \cite{FPSUZ2017}, that it gives the free energy and the capacity \cite[Conjecture~1]{ElAlaouiSellke2022}. In their terms, the model has no overlap gap if some minimizer has a Parisi measure supported on an interval on which its distribution function is strictly increasing \cite[Definition~2]{ElAlaouiSellke2022}. Under this hypothesis they constructed an efficient algorithm that, given oracle access to such a minimizer, produces vectors of norm close to $\sqrt{QN}$ with a small vector of constraint violations, where $Q$ is the top of the support of that minimizer \cite[Theorem~1]{ElAlaouiSellke2022}; they expect these vectors to approximate barycenters of pure states. They left open whether the hypothesis holds for any parameters. In Part~I we prove a Parisi formula for this model (Part~I, \Cref{I:thm:parisi}), and we deduce that, for $\alpha>2$, the free energy converges to the infimum of the Parisi functional below its critical margin and $\kappa_N$ converges to this critical margin (Part~I, \Cref{I:thm:capacity}). This part studies the variational problem itself. The structure of Parisi minimizers has been determined rigorously for spin glasses of Sherrington--Kirkpatrick (SK) type. Following Guerra's interpolation bound \cite{Guerra2003}, Talagrand proved the Parisi formula for the SK model and mixed even $p$-spin models \cite{Talagrand2006}, and Panchenko extended it to general mixtures \cite{Panchenko2014Parisi,PanchenkoBook}. For spherical models with even interactions, Talagrand proved the Crisanti--Sommers formula \cite{Talagrand2006Spherical} (see \cite{CrisantiSommers1992} for the formula), and W.-K.\ Chen proved it for general mixtures \cite{Chen2013}. Auffinger and W.-K.\ Chen proved that the Parisi functional is strictly convex, so that its minimizer is unique \cite{AuffingerChen2015Uniqueness}, and Jagannath and Tobasco gave another proof \cite[Theorem~2]{JagannathTobasco2016}. Auffinger and W.-K.\ Chen also established structural properties of Parisi measures \cite{AuffingerChenProperties}; see also \cite{TalagrandParisiMeasures}. Jagannath and Tobasco recast the Parisi functional through stochastic dynamic programming, which gives the well-posedness of the Parisi equation and its control representation \cite{JagannathTobasco2016}, and they derived first-variation criteria for its minimizer \cite{JagannathTobasco2017}. W.-K.\ Chen developed variational representations and optimality conditions \cite{Chen2017Variational}. By strict convexity, an order parameter that satisfies these first-order conditions is the Parisi measure. Toninelli proved that the replica-symmetric formula fails below the de Almeida--Thouless line of the SK model \cite{Toninelli2002}. Jagannath and Tobasco proved that every replica-symmetric Parisi measure of a mixed $p$-spin model satisfies the de Almeida--Thouless condition, and that replica symmetry holds up to the de Almeida--Thouless line in parts of the phase diagram \cite[Theorems~1.8, 1.9, and~1.13]{JagannathTobasco2017}. W.-K.\ Chen proved that the de Almeida--Thouless line is the exact transition line for the SK model with a centered Gaussian external field \cite{ChenAT2021}, and Lopatto proved that replica symmetry holds up to the de Almeida--Thouless line for the SK model with a uniform external field $h>0$ \cite{LopattoAT}. The support of the Parisi measure has since been determined in several regimes. Auffinger, W.-K.\ Chen, and Zeng proved that the zero-temperature Parisi measure of mixed $p$-spin models has infinitely many points in its support \cite{AuffingerChenZeng2020}. Zhou proved full replica symmetry breaking for the SK model at zero external field and inverse temperatures slightly above the critical value (one in the usual SK normalization) \cite{ZhouFRSB}. There the Parisi measure is supported on an interval, with a smooth density below its endpoint and an atom at the endpoint. A recent preprint extends this description to every inverse temperature above one \cite{LopattoFRSB}, and a recent preprint by H.-B.\ Chen shows that at zero temperature the SK order parameter has full support on $[0,1)$ \cite{Chen2026ZeroTempFRSB}. Another recent preprint constructs mixed even $p$-spin Ising models whose Parisi measures have exactly $k$ steps of replica symmetry breaking, for every $k$, together with transitions from $r$ to $r+1$ steps \cite{AronowLopattoDepth}. These structural results rely on the strict convexity of the Parisi functional and on the resulting characterization of the unique Parisi measure by its first-order conditions. The perceptron differs from these models in three ways. First, the terminal datum of its Parisi equation, \eqref{II:eq:intro-pde} below, is the logarithm of a Gaussian probability, and it becomes singular as the top of the support of the Parisi measure approaches one. This hard wall must be kept at a distance from every minimizer. Second, the functional is not known to be convex. Its first-order conditions are necessary for minimality but are not known to be sufficient, and minimizers are not known to be unique. A candidate that satisfies the first-order conditions therefore cannot be certified as a minimizer, and every minimizer has to be described. Third, the transition studied here takes place in a singular limit with several scales. For $\alpha=2+e$ the relevant margins lie within $O(e^2)$ of $\kRS$, the relevant overlaps lie within $O(e^2)$ of one, and the Parisi measure has a density of order $e^{-2}$. The scaling limit is itself a nontrivial variational problem, and the order parameter must be controlled both near the top of its support and far below it. To our knowledge, full replica symmetry breaking has not been established for any perceptron model. Our contribution is to determine the minimizers near the de Almeida--Thouless line for constraint densities slightly above two. Let $\alpha=2+e$ with $e>0$ small. We prove that the replica-symmetric order parameter is the unique minimizer at the de Almeida--Thouless margin $\kd$ and in a window of width of order $e^2$ below it. In a window of the same width above $\kd$, every minimizer has a Parisi measure supported on a single interval $[q_0,Q]$, with a density of order $e^{-2}$ on $(q_0,Q)$ and atoms at both endpoints (\Cref{II:thm:main}). Thus replica symmetry is lost exactly at the de Almeida--Thouless line, through a continuous transition that passes directly to full replica symmetry breaking, as predicted in \cite{FranzParisi2016,FPSUZ2017}. We also locate the three thresholds. The critical margin $\kc$ of the functional satisfies $\kd<\kc<\kRS$, the distance $\kRS-\kd$ is $\thd e^2+O(e^3\log(1/e))$ for an explicit $\thd>0$, and $\kRS-\kc$ is of exact order $e^3$ (\Cref{II:thm:thresholds}). Combined with the capacity theorem of Part~I, the last statement shows that Gardner's formula overestimates the capacity by an amount of exact order $(\alpha-2)^3$ (Part~I, \Cref{I:cor:thresholds}). The main step is a compactness theorem (\Cref{II:thm:compact}). In variables adapted to the scale $e^2$, it places the order parameter of every minimizer, at all margins $\kRS-\theta e^2$ with $\theta$ in a fixed compact subset of $(0,\infty)$, in one compact class. The obstruction is a thin tail of the support at large depth, which costs little in the value of the functional. We exclude it with a balance law, obtained by integrating the first-order conditions against an integrating factor. On the compact class the functional has a universal limit, the flat functional, whose unique minimizer at the crossing is replica symmetric (\Cref{II:thm:flat}). A local analysis of the first-order conditions near the crossing then gives \Cref{II:thm:main}. \section{The Parisi variational problem}\label{II:sec:intro-problem} An \emph{order parameter} is a right-continuous nondecreasing function $\gamma\colon[0,1]\to[0,1]$ with $\gamma=1$ on a neighborhood of one. We write $\cU$ for the set of order parameters and $\mu_\gamma$ for the probability measure on $[0,1]$ with distribution function $\gamma$, called the \emph{Parisi measure} of $\gamma$. For $\gamma\in\cU$ put \[ \begin{gathered} Q=\inf\{q:\gamma(q)=1\}<1,\qquad \delta=1-Q,\\ \lambda_\gamma(q)=\int_q^1\gamma(t)\dd t,\qquad q_0=\min\supp\mu_\gamma . \end{gathered} \] Thus $\supp\mu_\gamma\subset[q_0,Q]$ and $q_0,Q\in\supp\mu_\gamma$. We call $\lambda_\gamma(0)$ the \emph{mass} of $\gamma$, and we call $\gamma$ \emph{replica symmetric} if $\mu_\gamma$ is a point mass, that is, if $\gamma=\1_{[q,1]}$ for some $q\in[0,1)$. Let $\phi$ and $\Phi$ be the standard Gaussian density and distribution function, and put $\phi_0=\phi(0)=(2\pi)^{-1/2}$. Fix $\alpha>0$ and $\kappa\in\R$. For $\gamma\in\cU$ let $u=u_\gamma(\cdot,\cdot\,;\kappa)$ solve the Parisi equation \begin{equation}\label{II:eq:intro-pde} \partial_qu=-\frac12\bigl(\partial_{xx}u+\gamma(q)(\partial_xu)^2\bigr)\quad\text{on }[0,Q)\times\R,\qquad u(Q,x)=\log\Phi\Bigl(\frac{x-\kappa}{\sqrt\delta}\Bigr). \end{equation} On an interval where $\gamma$ is constant the equation is solved by a Cole--Hopf transformation \cite{Hopf1950,Cole1951}, and for general $\gamma$ the solution is the limit of the solutions for step order parameters (\Cref{II:sec:functional}). This construction goes back to Parisi \cite{Parisi1980Sequence}, the equation \eqref{II:eq:intro-pde} to Duplantier \cite{Duplantier1981}, and the passage to general $\gamma$ uses Lipschitz continuity in the order parameter, due to Guerra \cite{Guerra2001}; see also \cite[Lemma~14 and Remark~15]{JagannathTobasco2016}. For this terminal datum, El Alaoui and Sellke constructed the solution in the same way and proved its regularity \cite[Section~4]{ElAlaouiSellke2022}. The terminal datum is $\log\Prob(x+\sqrt\delta\,G\ge\kappa)$. It is the hard-wall form of the constraint \eqref{II:eq:constraints}, smoothed by the Gaussian fluctuations of the overlaps above $Q$. The \emph{Parisi functional} and its infimum are \begin{equation}\label{II:eq:intro-functional} \cP_\kappa(\gamma)=\alpha\,u_\gamma(0,0;\kappa)+\frac12\int_0^Q\frac{\dd q}{\lambda_\gamma(q)}+\frac12\log\delta,\qquad \cP_*(\kappa)=\inf_{\gamma\in\cU}\cP_\kappa(\gamma). \end{equation} The last two terms of $\cP_\kappa$ form the Crisanti--Sommers entropy of spherical spin glasses \cite{CrisantiSommers1992,Talagrand2006Spherical,Chen2013}. For the replica-symmetric order parameters the functional is explicit: \begin{equation}\label{II:eq:intro-rs} \PRS_\kappa(q):=\cP_\kappa(\1_{[q,1]})=\alpha\E\log\Phi\Bigl(\frac{\sqrt q\,G-\kappa}{\sqrt{1-q}}\Bigr)+\frac q{2(1-q)}+\frac12\log(1-q)\qquad(0\le q<1). \end{equation} This is Gardner's replica-symmetric functional \cite{Gardner1988,GardnerDerrida1988}, and its minimum over $q$ is Gardner's formula for the free energy. The \emph{critical margin} of the functional is \begin{equation}\label{II:eq:intro-kc} \kc(\alpha)=\sup\{\kappa\in\R:\cP_*(\kappa)>-\infty\}. \end{equation} By \Cref{II:fun:finite,II:fun:existence,II:fun:critical}, $\kc(\alpha)$ is finite, $\kc(\alpha)\le\kRS(\alpha)$, and $\kc(\alpha)<0$ for $\alpha>2$. For $\kappa<\kc$ the infimum $\cP_*(\kappa)$ is attained, and every minimizer satisfies $1-Q>\e^{2\cP_*(\kappa)}$. Moreover $\cP_*(\kappa)\to-\infty$ as $\kappa\uparrow\kc$. Every theorem of this part is a deterministic statement about $\cP_\kappa$ and its minimizers. The results of \Cref{II:fun:sec:value,II:fun:sec:thresholds} are proved in Part~I, \Cref{I:sec:hardwall}, and are only stated here. No other proof in this part uses Part~I. The random model enters only through the consequences of Part~I recalled after \Cref{II:thm:thresholds,II:thm:main}. Unlike the Parisi functionals of SK-type models, $\cP_\kappa$ is not known to be convex, and its minimizers are not known to be unique in general. Our theorems therefore describe every minimizer. By a minimizer we always mean a global minimizer of $\cP_\kappa$ over $\cU$. The first-order conditions for minimizers are expressed through the \emph{Parisi diffusion} \[ \dd X_q=\gamma(q)\,\partial_xu(q,X_q)\dd q+\dd B_q,\qquad X_0=0, \] where $B$ is a standard Brownian motion, and through the functions \begin{equation}\label{II:eq:intro-D} D(q)=\alpha\E\bigl[\partial_xu(q,X_q)^2\bigr]-\int_0^q\frac{\dd t}{\lambda_\gamma(t)^2},\qquad \cH_\gamma(t)=-\frac12\int_0^tD(q)\dd q . \end{equation} The right derivative of $\vartheta\mapsto\cP_\kappa(\gamma+\vartheta(\widetilde\gamma-\gamma))$ at $\vartheta=0$ equals $\int\cH_\gamma\dd\mu_{\widetilde\gamma}-\int\cH_\gamma\dd\mu_\gamma$. Consequently every minimizer $\gamma$ satisfies the obstacle condition $\supp\mu_\gamma\subset\argmin_{[0,1)}\cH_\gamma$ (\Cref{II:prop:kkt}); in particular $D=0$ on $\supp\mu_\gamma$. These conditions are necessary but, in the absence of convexity, not known to be sufficient. For a replica-symmetric order parameter $\1_{[q,1]}$ the diffusion is a Brownian motion up to the level $q$, and everything is explicit. Let $R=\phi/\Phi$, and let \[ V(z)=R(z)\bigl(z+R(z)\bigr)=-\frac{\dd^2}{\dd z^2}\log\Phi(z), \] which takes values in $(0,1)$. The derivative of $\PRS_\kappa$ at $q$ is $-\frac12D(q)$, where $D$ is the function \eqref{II:eq:intro-D} of $\1_{[q,1]}$, and one half of the curvature of the obstacle potential at the atom is \begin{equation}\label{II:eq:intro-replicon} \Lambda_{\mathrm R}=\frac12\cH_{\1_{[q,1]}}''(q)=\frac{1-\alpha\E V(Z)^2}{4(1-q)^2},\qquad Z=\frac{\sqrt q\,G-\kappa}{\sqrt{1-q}} \end{equation} (\Cref{II:rs:lem-formulas}). This is the replicon eigenvalue of the replica-symmetric solution. The equation $\alpha\E V(Z)^2=1$ is the de Almeida--Thouless condition of the physics analysis of this model \cite[(52)]{FPSUZ2017}; for perceptrons, the stability of the replica-symmetric solution was studied by Gardner and Derrida \cite{GardnerDerrida1988} and by Bouten \cite{Bouten1994}. If $q$ is a critical point of $\PRS_\kappa$ and $\Lambda_{\mathrm R}<0$, then moving a small amount of mass from the atom at $q$ to a nearby level lowers the functional, so $\1_{[q,1]}$ is not a minimizer. \section{Main results}\label{II:sec:intro-results} Three margins organize our results: Gardner's prediction $\kRS$, the critical margin $\kc$, and the de Almeida--Thouless margin $\kd$, which we now define. For $0<\alpha<\pi^2/4$ and $\kappa<\kRS(\alpha)$, the function $\PRS_\kappa$ has exactly one critical point $q^*(\kappa)$ in $(0,1)$, which is its unique minimizer on $[0,1)$ (\Cref{II:rs:lem-unique} and \Cref{II:rs:lem-formulas}(c)). Thus $\1_{[q^*(\kappa),1]}$ is the only replica-symmetric order parameter whose atom satisfies $D=0$, and $\PRS_\kappa(q^*(\kappa))=\min_q\PRS_\kappa(q)$ is Gardner's formula for the free energy. Let $\Lambda_{\mathrm R}(\kappa)$ denote the replicon eigenvalue \eqref{II:eq:intro-replicon} at $q=q^*(\kappa)$. For $\alpha=2+e$ with $e>0$ sufficiently small, there is exactly one margin $\kd(\alpha)\in(\kRS-c_2,\kRS)$ with $\Lambda_{\mathrm R}(\kd)=0$, where $c_2>0$ is a numerical constant, and \[ \Lambda_{\mathrm R}(\kappa)<0\quad(\kd<\kappa<\kRS),\qquad \Lambda_{\mathrm R}(\kappa)>0\quad(\kRS-c_2\le\kappa<\kd) \] (\Cref{II:prop:crossing}). We call $\kd(\alpha)$ the \emph{de Almeida--Thouless margin}, and put $\qd=q^*(\kd)$ and $\gamma_{\mathrm d}=\1_{[\qd,1]}$. On $(\kd,\kRS)$ the replica-symmetric solution is unstable. Finally, put \[ c_0=\int_\R V(z)\bigl(1-V(z)\bigr)\dd z\in(0,\infty),\qquad \thd=\frac{(2\pi)^{3/2}}{128\,c_0^2} . \] Our first result is the ordering of the three margins near $\alpha=2$. \begin{theorem}\label{II:thm:thresholds} There are numerical constants $e_0,c>0$ and $C<\infty$ such that for every $00$, yet its error is one order smaller than the length of the unstable interval. The expansion of $\kRS-\kd$ is part of \Cref{II:prop:crossing}. The two cubic bounds use a zero-temperature characterization of $\kc$, as the threshold at which a functional $J^k$ with the quadratic terminal penalty $-\frac12(k-x)_+^2$ first takes a negative value (\Cref{II:thm:kJ}). The upper bound on $\kRS-\kc$ comes from an explicit dual certificate for $J^k$: a single function $g\ge k-x$ with $\alpha\E g(G)^2\le1$ and $\alpha\E g'(G)^2\le1$ forces $J^k\ge0$ (\Cref{II:rs:thm-dual,II:thm:ordering}). The lower bound comes from an explicit coefficient, carried by a layer of width of order $e^2$ just below the top, at which $J^k$ is negative for $k=\kRS-ce^3$ (\Cref{II:thm:cubic}). Part~I, \Cref{I:thm:gardner}, proves the same lower bound by evaluating $\cP_\kappa$ directly at an order parameter with two support points near one, whose top layer plays the role of this coefficient. The constants $c$ and $C$ are not matched, and we do not identify the limit of $(\kRS-\kc)/e^3$. \Cref{II:thm:thresholds} concerns the functional. In Part~I we prove that, for every $\alpha>2$, \begin{equation}\label{II:eq:intro-capacity} \frac1N\log V_N(\kappa)\longrightarrow\cP_*(\kappa)\quad(\kappa<\kc(\alpha)),\qquad \kappa_N\longrightarrow\kc(\alpha), \end{equation} in probability, and that $\Prob\{N^{-1}\log V_N(\kappa)>-C\}\to0$ for every $\kappa\ge\kc(\alpha)$ and every $C<\infty$ (Part~I, \Cref{I:thm:capacity}). In Part~I this capacity theorem is combined with \Cref{II:thm:thresholds}, \Cref{II:cor:nonrs} and \Cref{II:fun:rates} (Part~I, \Cref{I:cor:thresholds,I:cor:logvolume}). For $\alpha=2+e$ with $e>0$ small enough, the capacity $\kc(\alpha)$ satisfies $c\,e^3\le\kRS(\alpha)-\kc(\alpha)\le C\,e^3$, and Gardner's formula for the free energy fails at every margin in $(\kd,\kRS)$. For every $\alpha>2$, the free energy diverges only logarithmically at the capacity. The limit $N\to\infty$ is taken at fixed $e$, and the bounds describe how this limit behaves as $e\downarrow0$. Our main theorem describes all minimizers at, below, and above the de Almeida--Thouless margin. Put \[ B_V=\int_\R V(z)^2\bigl(1-V(z)\bigr)\dd z,\qquad x_{\mathrm f}=\frac12-\frac{c_0}{4B_V} . \] By \Cref{II:frsb:lem-flatconst}, $x_{\mathrm f}=\int_\R V'(z)^2\dd z\big/(2B_V)$ and $00$ and $C<\infty$ such that the following holds for every $00$ put \[ \kappa_e(\theta)=\kRS(2+e)-e^2\theta , \] and measure overlaps from the top through the rescaled level $s=(1-q)/e^2$. For $\gamma\in\cU$ let \[ \zeta(s)=\gamma(1-e^2s),\qquad d=\frac{1-Q}{e^2},\qquad s_0=\frac{1-q_0}{e^2},\qquad \ell_0=\frac{\lambda_\gamma(0)}{e^2} \] be the \emph{rescaled profile}, the rescaled \emph{top gap}, the rescaled \emph{depth}, and the rescaled \emph{mass}. The profile $\zeta$ is nonincreasing, equals one on $[0,d]$, vanishes beyond $s_0$, and has integral $\ell_0$. A replica-symmetric order parameter has $d=s_0=\ell_0$. By \Cref{II:thm:thresholds}, $\kd=\kappa_e(\thd+O(e\log(1/e)))$, and by \Cref{II:thm:ordering}(iii) the margins $\kappa_e(\theta)$, $\theta$ in a compact interval $\Theta\subset(0,\infty)$, lie below $\kc$ by at least $\frac12e^2\min\Theta$ for small $e$, so that minimizers exist. \begin{theorem}\label{II:thm:compact} Let $\Theta\subset(0,\infty)$ be a compact interval. There are constants \[ 00, \] depending only on $\Theta$, such that for every $00 . \] \end{theorem} In the original variables, the support of every minimizer lies in $[1-S_*e^2,\,1-d_1e^2]$, its mass lies between $\ellmin e^2$ and $Le^2$, and the jump of $\gamma$ at the top of the support is at least $1-m_*$. The difficult statement is the bound on the depth $s_0$; the other bounds use it. Comparison with replica-symmetric trial order parameters and the first-order condition at the bottom of the support give only $s_0\le C_0e^{-2/3}$ (\Cref{II:thm:upper}). This leaves room for minimizers with a large atom near the top, above a thin tail of small mass that extends to depths much larger than $e^2$. Such a tail costs little in the value of the functional, and indeed near-minimizers can have tails of vanishing mass at arbitrary depth (\Cref{II:rem:nearmin}). Thus no estimate of the minimal value alone yields \Cref{II:thm:compact}; the proof uses the first-order conditions of exact minimizers. Finally, the rescaled problem has a universal limit as $e\to0$. A \emph{profile} is a nonincreasing function $\zeta\colon[0,\infty)\to[0,1]$; it is \emph{normalized} if $\int\zeta=1$. The \emph{normalized profile} of $\gamma$ is $\widetilde\zeta(s)=\zeta(\ell_0s)=\gamma(1-e^2\ell_0s)$, and the replica-symmetric normalized profile is $\1_{[0,1]}$. For normalized profiles $\zeta$ with bounded support that equal one on some interval $[0,d_1]$, we define in \Cref{II:def:flat} the \emph{flat functional} \[ F_t(\zeta)=A(\zeta)+t\,I(\zeta),\qquad t\in\R . \] Here $A(\zeta)$ is an explicit entropy term, and $I(\zeta)$ is a renormalized integral over $y\in\R$ of the solution $U=U_\zeta$ of \begin{equation}\label{II:eq:intro-flatpde} \partial_sU=\frac12\bigl(\partial_{yy}U+\zeta(s)(\partial_yU)^2\bigr),\qquad U(s,y)=\log\Phi\bigl(y/\sqrt s\bigr)\quad(00$, depending only on $\Theta$, $d_1$, and $S$, such that the following holds for $00$ (\Cref{II:thm:kJ}). For every $\alpha>2$ the infimum diverges logarithmically at the critical margin, $\alpha\log\eta-C_\alpha\le\cP_*(\kc-\eta)\le\frac12\log\eta+C_\alpha$ for small $\eta>0$ (\Cref{II:fun:rates}). Finally, for every $\alpha>0$, every margin, and every order parameter with $q_0>0$, the value $\cP_\kappa(\gamma)$ equals the replica-symmetric value $\PRS_\kappa(1-\lambda_\gamma(0))$ of the same mass plus an integral over the levels $q\in[q_0,Q]$, with weight $(1-\gamma(q))/(2\lambda_\gamma(q))$, of an explicit coefficient (\Cref{II:cpt:lem-jam} and \eqref{II:cpt:eq-samemass}). We call this the jamming identity. Throughout, $c$ and $C$ denote positive numerical constants, and constants with subscripts, such as $C_\alpha$, depend only on the indicated quantities. Their values may change from one occurrence to the next. Unless stated otherwise, $O(\cdot)$ refers to $e\downarrow0$ with a numerical implied constant. No constant depends on the order parameter under consideration. \section{Ideas of the proof}\label{II:sec:intro-ideas} The proof has two stages. The first is global: it places the rescaled profiles of all minimizers, at all margins $\kappa_e(\theta)$ with $\theta$ in a compact interval, in one compact class (Steps~1--3). The second is local: on this class the functional has a universal limit, which forces minimizers near $\kd$ to be close to $\gamma_{\mathrm d}$, and the first-order conditions can then be analyzed near $\gamma_{\mathrm d}$ (Steps~4 and~5). \Cref{II:fig:map} shows how the steps depend on one another. Throughout we use the function $v=-\lambda_\gamma\partial_{xx}u$, which takes values in $(0,1)$, and the scalar functions \[ S(q)=\alpha\E v(q,X_q)^2,\qquad \chi(q)=\frac{\alpha\E v(q,X_q)-1}{\lambda_\gamma(q)},\qquad w(q)=\alpha\E\bigl[v(1-v)\bigr](q,X_q). \] The function $D$ of \eqref{II:eq:intro-D} is continuously differentiable with $D'=(S-1)/\lambda_\gamma^2$, and every minimizer satisfies $D=0$ and $S\le1$ on its support (\Cref{II:lem:generator,II:prop:kkt}). For a minimizer whose support is an interval, El Alaoui and Sellke proved $D=0$ and $S=1$ on the support; see the remark after \Cref{II:prop:kkt}. \begin{figure}[htbp] \centering \input{figures/part2-map} \caption{Selected dependencies in the two stages of the proof. Each solid box names a step (or, at the left, \Cref{II:sec:rs}), the chapters or sections that carry it out and, where there is one, the theorem it proves; an arrow runs from a box to a box that uses it. The dashed box is the consequence for the random model, proved in Part~I. The general facts about the Parisi functional in \Cref{II:fun:sec:gauss,II:fun:sec:def,II:fun:sec:reg,II:fun:sec:gen,II:fun:sec:kkt} and \Cref{II:app:functional} are used throughout and are not drawn. The results of \Cref{II:fun:sec:value,II:fun:sec:thresholds} are proved in Part~I, \Cref{I:sec:hardwall}.} \label{II:fig:map} \end{figure} \paragraph{Step 1: A priori bounds.} Two inputs give the preliminary bounds. The first is the zero-temperature functional $J^k$. The certified margin $k_*\ge\kRS-Ce^3$ of the dual certificate lies below $\kc$. The resulting slack of $J^k$ (\Cref{II:fun:slack}), together with a universal bound on the mass of minimizers in terms of their top gap (\Cref{II:lem:weaktop}), keeps the top away from the wall: $d\ge ce^2$ (\Cref{II:thm:gapD}). The second input is the first-order condition $D(q_0)=0$ at the bottom of the support. Since the Parisi diffusion is a Brownian motion up to the level $q_0$, this condition is an exact Gaussian identity at a single level. In the rescaled space variable $y=(x-\kappa)/e$, the negative half-line controls the mass, and a window of width of order $\sqrt{s_0}$ to the right of zero controls the depth. This gives $\ell_0\le L$ and $s_0\le C_0e^{-2/3}$ (\Cref{II:lem:bottom,II:thm:upper}). These bounds are proved in \Cref{II:sec:functional,II:sec:rs,II:sec:apriori}. \paragraph{Step 2: Bounded depth.} To exclude thin tails at large depth we use the balance law \begin{equation}\label{II:eq:intro-balance} \bigl(\lambda_\gamma D\bigr)'=\chi(0)-\frac w{\lambda_\gamma}\qquad\text{almost everywhere on }[0,1), \end{equation} valid for every order parameter that satisfies the first-order conditions (\Cref{II:lem:if}). It follows from $D'=(S-1)/\lambda_\gamma^2$, the identity $S-1=\lambda_\gamma\chi-w$, and the relation $\chi=\chi(0)+\gamma D$, which holds because $D$ vanishes on the support. The integrating factor $\lambda_\gamma$ cancels the unknown term $\gamma D$ exactly. Integrating \eqref{II:eq:intro-balance} between $q_0$ and $Q$, where $D$ vanishes, and passing to rescaled variables gives \begin{equation}\label{II:eq:intro-balance-int} \int_d^{s_0}\frac{w(s)}{\ell(s)}\dd s=\widetilde\chi\,(s_0-d),\qquad \widetilde\chi=e^2\chi(0)=\frac{\alpha\E v(s_0,Y_0)-1}{\ell_0}, \end{equation} where $\ell(s)=\lambda_\gamma(1-e^2s)/e^2\le\ell_0$ and $Y_0$ is the rescaled position of the diffusion at the bottom of the support, which is Gaussian. The right side is computed through an exact cancellation. At every level, $v-\1_{\{y<0\}}$ has integral zero in the rescaled space variable $y$ (\Cref{II:lem:cancel}), so the Gaussian average $\alpha\E v(s_0,Y_0)$ equals $\alpha\Prob(Y_0<0)$ up to $O(e^2(1+s_0))$, and $\widetilde\chi\ell_0=\frac e4+o(e)$. For the left side, a heat-semigroup lower bound for the rescaled derivative of the solution and a density floor for the diffusion (\Cref{II:lem:tilted}) give $w(s)\ge c_we\sqrt s$ at every level $s\in[d,s_0]$. Hence \[ \frac{2c_we}{3\ell_0}\bigl(s_0^{3/2}-d^{3/2}\bigr)\le\frac e{3\ell_0}(s_0-d), \] and for a minimizer that is not replica symmetric, $s_0>d$ and this forces $s_0\le1/(4c_w^2)$. The factors $e$ and $\ell_0$ cancel, so no lower bound on the mass or the top gap is needed. This is carried out in \Cref{II:sec:depth}. \paragraph{Step 3: Compactness.} It remains to bound the mass and the top gap from below and the mass below the top atom from above. The main tool is the jamming identity (\Cref{II:cpt:lem-jam}). In rescaled variables it reads \begin{equation}\label{II:eq:intro-jam} \cP_\kappa(\gamma)=\Pi(\ell_0;q_0)+\int_d^{s_0}\frac{1-\zeta(t)}{2\ell(t)}\,\mathcal J(t)\dd t , \end{equation} where $\Pi(\ell_0;q_0)=\alpha\E\log\Phi((\sqrt{q_0}\,G-\kappa)/\sqrt{\lambda_\gamma(0)})+q_0/(2\lambda_\gamma(0))+\frac12\log\lambda_\gamma(0)$ is the replica-symmetric expression \eqref{II:eq:intro-rs} with the bottom and the mass decoupled, and $\mathcal J$ is an explicit jamming coefficient with values in $(1-\alpha,1)$. Near $\alpha=2$, for profiles of bounded depth, \[ \mathcal J(t)\ge c\,e\min\Bigl\{\frac{\ell_0}{\sqrt{\ell(t)}},\frac1{2e}\Bigr\}-Ce \] (\Cref{II:cpt:lem-J}). Thus the coefficient is at least $-Ce$ at every level, and it is large near the top, where $\ell(t)$ is small. This gives three bounds. A profile of bounded depth and small mass has a value above the replica-symmetric value, so $\ell_0\ge\ellmin$ (\Cref{II:prop:massLB}). The first-order conditions express the mass $\gamma(Q-)$ below the top atom through a ratio of Gaussian averages of derivatives of $u$ near the top of the support (\Cref{II:cpt:lem-topKKT}). The tilted-density representation of the law of the diffusion and a Chebyshev association inequality \cite{BoucheronLugosiMassart2013} bound this ratio, so that $\gamma(Q-)\le m_*<1$ (\Cref{II:thm:topjump,II:app:compact}). This bound is global: it does not assume closeness to a replica-symmetric order parameter. Finally, if the top gap is small and $\gamma(Q-)\le m_*$, the profile is jammed just below the wall, where $\mathcal J$ is large, and \eqref{II:eq:intro-jam} shows that the replica-symmetric order parameter of the same mass has a smaller value (\Cref{II:cpt:prop-gap}). Hence $d>d_1$. This is carried out in \Cref{II:sec:compact}, which completes the proof of \Cref{II:thm:compact}. \paragraph{Step 4: The flat limit.} On the compact class, \Cref{II:thm:flat}(ii) holds (\Cref{II:thm:compactexp}). To prove \Cref{II:thm:flat}(i) we parametrize a normalized profile by the inverse $T=\ell^{-1}$ of its primitive $\ell$. In this coordinate the entropy term $A$ is affine (\Cref{II:flat:prop-inverse}), and $I$ is convex along every segment from the replica-symmetric profile to a normalized step profile (\Cref{II:flat:cor-rayconv}). The convexity comes from an exact formula for the second variation (\Cref{II:flat:prop-zform}), whose integrand is nonnegative because an associated flat density is monotone (\Cref{II:flat:prop-mono}). It then suffices that the first variation of $F_{\td}$ at $\1_{[0,1]}$ is positive in every admissible direction. Along the segment in the inverse coordinate from $\1_{[0,1]}$ toward a normalized step profile with inverse profile $T$, this first variation equals $\int_{(0,1]}\mathcal E(r)\dd\dot\sigma(r)$, where $\dot\sigma=(T-\mathrm{id})'$ is nonnegative and nondecreasing. The kernel $\mathcal E(r)$ is a positive weighted integral, over $\rho\in(r,1)$, of the decrease of $\int V^2$ when Lebesgue measure is transported for the time $\log(1/\rho)$ by the replica-symmetric top diffusion \[ \dd Z_\varsigma=\Bigl(R(Z_\varsigma)+\frac{Z_\varsigma}2\Bigr)\dd\varsigma+\dd B_\varsigma . \] A synchronous coupling with an Ornstein--Uhlenbeck process gives $\mathcal E(r)\ge(1-r)^3/96$ (\Cref{II:flat:prop-fs}), and convexity turns this cubic contact into the modulus $\frac9{1024}$ of \eqref{II:eq:intro-flatmod}. Combining the expansion with the modulus gives a stability estimate that is uniform in $\theta$ near $\thd$ (\Cref{II:flat:cor-nearmin}): every minimizer at the margin $\kappa_e(\theta)$ satisfies \[ \bigl(\ell_0-2\phi_0\theta\bigr)^2+e\,\bigl\|\widetilde\zeta-\1_{[0,1]}\bigr\|_{L^1}^3\le C\bigl(e^2\log(1/e)+e|\theta-\thd|\bigr). \] Consequently, for every $\varepsilon>0$ there is $\rho>0$ such that for small $e$, every minimizer at every margin within $\rho e^2$ of $\kd$ is within $\varepsilon(1-\qd)$ of $\gamma_{\mathrm d}$ in $L^1$ (\Cref{II:ext:cor-flat}). The flat-limit analysis is carried out in \Cref{II:sec:flat,II:app:flat}; the last consequence is proved in \Cref{II:frsb:ss-loc}. \paragraph{Step 5: Local analysis at the crossing.} The last step turns closeness in $L^1$ into the structure asserted in \Cref{II:thm:main}. First, the supports localize (\Cref{II:prop:localize}): for margins within a small multiple of $e^2$ of $\kd$, every support point of every minimizer lies within a small multiple of $e^2$ of $\qd$. The mechanism is the obstacle condition, written through the obstacle potential anchored at a support point and compared with the obstacle potential of $\gamma_{\mathrm d}$ at $\kd$. The latter is explicit and has a strict cubic minimum at $\qd$ of depth proportional to $e$, because the replicon eigenvalue vanishes at the crossing (\Cref{II:frsb:prop-rsobstacle}). The comparison uses continuity of $S$ at the natural scale, uniformly over order parameters close to $\gamma_{\mathrm d}$ (\Cref{II:frsb:lem-natcont}) and over the margins in the window (\Cref{II:ext:lem-margin}). Second, window rigidity (\Cref{II:prop:window}) describes every order parameter that satisfies the first-order conditions and has its support in such a window. On a gap of the support the order parameter is constant, and the first-order conditions force a local curvature $\widetilde\Sigma$ to be nonpositive somewhere in the gap (\Cref{II:lem:gap} and \Cref{II:frsb:lem-gapint}). In the window, the ratio of $\widetilde\Sigma$ to the positive function $\E[v^2(1-v)](q,X_q)$ is close to a positive constant of the flat limit (\Cref{II:frsb:lem-natcont}), so $\widetilde\Sigma>0$ and the support is an interval. On an interval of the support, $D\equiv0$ and $S\equiv1$, which give the density of $\mu_\gamma$ explicitly; it is of order $e^{-2}$. Finally, if $Q>q_0$, comparing the bottom of the interval with the replica-symmetric order parameter of the same mean (\Cref{II:frsb:lem-bottom}) gives the width estimate \[ c\,e\,\frac{Q-q_0}{e^2}\le-\tau(\kappa),\qquad \tau(\kappa)=4\bigl(1-q^*(\kappa)\bigr)^2\Lambda_{\mathrm R}(\kappa), \] where $\tau(\kappa)$ has the sign of $\kd-\kappa$. At and below $\kd$ this forces $Q=q_0$, and the minimizer is replica symmetric. Above $\kd$ it gives $Q-q_0\le C(\kappa-\kd)$, while $q_0=Q$ is excluded because the replicon eigenvalue is negative. This is carried out in \Cref{II:sec:frsb,II:app:local}, and the main theorems are assembled in \Cref{II:sec:proofs}. Steps~3 and~5 follow arguments for the SK model. The top-jump bound of Step~3 follows the endpoint strategy of \cite[Theorem~4]{AuffingerChenProperties}, as in \cite[Proposition~9.3]{LopattoFRSB}: the first-order conditions at the top of the support relate the mass below the top atom to Gaussian averages of derivatives of $u$ there, and these averages are then bounded. Step~5 follows the near-critical architecture of Zhou \cite{ZhouFRSB}: closeness to the replica-symmetric order parameter, then positivity of a local curvature, then exclusion of gaps of the support through the first-order conditions at their endpoints. In \cite[Section~9]{LopattoFRSB} gaps are excluded through the same endpoint conditions, combined with a sign condition inside the gap. Compared to these works, the local analysis differs in three ways. The functional here is not known to be convex, so the first-order conditions do not identify a minimizer, and every minimizer has to be localized before the local analysis applies. The windows of margins have width $ce^2$ with numerical constants, uniformly over all minimizers. The Parisi measures have atoms at both endpoints of the support, not only at the top. In every step, the terminal datum is the hard wall, which becomes singular as the top of the support approaches one. The differences in Step~3 are listed in \Cref{II:cpt:ssec-top}. \medskip The steps are arranged so that no step uses a later conclusion; the logical order is recorded in \Cref{II:proofs:ss-order}. \Cref{II:app:functional} proves the regularity of the Parisi equation used in \Cref{II:sec:functional}. \chapter{The Parisi functional and its optimality conditions}\label{II:sec:functional} This chapter collects the deterministic facts about the hard-wall Parisi functional on which the rest of this part is built. \Cref{II:fun:sec:gauss,II:fun:sec:def,II:fun:sec:reg,II:fun:sec:gen,II:fun:sec:kkt} contain general structural facts, valid for every order parameter and every margin. \Cref{II:fun:gauss} collects the facts about Gaussian functions used throughout this part. We define the functional through a finite-cutoff form of the Parisi equation (\Cref{II:def:orderparam,II:fun:cutoff}), prove that its solution is regular and depends Lipschitz continuously on the order parameter (\Cref{II:lem:regularity}), and record the Cole--Hopf description on constant blocks together with a comparison principle (\Cref{II:fun:colehopf}). We then introduce the diffusion along which all first-order quantities are evaluated (\Cref{II:fun:diffusion}), describe its law (\Cref{II:lem:tilted}), and state the stochastic control representation (\Cref{II:fun:control}). \Cref{II:lem:generator} gives the generator identities for the derivatives of the solution along the diffusion, and \Cref{II:lem:sandwich} gives a pointwise sandwich inequality. Finally, \Cref{II:prop:kkt} computes the first variation and derives the obstacle conditions, of Karush--Kuhn--Tucker (KKT) type, satisfied by minimizers, and \Cref{II:lem:gap} turns these conditions into a principle for gaps and intervals of the support. The remaining sections, \Cref{II:fun:sec:value,II:fun:sec:thresholds,II:fun:sec:J,II:fun:sec:rates}, study the variational problem itself; their contents are described at the beginning of \Cref{II:fun:sec:value}. Besides the first-order conditions, Step~1 in \Cref{II:sec:intro-ideas} uses three results of these sections: the value inequality (\Cref{II:lem:value}), the slack of the zero-temperature functional (\Cref{II:fun:slack}) and the bound on the mass in terms of the top gap (\Cref{II:lem:weaktop}). Every statement of \Cref{II:fun:sec:gauss,II:fun:sec:def,II:fun:sec:reg,II:fun:sec:gen,II:fun:sec:kkt} holds for general order parameters, not only for step order parameters. The one exception is the control representation of \Cref{II:fun:control}, which is stated for step coefficients and is used only through inequalities that extend to general order parameters by continuity. The functional $\cP_\kappa$ defined in \eqref{II:fun:P} below is the Parisi functional of the negative spherical perceptron with hard-wall activation \cite{FranzParisi2016,ElAlaouiSellke2022}; see also Part~I, \Cref{I:sec:model}. Nothing in this part depends on that interpretation. Many of the facts in \Cref{II:fun:sec:def,II:fun:sec:reg,II:fun:sec:gen,II:fun:sec:kkt} have counterparts in earlier work, which are credited in remarks next to the individual results. What is specific here is the following. The terminal datum is the hard wall $\log\Phi((x-\kappa)/\sqrt\delta)$, which is increasing and concave in $x$ rather than even, and which is singular as $Q\to1$; this is why all estimates are stated in the classes $\cU_r$. The first variation is computed at every $\gamma\in\cU_r$ in every direction of $\cU_r$, and the obstacle conditions are derived for every global minimizer and every KKT point, with no assumption on the support; the remark after \Cref{II:prop:kkt} compares this with \cite[Proposition~16]{ElAlaouiSellke2022}. Part~I, \Cref{I:sec:functional}, derives several of its lemmas on the hard-wall functional from the results of \Cref{II:fun:sec:gauss,II:fun:sec:def,II:fun:sec:reg,II:fun:sec:gen,II:fun:sec:kkt} and \Cref{II:app:functional}. \section{Gaussian functions and Mills ratios}\label{II:fun:sec:gauss} Throughout, $G$ denotes a standard Gaussian variable, $\phi$ and $\Phi$ denote its density and distribution function, and $\phi_0=\phi(0)=(2\pi)^{-1/2}$. For $q>0$ let $\phi_q(x)=q^{-1/2}\phi(x/q^{1/2})$ be the $N(0,q)$ density. We use the functions \begin{equation}\label{II:fun:RetaV} R(z)=\frac{\phi(z)}{\Phi(z)},\qquad \Rbar(z)=z+R(z),\qquad V(z)=R(z)\Rbar(z),\qquad h(z)=\log\Phi(z)+\tfrac12R(z)^2 , \end{equation} and \[ m(k)=\E(k-G)_+^2,\qquad m_1(k)=\E(k-G)_+,\qquad \varrho(z)=R(z)^2-z_-^2 . \] The function $\log\Phi((x-\kappa)/\sqrt\delta)$ will be the terminal datum of the Parisi equation. Its first two derivatives in $x$ are $R/\sqrt\delta$ and $-V/\delta$, which explains the role of $R$ and $V$ below. For $z\in\R$ let $T_z$ be the random variable $z+G$ conditioned on the event $\{z+G\ge0\}$; its density on $[0,\infty)$ is proportional to $\e^{zt-t^2/2}$. \begin{lemma}\label{II:fun:gauss} \begin{enumerate}[label=\textup{(\alph*)}] \item $R>0$, $\Rbar>0$, $00$, and \begin{equation}\label{II:fun:khintchine} 1-V(z)\ge\tfrac13\Rbar(z)^2\qquad(z\le0). \end{equation} \item For $t>0$, \[ \frac{t\phi(t)}{1+t^2}<\Phi(-t)<\frac{\phi(t)}{t}. \] Consequently $t0$, and $\Rbar(-t)\ge t^{-1}-3t^{-3}$ and $0<1-V(-t)<3t^{-2}$ for $t\ge2$. \item For $t\ge0$, $-\tfrac12t^2-\log(1+t)-2\le\log\Phi(-t)\le-\tfrac12t^2$. For $z\ge0$, $\log\Phi(z)\ge-2\Phi(-z)$. \item $h<0$ and $h'=R(1-V)>0$. \item $V^2\le 2R$. For $z\ge0$, $R(z)\le2\phi(z)$ and $V(z)\le2(1+z)\phi(z)$. \item The constant $c_0:=\int_\R V(1-V)\dd z$ lies in $(0,\infty)$, and \[ \int_\R\bigl(V-\1_{\{z<0\}}\bigr)\dd z=0,\qquad \int_\R\bigl(V^2-\1_{\{z<0\}}\bigr)\dd z=-c_0 . \] The function $\varrho$ is continuous and positive, $M_\varrho:=\sup\varrho<\infty$, and there is a numerical constant $C_{\rm g}$ with \[ |V(z)^2-\1_{\{z<0\}}|+|\varrho(z)-2\cdot\1_{\{z<0\}}|\le\frac{C_{\rm g}}{1+z^2}\qquad(z\in\R). \] \item $m(k)=(1+k^2)\Phi(k)+k\phi(k)$, $m_1(k)=k\Phi(k)+\phi(k)=\Phi(k)\Rbar(k)>0$, $m'=2m_1$, $m_1'=\Phi$, $m''=2\Phi$, and $\Phi-m=-k\,m_1$. \end{enumerate} \end{lemma} Most of these facts are classical. The two bounds in (b) are Gordon's bounds for the Mills ratio \cite{Gordon1941}, and the bound $V<1$ in (a) is due to Sampford \cite{Sampford1953}. The proof of \eqref{II:fun:khintchine} uses Khintchine's representation of unimodal distributions (see \cite{DharmadhikariJoagDev1988}). Parts of (a)--(e) are used in Part~I, \Cref{I:pf:gauss} and \Cref{I:hw:gauss}; in \Cref{I:hw:gauss}(v) the constant $2$ in the lower bound of (c) is replaced by $-\log\Phi(-1)<2$. \begin{proof} (a) Integrating against $\phi$ on $[-z,\infty)$ and using $\phi'(t)=-t\phi(t)$ gives $\E[G\mid G\ge-z]=R(z)$ and $\E[G^2\mid G\ge-z]=1-zR(z)$, so \[ \E T_z=\Rbar(z),\qquad \Var(T_z)=1-zR(z)-R(z)^2=1-V(z). \] Hence $\Rbar>0$, $V=R\Rbar>0$, and $V<1$, since $T_z$ is positive and not almost surely constant; also $\E T_z^2=1-V+\Rbar^2=1+\Rbar(\Rbar-R)=1+z\Rbar$. The identities $R'=-zR-R^2=-V$ and $\Rbar'=1-V$ follow by differentiation. In particular $R$ is decreasing, so $R(z)\le R(0)=2\phi_0$ for $z\ge0$, while $R(z)-R(0)=\int_z^0V\le|z|$ for $z<0$; and $R>-z$ (from $\Rbar>0$) gives $R>z_-$. Next \[ V'=R'\Rbar+R\Rbar'=R\big(1-V-\Rbar^2\big)=R\big\{\Var(T_z)-(\E T_z)^2\big\}, \] and differentiating once more, with $V(1-V-\Rbar^2)=\Rbar V'$, \[ V''=-V(1-V-\Rbar^2)+R\bigl(-V'-2\Rbar(1-V)\bigr)=-(\Rbar+R)V'-2V(1-V), \] which is \eqref{II:fun:Vpp} because $\Rbar+R=z+2R$. To see $V'<0$, it suffices that $\E T_z^2<2(\E T_z)^2$. The survival function $H(s)=\Prob(T_z>s)$ has hazard rate $\mathfrak m(s-z)$, where $\mathfrak m(s)=\phi(s)/\Phi(-s)$. Since $\mathfrak m'=\mathfrak m(\mathfrak m-s)$ and $\mathfrak m(s)>s$ (by (b) for $s>0$, and because $\mathfrak m>0$ for $s\le0$), the hazard rate is strictly increasing. Hence $H(s+t)0$, and integrating over $s,t>0$ gives \[ \tfrac12\E T_z^2=\int_0^\infty sH(s)\dd s<\Big(\int_0^\infty H\Big)^2=(\E T_z)^2 . \] Then $\Rbar''=-V'>0$. Finally let $z\le0$. The density $p(t)=c\,\e^{zt-t^2/2}$ of $T_z$ on $[0,\infty)$ is $C^1$, nonincreasing, and tends to zero at infinity. Hence \[ p(t)=\int_t^\infty\bigl(-p'(s)\bigr)\dd s=\int_{(0,\infty)}\frac{\1_{\{t0$ and $j\ge0$ set $I_j(t)=\int_t^\infty\phi(s)s^{-j}\dd s$. Since $\phi'(s)=-s\phi(s)$, integration by parts gives the recurrence \begin{equation}\label{II:fun:mills-rec} I_j(t)=\phi(t)t^{-j-1}-(j+1)I_{j+2}(t),\qquad 00$, this gives $R(-t)>t(1-\varepsilon+3\varepsilon^2)^{-1}\ge t(1+\varepsilon-3\varepsilon^2)$, because $(1-\varepsilon+3\varepsilon^2)(1+\varepsilon-3\varepsilon^2)=1-(\varepsilon-3\varepsilon^2)^2\le1$ and both factors are positive. Thus $\Rbar(-t)>t^{-1}-3t^{-3}$. For $t\ge2$ both lower bounds $R(-t)>t$ and $\Rbar(-t)>t^{-1}-3t^{-3}$ are positive, so $V(-t)=R(-t)\Rbar(-t)>1-3t^{-2}$. (c) The upper bound is the Chernoff bound $\Prob(G\ge t)\le \e^{-t^2/2}$. For $t\ge1$, (b) gives $\Phi(-t)\ge\phi(t)/(2t)\ge\phi(t)/(2(1+t))$, and $\log(2\sqrt{2\pi})<2$. For $0\le t\le1$, $\log\Phi(-t)\ge\log\Phi(-1)>-2$. The last claim follows from $\log(1-p)\ge-2p$ for $0\le p\le\frac12$. (d) By (a), $h'=R+RR'=R(1-V)>0$. As $z\to\infty$, $\log\Phi(z)\to0$ and $R(z)\to0$, so $h<0$. (e) For $z\le0$, $\Rbar(z)\le\Rbar(0)=2\phi_0<2$, so $V^2\le V=R\Rbar\le2R$. For $z\ge0$, $\Phi(z)\ge\frac12$ gives $R\le2\phi$, and $R(z)\le R(0)<1$ gives $\Rbar(z)\le z+1$; hence $V=R\Rbar\le2(1+z)\phi(z)$. Moreover $V^2/R=R\Rbar^2\le2\phi(z)(1+z)^2\le8\phi(1)<2$, because $(1+z)^2\phi(z)$ is maximal at $z=1$. (f) Positivity of $c_0$ follows from $00$; for $z<0$, $\varrho=\Rbar(z)(R(z)-z)>0$. For the pointwise bound, $1-V^2\le2(1-V)\le6/z^2$ for $z\le-2$, and $V^2\le V\le2(1+z)\phi(z)$ for $z\ge0$. For $z=-t\le-2$ we have $\varrho(-t)=\Rbar(-t)(R(-t)+t)$, which by (b) lies between $(\frac1t-\frac3{t^3})\cdot2t=2-\frac6{t^2}$ and $\frac1t(2t+\frac1t)=2+\frac1{t^2}$; for $z\ge0$, $\varrho=R^2\le4\phi^2$. On $[-2,0]$ both functions are bounded. These estimates also give $M_\varrho<\infty$. (g) These are direct computations: $m_1(k)=k\Phi(k)-\E[G;G0$ define the Gaussian recursion \begin{equation}\label{II:fun:recursion} \cT_{m,s}g(x)=\frac1m\log\E\exp\big(m\,g(x+\sqrt s\,G)\big)\quad(m>0),\qquad \cT_{0,s}g(x)=\E g(x+\sqrt s\,G). \end{equation} If $r=0$, then $\cU_0=\{\1_{[0,1]}\}$, and we set $u(0,\cdot)=f_0$. If $r>0$ and $\gamma\in\cU_r$ is a step order parameter, choose $0=t_00$, \begin{equation}\label{II:fun:Tprops} \cT_{m,s}g\le\cT_{m,s}g',\qquad \cT_{m,s}(g+C)=\cT_{m,s}g+C,\qquad \cT_{m,s}g=K^{-1}\,\cT_{m/K,s}(Kg). \end{equation} The first two properties are immediate from \eqref{II:fun:recursion}, and the third follows from $\frac Km\log\E\e^{(m/K)Kg}=\frac Km\log\E\e^{mg}$ if $m>0$, and from linearity if $m=0$. Consequently the recursion \eqref{II:fun:steprec} is nondecreasing in its terminal datum and commutes with the addition of constants. On a block where $\gamma=m>0$, the recursion \eqref{II:fun:steprec} says that $\e^{mu}$ is transported by the backward heat equation. This is the Cole--Hopf transform of the Parisi equation \begin{equation}\label{II:fun:pde} u_q=-\tfrac12\big(u_{xx}+\gamma\,u_x^2\big), \end{equation} since for constant $\gamma=m$ the function $\e^{mu}$ solves $\partial_q \e^{mu}=-\frac12\partial_{xx}\e^{mu}$ exactly when $u$ solves \eqref{II:fun:pde}. \Cref{II:fun:colehopf} below records the consequences used later. The next lemma shows that the cutoff $r$ is only a device. \begin{lemma}\label{II:fun:cutoff} Let $\gamma\in\cU_r$ and $r\le r'<1$. Then $\gamma\in\cU_{r'}$, and the solutions computed with the cutoffs $r$ and $r'$ agree on $[0,r]\times\R$. Consequently $u_\gamma(\cdot,\cdot\,;\kappa)$ is defined on $[0,1)\times\R$ for every $\gamma\in\cU$, independently of the cutoff, and \begin{equation}\label{II:fun:top} u_\gamma(q,x;\kappa)=\log\Phi\Big(\frac{x-\kappa}{\sqrt{1-q}}\Big)\qquad(Q\le q<1). \end{equation} \end{lemma} \begin{proof} Since $\gamma=1$ on $[r,r']$, the recursion with cutoff $r'$ performs one step with $m=1$ on $[r,r']$. With $G'$ an independent standard Gaussian, \[ \cT_{1,r'-r}f_{r'}(x)=\log\Prob\big(\sqrt{1-r'}\,G'\le x-\kappa+\sqrt{r'-r}\,G\big)=\log\Phi\Big(\frac{x-\kappa}{\sqrt{1-r}}\Big)=f_r(x). \] Below $r$ the two recursions coincide, which proves the claim for step order parameters. The general case follows by taking limits in \Cref{II:def:orderparam}(iii), since step approximations in $\cU_r$ also lie in $\cU_{r'}$. For $q\in[Q,1)$ we have $\gamma\in\cU_q$, and the cutoff $q$ gives $u(q,\cdot)=f_q$, which is \eqref{II:fun:top}. \end{proof} In particular $u(Q,x)=\log\Phi((x-\kappa)/\sqrt\delta)$, which is the usual terminal datum at the top of the support. We use the notation $u$, $b$, $c$, $v$ on $[0,1)\times\R$ without reference to a cutoff. On $[Q,1)$, with $z=(x-\kappa)/\sqrt{1-q}$, formula \eqref{II:fun:top} and $\lambda(q)=1-q$ give \begin{equation}\label{II:fun:topderiv} b=\frac{R(z)}{\sqrt{1-q}},\qquad c=\frac{V(z)}{1-q},\qquad v=V(z). \end{equation} For $\alpha>0$ and $\kappa\in\R$ the \emph{Parisi functional} and its infimum are \begin{equation}\label{II:fun:P} \begin{gathered} \cP_\kappa(\gamma)=\alpha\,u_\gamma(0,0;\kappa)+\CS(\gamma),\qquad \CS(\gamma)=\frac12\int_0^Q\frac{\dd q}{\lambda(q)}+\frac12\log\delta,\\ \cP_*(\kappa)=\inf_{\gamma\in\cU}\cP_\kappa(\gamma). \end{gathered} \end{equation} We minimize $\cP_\kappa$. When the dependence on $\alpha$ matters we write $\cP_\kappa(\gamma;\alpha)$ and $\cP_*(\kappa;\alpha)$. Since $\lambda(q)=1-q$ on $[Q,1]$, the Crisanti--Sommers entropy $\CS(\gamma)$ satisfies, for every $r\in[Q,1)$, \begin{equation}\label{II:fun:entropy} \CS(\gamma)=\frac12\int_0^r\frac{\dd q}{\lambda(q)}+\frac12\log(1-r)=\frac12\int_0^1\Big(\frac1{\lambda(q)}-\frac1{1-q}\Big)\dd q\ \ge0 , \end{equation} where the inequality holds because $\lambda(q)\le1-q$. Thus on $\cU_r$ the functional involves only the solution on $[0,r]$ and the entropy term $\frac12\int_0^r\lambda^{-1}+\frac12\log(1-r)$, in which $\lambda\ge1-r$. For $\gamma=\1_{[0,1]}$ one has $Q=0$, $\CS(\gamma)=0$, and $\cP_\kappa(\gamma)=\alpha\log\Phi(-\kappa)<0$; hence $\cP_*(\kappa)<0$ for every $\kappa$. \section{Regularity, constant blocks, and the diffusion}\label{II:fun:sec:reg} We first record the properties of $u$ proved in \Cref{II:app:functional}. All constants in the following lemma depend only on the cutoff $r$, on a bound $K_0\ge|\kappa|$, and on the order of differentiation. They do not depend on $\gamma$ or, for step order parameters, on the number of steps. \begin{lemma}\label{II:lem:regularity} Fix $r\in[0,1)$ and $K_0\ge0$. Let $\gamma\in\cU_r$, $|\kappa|\le K_0$, and $u=u_\gamma(\cdot,\cdot\,;\kappa)$. \begin{enumerate}[label=(\alph*)] \item \emph{Smoothness.} For every $k\ge0$, $\partial_x^ku$ exists and is continuous on $[0,r]\times\R$, and \[ u(q,x)=f_r(x)+\frac12\int_q^r\big(u_{xx}+\gamma\,u_x^2\big)(s,x)\dd s\qquad(q\in[0,r],\ x\in\R). \] If $\gamma$ is constant on an interval $[s,t)\subset[0,r)$, then $u$ is $C^\infty$ on $[s,t)\times\R$, its derivatives from this block extend continuously to $[s,t]\times\R$, and \eqref{II:fun:pde} holds classically for $sQ$ the integral is oriented, $\int_q^Q=-\int_Q^q$. There are $C,x_\infty$ such that for $q\in[0,r]$ and $x\le\kappa-x_\infty$, \[ \begin{gathered} \Big|u(q,x)+\frac{(x-\kappa)^2}{2\lambda(q)}+\log(\kappa-x)-\omega(q)\Big|\le\frac{C}{(\kappa-x)^2},\\ \Big|\lambda(q)b(q,x)-(\kappa-x)-\frac{\lambda(q)}{\kappa-x}\Big|\le\frac C{(\kappa-x)^3}. \end{gathered} \] \item \emph{Continuity in $\gamma$.} For $\gamma,\widetilde\gamma\in\cU_r$ and $k\ge0$, \[ \big|\partial_x^ku_{\widetilde\gamma}(q,x;\kappa)-\partial_x^ku_{\gamma}(q,x;\kappa)\big|\le C_k\,\|\widetilde\gamma-\gamma\|_{L^1(0,1)}\,(1+|x|^{p_k})\qquad(q\in[0,r]), \] where one may take $p_0=2$ and $p_1=1$. Moreover $\cU_r$ is compact in $L^1(0,1)$, and $|\cP_\kappa(\widetilde\gamma;\alpha)-\cP_\kappa(\gamma;\alpha)|\le C(1+\alpha)\|\widetilde\gamma-\gamma\|_{L^1}$. In particular $\cP_\kappa$ is continuous on $\cU_r$, uniformly for $|\kappa|\le K_0$. \item \emph{Translation.} $u_\gamma(q,x;\kappa)=u_\gamma(q,x-\kappa;0)$. \end{enumerate} \end{lemma} Every $\gamma\in\cU$ belongs to $\cU_r$ for all $r\in[Q,1)$, and by \Cref{II:fun:cutoff} the solution does not depend on $r$. Hence (a)--(e) describe $u$ on all of $[0,1)\times\R$, with constants that are uniform on $[0,r]$. The proofs of (a)--(d), (f), and (g) are in \Cref{II:appA:sec:step,II:appA:sec:cont}. The proof of (e) uses \Cref{II:lem:generator,II:lem:sandwich} below, whose proofs do not use (e), and is given in \Cref{II:appA:sec:asym}. On $[Q,1)$ all statements can be read off from \eqref{II:fun:top}. For instance $\omega(q)=\frac12\log(1-q)-\frac12\log(2\pi)$ there, which is the constant in the expansion $\log\Phi(z)=-\frac12z^2-\log|z|-\log\sqrt{2\pi}+O(z^{-2})$ at $z=(x-\kappa)/\sqrt{1-q}\to-\infty$. Below $Q$ the constant evolves by $\omega'=(1-2\gamma)/(2\lambda)$, which is what \eqref{II:fun:pde} forces on the expansion in (e). The construction of $u$ for general $\gamma$, growth bounds, the non-strict form of the bounds on $c$ in (c), and a local form of the Lipschitz bound (f), for $x$ in bounded sets, are in \cite[Propositions~12 and~13, Lemma~14]{ElAlaouiSellke2022}. The proof of (f) rests on a Feynman--Kac identity for the difference of two solutions (\Cref{II:appA:diff}). This identity and the resulting Lipschitz bound appear in \cite[Lemma~14]{JagannathTobasco2016}, and the Lipschitz bound goes back to Guerra \cite{Guerra2001} (see \cite[Remark~15]{JagannathTobasco2016}). Next we record the Cole--Hopf description of $u$ on an interval where $\gamma$ is constant, and the monotonicity of $u$ in the order parameter. Together they are the form in which Jensen's inequality enters later chapters. \begin{lemma}\label{II:fun:colehopf} Let $\gamma\in\cU$ and $\kappa\in\R$, and write $u=u_\gamma(\cdot,\cdot\,;\kappa)$. \begin{enumerate}[label=(\alph*)] \item \emph{Constant blocks.} If $0\le s0$ this reads $\e^{mu(q,x)}=\E\,\e^{mu(t,x+\sqrt{t-q}\,G)}$, so $\e^{mu}$ solves the backward heat equation on $[s,t]\times\R$. \item \emph{Comparison.} If $\widetilde\gamma\in\cU$, $q'\in[0,1)$ and $\widetilde\gamma\le\gamma$ on $[q',1]$, then $u_{\widetilde\gamma}(q',\cdot\,;\kappa)\le u_\gamma(q',\cdot\,;\kappa)$. In particular $u_\gamma(q',\cdot\,;\kappa)$ depends only on the restriction of $\gamma$ to $[q',1]$. \item \emph{Heat lower bound.} For $0\le q\le t<1$ and $x\in\R$, $u(q,x)\ge\E\,u(t,x+\sqrt{t-q}\,G)$. \end{enumerate} \end{lemma} \begin{proof} (a) For step order parameters this is the definition \eqref{II:fun:steprec}, after refining the grid so that $s$ and $t$ are grid points. For general $\gamma$ it is obtained in \Cref{II:appA:sec:cont} by passing to the limit along step order parameters equal to $m$ on $[s,t)$. The formula for $\lambda$ follows from $\lambda'=-m$ on $[s,t)$. (b) Choose $r\in(q',1)$ with $\gamma,\widetilde\gamma\in\cU_r$. \Cref{II:appA:diff}, applied with $\xi=\widetilde\gamma$ and $\widetilde\xi=\gamma$, gives \[ u_\gamma(q',x;\kappa)-u_{\widetilde\gamma}(q',x;\kappa)=\frac12\,\E\int_{q'}^r(\gamma-\widetilde\gamma)(t)\,b_{\widetilde\gamma}(t,Y_t)^2\dd t\ \ge0, \] where $Y$ is the diffusion of that lemma started at $(q',x)$. If $\widetilde\gamma=\gamma$ on $[q',1]$, the inequality holds in both directions. (c) We may assume $q0$, the drift is nonnegative, and therefore \begin{equation}\label{II:fun:XgeB} X_q\ge B_q\qquad\text{and}\qquad X^{s,x}_q\ge x+B_q-B_s\qquad(q\ge s). \end{equation} For a general $\gamma$ the function $u$ is not differentiable in $q$ at the discontinuities of $\gamma$, so It\^o's formula along $X$ is used in the integrated form below. Let $\mathscr D$ be the set of functions of the form \[ g(q,x)=\sum_i\pi_i\big(q,x,\lambda(q),\lambda(q)^{-1}\big)\prod_j\partial_x^{k_{ij}}u(q,x), \] finite sums with polynomials $\pi_i$ and integers $k_{ij}\ge0$. For $g\in\mathscr D$ we define $\cL g$ formally: we compute $\partial_q$ of each factor with the rules $\partial_q\partial_x^ku=-\frac12\partial_x^{k+2}u-\frac12\gamma\,\partial_x^k(u_x^2)$ (the $x$-derivatives of \eqref{II:fun:pde}) and $\lambda'=-\gamma$, and add $\frac12g_{xx}+\gamma bg_x$. Since these rules are affine in $\gamma$, we can write $\cL g=\cL_0g+\gamma\,\cL_1g$ with $\cL_0g,\cL_1g\in\mathscr D$ not involving $\gamma$. Where $\gamma$ is constant, $\cL g$ is the classical generator applied to $g$. The formal time derivative is a derivation on $\mathscr D$, and it commutes with $\partial_x$, because the rule for $\partial_q\partial_x^{k+1}u$ is the $x$-derivative of the rule for $\partial_q\partial_x^ku$. Hence the formal generator obeys the product rule and the commutation rule \begin{equation}\label{II:fun:Lrules} \cL(fg)=f\,\cL g+g\,\cL f+f_xg_x,\qquad \cL(g_x)=\partial_x(\cL g)+\gamma\,c\,g_x\qquad(f,g\in\mathscr D), \end{equation} where the second identity uses $\partial_x(\gamma bg_x)=\gamma bg_{xx}-\gamma cg_x$. \begin{lemma}\label{II:fun:diffusion} Fix $r\in[0,1)$, $K_0\ge0$, and $p\ge1$; let $\gamma\in\cU_r$ and $|\kappa|\le K_0$. \begin{enumerate}[label=(\alph*)] \item $\E\sup_{s\le q\le r}|X^{s,x}_q|^p\le C(1+|x|^p)$, with $C=C(r,K_0,p)$. \item If $\widetilde\gamma\in\cU_r$ and $\widetilde X^{s,x}$ is the diffusion of $\widetilde\gamma$ driven by the same Brownian motion, then $\E\sup_{s\le q\le r}|\widetilde X^{s,x}_q-X^{s,x}_q|^p\le C(1+|x|^p)\|\widetilde\gamma-\gamma\|_{L^1}^p$. \item \emph{It\^o's formula along $X$.} For $g\in\mathscr D$ and $s\le t\le r$, \[ \E\,g(t,X^{s,x}_t)-g(s,x)=\int_s^t\E\big[\cL_0g+\gamma(q)\cL_1g\big](q,X^{s,x}_q)\dd q, \] and $q\mapsto\E h(q,X^{s,x}_q)$ is continuous on $[s,r]$ for every $h\in\mathscr D$. Consequently $e_g(q)=\E g(q,X_q)$ is Lipschitz on $[0,r]$; its right derivative at $q$ is $\E[\cL_0g+\gamma(q)\cL_1g](q,X_q)$ and its left derivative is $\E[\cL_0g+\gamma(q-)\cL_1g](q,X_q)$. If $\gamma$ is constant on $[s',t')$, then $e_g$ is $C^\infty$ on $[s',t']$ (with one-sided derivatives at the endpoints), and its derivatives are obtained by iterating $\cL$ with $\gamma$ frozen at its value on $[s',t')$. \end{enumerate} \end{lemma} The proof is in \Cref{II:appA:sec:ito}. In particular, if $\cL g=0$, then $g(q,X_q)$ is a martingale on $[0,1)$: applying (c) to the diffusion started from $(s,X_s)$ and using the Markov property gives $\E[g(t,X_t)\mid X_s]=g(s,X_s)$ for $s\le t<1$. The first-order quantities of \Cref{II:fun:sec:gen,II:fun:sec:kkt} are expectations over the law of $X_q$. The following structural fact about this law is used for lower bounds on densities in \Cref{II:sec:depth} and for the control of the top of the support in \Cref{II:sec:compact}. At $q=0$ the law of $X_0$ is a point mass, so the statement concerns $q>0$ only. \begin{lemma}\label{II:lem:tilted} Let $\gamma\in\cU$, $\kappa\in\R$, and $q\in(0,1)$. The law of $X_q$ has a density $p_q$ of the form \begin{equation}\label{II:fun:tilted} p_q(x)=\e^{\gamma(q)u(q,x)}\,\phi_q(x)\,g_q(x), \end{equation} where $g_q:\R\to[0,\infty)$ is nonincreasing. Moreover $g_q\ge1$ and $\e^{\gamma(q)u(q,x)}\ge\Phi\big((x-\kappa)/\sqrt{1-q}\big)$. Consequently \begin{equation}\label{II:fun:floor} p_q(x)\ge\Phi\Big(\frac{x-\kappa}{\sqrt{1-q}}\Big)\phi_q(x)\qquad(x\in\R). \end{equation} \end{lemma} \begin{proof} We divide the proof into three steps. First we compute the density when $\gamma$ is piecewise constant below $q$, then we approximate, and finally we prove the two lower bounds. \proofstep{1}{$\gamma$ piecewise constant on $[0,q)$.} Suppose that $\gamma=m_j$ on $[t_{j-1},t_j)$ for $0=t_0<\dots0$, the function $H=\e^{m_ju}$ is positive, bounded by one, and solves $H_q+\frac12H_{xx}=0$ by \Cref{II:fun:colehopf}(a). The drift of $X$ on this block is $m_jb=H_x/H$. Thus $X$ is the Doob transform \cite{Doob1957,RevuzYor1999} of Brownian motion by $H$ on this block: for $t_{j-1}\le s0$. Monotonicity gives $g_n(x)\mathfrak P(x)\le\int_{-\infty}^x\mathfrak p g_n\le1$, so $\sup_ng_n\le1/\mathfrak P(x_0)$ on $[x_0,\infty)$. By Helly's selection theorem and a diagonal argument, a subsequence converges pointwise on $\R$ to a nonincreasing function $g\ge0$. For continuous $F$ with compact support in $(x_0,\infty)$, dominated convergence gives $\E F(X^n_q)=\int F\mathfrak p g_n\to\int F\mathfrak p g$, while $\E F(X^n_q)\to\E F(X_q)$. Hence the law of $X_q$ coincides with $\mathfrak p g\dd x$ on every half-line $(x_0,\infty)$, and therefore on $\R$. \proofstep{3}{The lower bounds.} Since $u\le0$ and $g_q$ is nonincreasing, for every $x$ \[ \Prob(X_q>x)=\int_x^\infty \e^{\gamma(q)u}\phi_q\,g_q\le g_q(x)\,\Prob(B_q>x), \] while $\Prob(X_q>x)\ge\Prob(B_q>x)$ by \eqref{II:fun:XgeB}. Hence $g_q\ge1$. For the second bound put $m=\gamma(q)$. For $q\ge Q$ equality holds by \eqref{II:fun:top}, so let $q0$, a nondecreasing right-continuous step function $\xi:[0,T)\to[0,\infty)$ and $f\in\mathfrak F$, let $\Upsilon=\Upsilon_{\xi,f}$ be defined on $[0,T]\times\R$ by \eqref{II:fun:steprec} with $r$ replaced by $T$, $f_r$ by $f$, and $\gamma$ by $\xi$. A \emph{control} is a progressively measurable process $\theta$ on $[0,T]$ with $\theta=0$ almost everywhere on $\{\xi=0\}$ and $\E\int_0^T\theta_q^2/\xi(q)\dd q<\infty$, with the convention $0/0=0$. \begin{lemma}\label{II:fun:control} In this setting, $\Upsilon$ and $\Upsilon_x$ are continuous on $[0,T]\times\R$, $\Upsilon$ is smooth and satisfies $\Upsilon_q=-\frac12(\Upsilon_{xx}+\xi\Upsilon_x^2)$ on each open interval of constancy of $\xi$, and \[ \begin{gathered} \E f(x+\sqrt{T-q}\,G)\le\Upsilon(q,x)\le0,\qquad 0\le\Upsilon_x(q,x)\le C(1+|x|),\\ 0\le-\Upsilon_{xx}\le\Lip(f'). \end{gathered} \] These bounds hold for $0\le q0$, the lemma applies with $T=r$, $\xi=\gamma$, and $f=f_r$, and then $\Upsilon=u_\gamma(\cdot,\cdot\,;\kappa)$ and $X^*=X$ when $x=0$. For the SK and mixed $p$-spin models, Auffinger and W.-K.\ Chen proved the control representation with the Bou\'e--Dupuis formula \cite[Theorem~3]{AuffingerChen2015Uniqueness}, \cite{BoueDupuis1998}, and Jagannath and Tobasco proved it by dynamic programming \cite[Lemma~18]{JagannathTobasco2016}. \section{Generator identities and the sandwich inequality}\label{II:fun:sec:gen} The first-order optimality conditions are expressed through expectations of derivatives of $u$ along $X$. We collect the identities for these expectations here. Besides $b,c,v$ we use the function \begin{equation}\label{II:fun:psi} \psi(q,x)=\lambda(q)\,b(q,x)+x-\kappa , \end{equation} which is harmonic for $\cL$, and the scalar functions of $q\in[0,1)$ \begin{equation}\label{II:fun:scalars} \begin{gathered} D(q)=\alpha\E b(q,X_q)^2-\int_0^q\frac{\dd t}{\lambda(t)^2},\qquad S(q)=\alpha\E v(q,X_q)^2,\qquad \chi(q)=\frac{\alpha\E v(q,X_q)-1}{\lambda(q)},\\ w(q)=\alpha\E\big[v(1-v)\big](q,X_q),\qquad \cN(q)=\E\,v_x(q,X_q)^2,\qquad \widetilde B(q)=\E\big[v^2(1-v)\big](q,X_q). \end{gathered} \end{equation} Thus $S=\alpha\lambda^2\E c(q,X_q)^2$, and $\cN>0$ and $\widetilde B>0$ by \Cref{II:lem:regularity}(c). For $K\in\R$ put \begin{equation}\label{II:fun:Sigma} \Sigma(q;K)=\E\Big[v_{xx}^2+4K(3v-1)v_x^2+2K^2v^2(1-v)(1-3v)\Big](q,X_q),\qquad \widetilde x(q)=\frac{\lambda(q)\cN(q)}{2\widetilde B(q)} . \end{equation} \begin{lemma}\label{II:lem:generator} Let $\gamma\in\cU$ and $\kappa\in\R$. \begin{enumerate}[label=(\alph*)] \item With $K=\gamma/\lambda$, in the sense of the formal generator of \Cref{II:fun:diffusion} (and classically where $\gamma$ is constant), \[ \begin{gathered} \cL u=\tfrac12\gamma\,b^2,\qquad \cL b=0,\qquad \cL c=\gamma\,c^2,\qquad \cL c_x=3\gamma\,c\,c_x,\qquad \cL\psi=0,\\ \cL v=-K\,v(1-v),\qquad \cL v_x=K(3v-1)v_x,\qquad \cL v_{xx}=K\big(3v_x^2+(4v-1)v_{xx}\big). \end{gathered} \] \item $\psi_x=1-v\in(0,1)$ and $\psi_{xx}=-v_x>0$, so $\psi(q,\cdot)$ is increasing and strictly convex for every $q\in[0,1)$. The processes $b(q,X_q)$ and $\psi(q,X_q)$ are martingales on $[0,1)$: for $0\le s\le t<1$, $b(s,x)=\E b(t,X^{s,x}_t)$ and $\psi(s,x)=\E\psi(t,X^{s,x}_t)$. For $s\le Q$ this gives \begin{equation}\label{II:fun:psirep} \psi(s,x)=\E\,\psi(Q,X^{s,x}_Q)=\sqrt\delta\,\E\,\Rbar\Big(\frac{X^{s,x}_Q-\kappa}{\sqrt\delta}\Big)\ \ge\ \sqrt\delta\,\Rbar\Big(\frac{\sqrt\delta\,(x-\kappa)}{\lambda(s)}\Big)>0 , \end{equation} and $b(0,0)=\delta^{-1/2}\E R(Z)$ with $Z=(X_Q-\kappa)/\sqrt\delta$. On $[Q,1)$, $\psi=\sqrt{1-q}\,\Rbar\big((x-\kappa)/\sqrt{1-q}\big)>0$. \item The functions $\E b(q,X_q)^2$, $S$, $\E v(q,X_q)$, $\chi$, $w$, $\cN$, and $\widetilde B$ are Lipschitz on $[0,r]$ for every $r<1$, and the following hold. \begin{enumerate}[label=(\roman*)] \item $\frac{\dd}{\dd q}\E b(q,X_q)^2=\E c(q,X_q)^2$, a continuous function. Hence $D\in C^1([0,1))$ and $D'=(S-1)/\lambda^2$. \item $S/\lambda^2=\alpha\E c(q,X_q)^2$ is nondecreasing on $[0,1)$. \item $q\mapsto\E v(q,X_q)$ is absolutely continuous with $\frac{\dd}{\dd q}\E v(q,X_q)=-\frac{\gamma(q)}{\lambda(q)}\E[v(1-v)](q,X_q)$ for almost every $q$, and $\chi'=\gamma D'$ almost everywhere. \item At every $q\in[0,1)$ the functions $S$, $\cN$, and $\widetilde B$ have right derivatives \[ \begin{gathered} S'(q+)=\alpha\big(\cN(q)-2K\widetilde B(q)\big),\qquad \cN'(q+)=\E\big[v_{xx}^2+2K(3v-1)v_x^2\big](q,X_q),\\ \widetilde B'(q+)=\E\big[(1-3v)v_x^2-Kv^2(1-v)(2-3v)\big](q,X_q), \end{gathered} \] with $K=\gamma(q)/\lambda(q)$, and left derivatives given by the same formulas with $K=\gamma(q-)/\lambda(q)$. \item $S-1=\lambda\chi-w$. \end{enumerate} \item If $\gamma=m$ on $[s,t)$ with $0\le s0$, since $\Rbar>0$. For the lower bound, let $M_q=(X^{s,x}_q-\kappa)/\lambda(q)$ for $q\in[s,Q]$. It\^o's formula and $\lambda'=-\gamma$ give \begin{equation}\label{II:fun:Mq} \dd M_q=\frac{\gamma(q)}{\lambda(q)^2}\,\psi(q,X^{s,x}_q)\dd q+\frac{\dd B_q}{\lambda(q)} . \end{equation} The drift is nonnegative because $\psi>0$, and the martingale part has a deterministic bounded integrand. Hence $\E M_Q\ge M_s=(x-\kappa)/\lambda(s)$. Since $\lambda(Q)=\delta$, we have $(X^{s,x}_Q-\kappa)/\sqrt\delta=\sqrt\delta M_Q$, and Jensen's inequality for the convex increasing function $\Rbar$ gives the lower bound. The formula on $[Q,1)$ is \eqref{II:fun:topderiv}. (c) By \Cref{II:fun:diffusion}(c) each of the listed functions is Lipschitz on $[0,r]$; for $\chi$ also use $\lambda\ge1-r$. By \eqref{II:fun:Lrules}, $\cL(b^2)=2b\cL b+b_x^2=c^2$ contains no factor $\gamma$, so $\E b^2$ is $C^1$ with derivative $\E c^2$, which is continuous by \Cref{II:fun:diffusion}(c). As $S=\alpha\lambda^2\E c^2$, this gives $D'=\alpha\E c^2-\lambda^{-2}=(S-1)/\lambda^2$, which is (i). Next $\cL(c^2)=2c\cL c+c_x^2=2\gamma c^3+c_x^2\ge0$, so the Lipschitz function $\E c^2$ has a nonnegative right derivative everywhere and is nondecreasing; this is (ii). By (a), $\cL v=-Kv(1-v)$, which gives the derivative of $\E v$ at every continuity point of $\gamma$, hence almost everywhere. Then, using $(1/\lambda)'=\gamma/\lambda^2$, \[ \chi'=\frac{\alpha\frac{\dd}{\dd q}\E v}{\lambda}+\frac{\gamma(\alpha\E v-1)}{\lambda^2}=\frac{\gamma}{\lambda^2}\big(-\alpha\E[v(1-v)]+\alpha\E v-1\big)=\frac{\gamma(S-1)}{\lambda^2}=\gamma D' \] almost everywhere, which is (iii). For (iv), (a) and \eqref{II:fun:Lrules} give \[ \begin{gathered} \cL(v^2)=v_x^2-2Kv^2(1-v),\qquad \cL(v_x^2)=v_{xx}^2+2K(3v-1)v_x^2,\\ \cL(v^3)=3v\,v_x^2-3Kv^3(1-v), \end{gathered} \] and $\widetilde B=\E v^2-\E v^3$. Since each of these expressions is affine in $\gamma$ through $K$, \Cref{II:fun:diffusion}(c) gives the one-sided derivatives. Identity (v) is $\lambda\chi-w=\alpha\E v-1-\alpha\E v+\alpha\E v^2$. (d) On $[s,t)$ we have $\lambda'=-m$, so $K'=K^2$. By \Cref{II:fun:diffusion}(c) the functions $S$, $\cN$, and $\widetilde B$ are $C^\infty$ on $[s,t]$, and their derivatives are given by (c)(iv) with $\gamma=m$. The first formula for $S'$ is (c)(iv), and the second is its rewriting with \eqref{II:fun:Sigma}. Differentiating $S'/\alpha=\cN-2K\widetilde B$ and using $K'=K^2$, \begin{align*} \frac{S''}\alpha={}&\E v_{xx}^2+2K\E[(3v-1)v_x^2]-2K^2\widetilde B\\ &-2K\E[(1-3v)v_x^2]+2K^2\E[v^2(1-v)(2-3v)]=\Sigma(q;K), \end{align*} since $2-3v-1=1-3v$. This proves (d). \end{proof} The identities $\cL b=0$ and $\cL c=\gamma c^2$ in (a) are the hard-wall forms of \cite[Lemma~2]{AuffingerChen2015Uniqueness}, and the one-sided derivatives in (c)(iv) are the analogue of \cite[Proposition~3]{AuffingerChenProperties}. The process $X_q+\lambda(q)b(q,X_q)$ is the conditional expectation of the terminal position in \cite[(3.19)]{ElAlaouiSellke2022}, and (b) extends the positivity $\psi(q,X_q)\ge0$ at levels $q\ge Q$ of \cite[Lemma~10]{ElAlaouiSellke2022} to all levels. The next inequality compares $b\psi$ with $v$. It is the main pointwise tool for the rescaled analysis of later chapters: it controls $v$, $b$, and $\psi$ far from $\kappa$ in terms of the single parameter $1-q$. \begin{lemma}\label{II:lem:sandwich} Let $\gamma\in\cU$ and $\kappa\in\R$, and write $\tau=1-q$. \begin{enumerate}[label=(\alph*)] \item For $(q,x)\in[0,1)\times\R$, \begin{equation}\label{II:fun:sandwich} v\le b\,\psi\le\frac\tau\lambda\,v . \end{equation} The upper inequality is equivalent to $F:=\tau c-\lambda b^2-(x-\kappa)b=\tau c-b\psi\ge0$. More precisely, for $q\le Q$, \[ \begin{gathered} b\psi-v=\E\int_q^Q\big[(1-\gamma)\,c\,(1-v)\big](t,X^{q,x}_t)\dd t,\\ F(q,x)=\E\int_q^Q\big[(\lambda-\gamma\tau)\,c^2\big](t,X^{q,x}_t)\dd t , \end{gathered} \] and both $b\psi-v$ and $F$ vanish on $[Q,1)$. The function $F$ satisfies $\cL F=(\gamma\tau-\lambda)c^2\le0$. \item For $q\in[0,1)$: $0<\psi(q,x)\le\tau/(\kappa-x)$ for $x<\kappa$; $0<\lambda b(q,x)\le\tau/(x-\kappa)$ for $x>\kappa$; $\psi(q,\kappa)=\lambda b(q,\kappa)\le\sqrt\tau$, and $\psi(q,\kappa)\ge2\phi_0\sqrt\delta$ if $q\le Q$; \[ \big|v(q,x)-\1_{\{x<\kappa\}}\big|\le\min\Big\{1,\frac{4\tau}{(x-\kappa)^2}\Big\}; \] and \[ \int_{-\infty}^\kappa(1-v)(q,x)\dd x=\int_\kappa^\infty v(q,x)\dd x=\psi(q,\kappa),\qquad \int_\R\big(v(q,x)-\1_{\{x<\kappa\}}\big)\dd x=0 . \] \end{enumerate} \end{lemma} \begin{proof} (a) The product rule \eqref{II:fun:Lrules} and \Cref{II:lem:generator}(a) give $\cL(b\psi)=b_x\psi_x=-c(1-v)$, hence \[ \cL(v-b\psi)=-\frac\gamma\lambda v(1-v)+c(1-v)=(1-\gamma)\,c\,(1-v), \] because $\frac\gamma\lambda v=\gamma c$. Similarly, $\tau'=-1$ and $\lambda'=-\gamma$ give \[ \cL(\tau c)=-c+\gamma\tau c^2,\qquad \cL(\lambda b^2)=-\gamma b^2+\lambda c^2,\qquad \cL\big((x-\kappa)b\big)=\gamma b^2-c, \] so $\cL F=(\gamma\tau-\lambda)c^2$, which is nonpositive because $\lambda(q)=\int_q^1\gamma\ge\gamma(q)\tau$. On $[Q,1)$, \eqref{II:fun:topderiv} and $\lambda=\tau$ give $b\psi=R\Rbar=V=v$ and $F=V-R\Rbar=0$. The functions $v-b\psi$ and $F$ belong to $\mathscr D$, so \Cref{II:fun:diffusion}(c) on $[q,Q]$, started at $(q,x)$, gives the two representations. Their integrands are nonnegative, since $\gamma\le1$, $c>0$, $v<1$, and $\lambda\ge\gamma\tau$. This proves \eqref{II:fun:sandwich}, the upper bound being $F\ge0$, because $\tau c=\frac\tau\lambda v$. (b) Put $f=\lambda b$, so that $f_x=-v<0$, $f>0$, $\psi=f+x-\kappa$ and, by \eqref{II:fun:sandwich}, $f\psi=\lambda b\psi\le\tau v<\tau$. For $x<\kappa$, $f=\psi+\kappa-x>\kappa-x$ since $\psi>0$, so $\psi\le\tau/f\le\tau/(\kappa-x)$. For $x>\kappa$, $\psi>x-\kappa$, so $f\le\tau/(x-\kappa)$. At $x=\kappa$, $\psi=f$ and $\psi^2\le\tau$. The lower bound $\psi(q,\kappa)\ge\sqrt\delta\,\Rbar(0)=2\phi_0\sqrt\delta$ for $q\le Q$ is \eqref{II:fun:psirep}. For $x<\kappa$ let $x'=(x+\kappa)/2$. By convexity of $\psi$ and $\psi>0$, \[ 1-v(q,x)=\psi_x(q,x)\le\frac{\psi(q,x')}{x'-x}\le\frac{2\tau}{\kappa-x}\cdot\frac2{\kappa-x}. \] For $x>\kappa$ let $x'=(x+\kappa)/2$. Since $v$ is decreasing in $x$, \[ v(q,x)\,(x-x')\le\int_{x'}^xv=f(x')-f(x)\le f(x')\le\frac{2\tau}{x-\kappa}. \] Together with $00$, $\cH$ is strictly decreasing near $0$. \begin{proposition}\label{II:prop:kkt} Let $\alpha>0$ and $\kappa\in\R$. \begin{enumerate}[label=(\alph*)] \item \emph{First variation.} Let $r\in[0,1)$ and $\gamma,\widetilde\gamma\in\cU_r$. Then \begin{equation}\label{II:fun:variation} \frac{\dd}{\dd\vartheta}\cP_\kappa\big(\gamma+\vartheta(\widetilde\gamma-\gamma)\big)\Big|_{\vartheta=0+}=\frac12\int_0^r(\widetilde\gamma-\gamma)(q)\,D(q)\dd q=\int_{[0,r]}\cH\dd\mu_{\widetilde\gamma}-\int_{[0,r]}\cH\dd\mu_\gamma , \end{equation} where $D$ and $\cH$ are those of $\gamma$. \item \emph{Minimizers at a fixed cutoff.} Let $r\in(0,1)$ and let $\gamma$ minimize $\cP_\kappa$ over $\cU_r$. Then $\supp\mu_\gamma\subset\argmin_{[0,r]}\cH$ and $0\notin\supp\mu_\gamma$. Every $t\in\supp\mu_\gamma$ with $t0$ on $[0,q_0)$. \item \emph{The scalar $\chi$.} If $\gamma$ is a KKT point (in particular, a global minimizer), then $\chi(q)=\chi(0)+\gamma(q)D(q)$ for every $q\in[0,1)$. In particular $\chi$ is constant on $[0,q_0]\cup\supp\mu_\gamma$. \end{enumerate} \end{proposition} The conditions in (b)--(c) are necessary conditions only. The inequality $S\le1$ at support points uses that each support point minimizes $\cH$; it does not follow from $D=0$ on the support alone. For a global minimizer the conditions are obtained in $\cU_r$ with $Q0$, the point $0$ is not a minimum point, so $0\notin\supp\mu_\gamma$. A support point $t\in(0,r)$ is an interior minimum point of the $C^2$ function $\cH$, so $\cH'(t)=0$ and $\cH''(t)\ge0$, that is, $D(t)=0$ and $S(t)\le1$. Since $Q\in\supp\mu_\gamma$, this applies to $Q$ if $Q0$ and $\cH'(0)<0$, the point $0$ does not minimize $\cH$ on $[0,r]$, so $0\notin\supp\mu_\gamma$ and $q_0>0$. Every support point $t$ lies in $(0,Q]\subset(0,r)$ and is therefore an interior minimum point of $\cH$ on $[0,r]$; hence $D(t)=0$ and $S(t)\le1$, as in (b). For $q0$ for $q\in[0,q_0)$. A continuous function with positive right derivative is strictly increasing, so $S$ is strictly increasing on $[0,q_0]$. Hence $SD(q_0)=0$ on $[0,q_0)$. (d) By \Cref{II:lem:generator}(c)(iii), $\chi(q)-\chi(0)=\int_0^q\gamma D'\dd t$. The integration by parts formula for the right-continuous function $\gamma$ of bounded variation and the $C^1$ function $D$ reads \[ \gamma(q)D(q)-\gamma(0)D(0)=\int_{(0,q]}D\dd\mu_\gamma+\int_0^q\gamma(t)D'(t)\dd t . \] Here $\gamma(0)=\mu_\gamma(\{0\})=0$ because $0\notin\supp\mu_\gamma$, and the Stieltjes integral vanishes because $D=0$ on $\supp\mu_\gamma$ by (c). Hence $\chi(q)-\chi(0)=\gamma(q)D(q)$. The right side vanishes on $[0,q_0)$, where $\gamma=0$, and on $\supp\mu_\gamma$, where $D=0$. \end{proof} A gap of the support is an interval on which $\gamma$ is constant. On such an interval the obstacle conditions force a critical point of $S$ with a sign condition on $S''$, which by \Cref{II:lem:generator}(d) is a sign condition on $\Sigma$. This is the mechanism by which gaps are excluded in \Cref{II:sec:frsb}. For the SK model, Zhou excluded gaps of the support through the first-order conditions at their endpoints \cite{ZhouFRSB}, and Lopatto added a sign condition inside the gap \cite[Section~9]{LopattoFRSB}; \Cref{II:lem:gap} combines the two. \begin{lemma}\label{II:lem:gap} Let $\gamma$ be a KKT point of $\cP_\kappa$ in the sense of \Cref{II:prop:kkt}(c), with $r\in(Q,1)$ as in the definition; for instance, let $\gamma$ be a global minimizer. \begin{enumerate}[label=(\alph*)] \item Let $p1$. If $D\equiv0$ on $[p,p']$, then $S\equiv1$, $\widetilde x\equiv m$, and $\Sigma(\cdot\,;m/\lambda)\equiv0$ on $[p,p']$. \item If $[c_1,c_2]\subset\supp\mu_\gamma$ with $c_10$. As $p'\in\supp\mu_\gamma$, $\mu_\gamma([p',1])>0$ and $m=1-\mu_\gamma([p',1])<1$. The statements about $D$ and $S$ at $p,p'$ are \Cref{II:prop:kkt}(c). Since $p,p'\in\supp\mu_\gamma\subset\argmin_{[0,r]}\cH$ and $[p,p']\subset[0,Q]\subset[0,r]$, we have $\cH\ge\cH(p)=\cH(p')$ on $[p,p']$. The formula for $\lambda$ follows from $\lambda'=-m$. Smoothness of $S$ is \Cref{II:lem:generator}(d) with $[s,t]=[p,p']$. Put $\mathcal G=\cH-\cH(p)\ge0$ on $[p,p']$, so $\mathcal G(p)=\mathcal G(p')=0$ and $\mathcal G'=-D/2$. If $\mathcal G\equiv0$, then $D\equiv0$ and $D'=(S-1)/\lambda^2\equiv0$, so $S\equiv1$ and $S'\equiv S''\equiv0$ on $[p,p']$; \Cref{II:lem:generator}(d) and $\widetilde B>0$ give $\widetilde x\equiv m$ and $\Sigma(\cdot\,;m/\lambda)\equiv0$. Otherwise $\mathcal G$ has a positive maximum at some $s_M\in(p,p')$. The mean value theorem gives $t_1\in(p,s_M)$ and $t_2\in(s_M,p')$ with $\mathcal G'(t_1)>0>\mathcal G'(t_2)$, that is, $D(t_1)<00$, that is, $S(x_2)>1$. Since $S(p),S(p')\le1$, the maximum of $S$ on $[p,p']$ exceeds $1$ and is attained at an interior point $x^*$, where $S'(x^*)=0$ and $S''(x^*)\le0$. The equivalent form follows from \Cref{II:lem:generator}(d) with $K=m/\lambda$ and $\widetilde B>0$. (b) By \Cref{II:prop:kkt}(c), $D=0$ on $[c_1,c_2]$; hence $D'=0$ and $S=1$ there. For $q\in[c_1,c_2)$ the right derivative of $S$ vanishes, and \Cref{II:lem:generator}(c)(iv) gives $\cN(q)=2\frac{\gamma(q)}{\lambda(q)}\widetilde B(q)$, that is, $\gamma(q)=\widetilde x(q)$. The left derivative at $q\in(c_1,c_2]$ gives $\gamma(q-)=\widetilde x(q)$ in the same way. Since $\widetilde x$ is continuous, so is $\gamma$ on $(c_1,c_2)$. \end{proof} \section{The value identity, the slope in the margin, and weak top stability}\label{II:fun:sec:value} The rest of this chapter studies the variational problem $\cP_*(\kappa)=\inf_{\gamma\in\cU}\cP_\kappa(\gamma)$. \Cref{II:fun:sec:value,II:fun:sec:thresholds} contain the value identity (\Cref{II:lem:value}), the universal slope bound (\Cref{II:lem:slope}), weak top stability (\Cref{II:lem:weaktop}), the existence of minimizers (\Cref{II:fun:existence}), the finiteness of the critical margin (\Cref{II:fun:finite}) and the divergence of $\cP_*$ at the critical margin (\Cref{II:fun:critical}). They appear with the same statements in Part~I, \Cref{I:fun:value,I:fun:slope,I:fun:top,I:fun:existence,I:fun:finite,I:fun:critical}, and we refer to Part~I, \Cref{I:sec:hardwall}, for their proofs. \Cref{II:fun:sec:J} introduces the zero-temperature functional $J^k$, compares $\cP_\kappa$ with it through an order-preserving rescaling of the Cole--Hopf recursion (\Cref{II:fun:compare}), identifies $\kc$ with the threshold of $J^k$ (\Cref{II:thm:kJ}), and records the slack of $J^k$ below $\kc$ (\Cref{II:fun:slack}), which enters the top-gap bound of \Cref{II:sec:apriori}. \Cref{II:fun:sec:rates} proves that $\cP_*(\kc-\eta)$ is of logarithmic order in $\eta$ for $\alpha>2$ (\Cref{II:fun:rates}); no later chapter uses it, and it enters only the consequences for the random model stated in \Cref{II:sec:intro-results}. \paragraph{Conventions.} Throughout the rest of this chapter $\alpha>0$ is fixed. For $\kappa\in\R$ and $\gamma\in\cU$ we use the notation of \Cref{II:def:orderparam}: $Q=Q_\gamma$, $\delta=1-Q$, $\lambda=\lambda_\gamma$, $u=u_\gamma(\cdot,\cdot\,;\kappa)$, $b=\partial_xu$, $c=-\partial_{xx}u$, and $v=\lambda c$. The diffusion $X$ of \eqref{II:fun:X} is driven by the Brownian motion $B$, and $D$ and $S$ are the functions of \eqref{II:fun:scalars}. When the order parameter or the margin varies, we write $u_\gamma(q,x;\kappa)$ as in \Cref{II:def:orderparam}, and $b_{\gamma,\kappa}=\partial_xu_\gamma(\cdot,\cdot\,;\kappa)$, $c_{\gamma,\kappa}$, and so on. The \emph{normalized top position} is \[ Z=\frac{X_Q-\kappa}{\sqrt\delta},\qquad\text{so that}\qquad u(Q,X_Q)=\log\Phi(Z),\qquad b(Q,X_Q)=\frac{R(Z)}{\sqrt\delta} \] by \eqref{II:fun:top} and \eqref{II:fun:topderiv}, and $\sqrt\delta\,b(0,0)=\E R(Z)$ by \Cref{II:lem:generator}(b). Here and below, $G$ is a standard Gaussian variable independent of $B$. Besides \Cref{II:fun:gauss}(c) we use the following shifted form of the Gaussian tail bound: for every $t\in\R$ and $\theta>0$, with $\log_+y=\max\{\log y,0\}$, \begin{equation}\label{II:fun:tailshift} \log\Phi(-t)\ge-\frac12(t+\theta)_+^2-\log_+\frac1\theta-2 . \end{equation} Indeed, if $t\le0$ the left side is at least $\log\frac12>-2$. If $t>0$, then $\frac12(t+\theta)^2\ge\frac12t^2+\theta t$, and by \Cref{II:fun:gauss}(c) it suffices to show $\log(1+t)\le\theta t+\log_+(1/\theta)$. For $\theta\ge1$ this follows from $\log(1+t)\le t$; for $\theta<1$ the concave function $t\mapsto\log(1+t)-\theta t$ is maximal at $t=1/\theta-1$, where it equals $\log(1/\theta)-1+\theta<\log(1/\theta)$. Part~I, \eqref{I:hw:tailshift}, states the same bound with $-\log\Phi(-1)<2$ in place of $2$. The next identity expresses $\cP_\kappa(\gamma)$ through the terminal law of $X$ and the function $D$. For a minimizer the $D$-terms vanish or have a sign, and the identity then bounds the top gap $\delta$ from below by the value. \begin{lemma}\label{II:lem:value} For every $\kappa\in\R$ and $\gamma\in\cU$, \begin{equation}\label{II:fun:valueid} \cP_\kappa(\gamma)=\frac12\log\delta+\alpha\E h(Z)-\frac\delta2D(Q) -\frac12\int_0^Q\gamma(q)D(q)\dd q . \end{equation} If $r\in[0,1)$ and $\gamma$ minimizes $\cP_\kappa$ over $\cU_r$, then $\int_0^Q\gamma D\dd q=0$ and $D(Q)\ge0$, with $D(Q)=0$ if $Q1$ and \[ S(Q)\ge s_\alpha:=\min\Big\{\frac12,\frac1{16\alpha}\Big\}. \] Moreover, every $\gamma\in\cU$ satisfies, at every margin $\kappa$, \begin{equation}\label{II:fun:topslope} \sqrt\delta\,b(0,0)\ge\frac{S(Q)}{2\alpha}. \end{equation} Consequently, every minimizer $\gamma$ of $\cP_\kappa$ over $\cU_r$ satisfies \begin{equation}\label{II:fun:lambdatop} \lambda(0)\le\max\Big\{1,\frac{2\alpha C_\kappa}{s_\alpha}\Big\}\sqrt\delta . \end{equation} \end{lemma} \begin{proof} This is Part~I, \Cref{I:fun:top}. \end{proof} \section{The critical margin and existence of minimizers}\label{II:fun:sec:thresholds} By \Cref{II:lem:slope}, $\cP_*(\cdot\,;\alpha)$ is nonincreasing. By \eqref{II:fun:P} and $u\le0$, $\cP_\kappa(\gamma;\alpha)$ and $\cP_*(\kappa;\alpha)$ are nonincreasing in $\alpha$. The \emph{critical margin} is \begin{equation}\label{II:fun:kc} \kc(\alpha)=\sup\{\kappa\in\R:\cP_*(\kappa;\alpha)>-\infty\}. \end{equation} Thus $\cP_*>-\infty$ on $(-\infty,\kc)$ and $\cP_*=-\infty$ on $(\kc,\infty)$, and $\alpha\mapsto\kc(\alpha)$ is nonincreasing. For $r\in[0,1)$ we write \begin{equation}\label{II:fun:constrained} \cP_*^r(\kappa)=\min_{\gamma\in\cU_r}\cP_\kappa(\gamma) \end{equation} for the constrained minimum, which is attained because $\cU_r$ is compact and $\cP_\kappa$ is continuous on it (\Cref{II:lem:regularity}(f)). Since the classes $\cU_r$ increase with $r$ and their union is $\cU$, $\cP_*^r(\kappa)$ decreases to $\cP_*(\kappa)$ as $r\uparrow1$. Also $\cP_*^0(\kappa)=\alpha\log\Phi(-\kappa)<0$, since $\cU_0=\{\1_{[0,1]}\}$; in particular $\cP_*(\kappa)<0$. \begin{theorem}\label{II:fun:existence} Let $\kappa\in\R$ with $\cP_*(\kappa)>-\infty$. Then the infimum defining $\cP_*(\kappa)$ is attained. Every minimizer $\gamma$ satisfies \[ 1-Q_\gamma>\e^{2\cP_*(\kappa)}, \] that is, every minimizer lies in $\cU_r$ with $r=1-\e^{2\cP_*(\kappa)}\in(0,1)$, and the set of minimizers is compact in $L^1([0,1])$. \end{theorem} \begin{proof} This is Part~I, \Cref{I:fun:existence}. \end{proof} The theorem concerns minimizers. It does not assert that the tops of an arbitrary minimizing sequence stay away from one: adding a vanishing mass near one changes the top without changing the $L^1$ limit. For the replica-symmetric order parameter $\1_{[q,1]}$, $0\le q<1$, the order parameter vanishes on $[0,q)$, so $X=B$ there and $\lambda=1-q$ on $[0,q]$. Hence \begin{equation}\label{II:fun:PRS} \PRS_\kappa(q):=\cP_\kappa(\1_{[q,1]})=\alpha\E\log\Phi\Big(\frac{\sqrt q\,G-\kappa}{\sqrt{1-q}}\Big) +\frac q{2(1-q)}+\frac12\log(1-q). \end{equation} \begin{proposition}\label{II:fun:finite} For every $\alpha>0$, \[ -\infty<\kc(\alpha)\le\kRS(\alpha)<\infty, \] where $\kRS(\alpha)$ is the unique solution of $\alpha\,m(\kRS)=1$. In particular $\kc(\alpha)<0$ if $\alpha>2$. \end{proposition} \begin{proof} This is Part~I, \Cref{I:fun:finite}. \end{proof} The bound $\kc\le\kRS$ comes from the limit $q\uparrow1$ in \eqref{II:fun:PRS}: if $\alpha m(\kappa)>1$, then $\PRS_\kappa(q)\to-\infty$ as $q\uparrow1$. This is the limit in which Gardner computed the replica-symmetric capacity \cite{Gardner1988}. For this functional, El Alaoui and Sellke made the same observation in the remark after \cite[Conjecture~1]{ElAlaouiSellke2022}. \begin{theorem}\label{II:fun:critical} For every $\alpha>0$ the function $\cP_*$ is finite, nonincreasing, concave, and continuous on $(-\infty,\kc)$, and $\cP_*(\kappa)=-\infty$ for $\kappa\ge\kc$. More precisely, with $C_*=C_{\kc}$, for all sufficiently small $\eta>0$, \begin{equation}\label{II:fun:critbound} \cP_*(\kc-\eta)\le\frac12\log\eta+\frac12\log(2\alpha C_*)+\frac12 . \end{equation} \end{theorem} \begin{proof} This is Part~I, \Cref{I:fun:critical}. \end{proof} \section{The zero-temperature functional}\label{II:fun:sec:J} The critical margin also has a zero-temperature characterization, which is used to certify bounds on $\kc$ in \Cref{II:sec:rs}. It involves the Parisi equation with a quadratic terminal penalty in place of $\log\Phi$. Zero-temperature Parisi functionals were introduced for the ground state energy of mixed $p$-spin models by Auffinger and W.-K.\ Chen \cite{AuffingerChen2017GroundState}, and for spherical models by W.-K.\ Chen and Sen \cite{ChenSen2017Spherical} and by Jagannath and Tobasco \cite{JagannathTobasco2017Spherical}; see also \cite{JagannathSen2021}. For the negative perceptron, the quadratic penalty is the boundary condition of the jamming limit in the replica analysis of Franz and Parisi \cite[(15)]{FranzParisi2016}; see also \cite{FPSUZ2017}. For $k\in\R$ let \[ f_k(x)=-\frac12(k-x)_+^2 , \] which belongs to the class $\mathfrak F$ of \Cref{II:fun:control}, with $f_k'(x)=(k-x)_+$. For $T>0$ let $\mathfrak X_T$ be the set of bounded, nondecreasing, right-continuous functions $\xi:[0,T)\to[0,\infty)$, and put $\mathfrak X=\mathfrak X_1$; the values of $\xi$ may exceed one. For a step function $\xi\in\mathfrak X_T$ let $\Upsilon=\Upsilon^k_{T,\xi}=\Upsilon_{\xi,f_k}$ be the solution of the recursion of \Cref{II:fun:control}, that is, of \begin{equation}\label{II:fun:quadpde} \Upsilon_q=-\frac12\big(\Upsilon_{xx}+\xi\,\Upsilon_x^2\big),\qquad\Upsilon(T,x)=f_k(x), \end{equation} and for general $\xi\in\mathfrak X_T$ let $\Upsilon^k_{T,\xi}$ be the $L^1$-limit of step approximations (\Cref{II:fun:quad}(a)). Define \begin{equation}\label{II:fun:J} J^k_T(\xi)=\alpha\,\Upsilon^k_{T,\xi}(0,0)+\frac12\int_0^T\frac{\dd q}{\Lambda_\xi(q)},\qquad \Lambda_\xi(q)=1+\int_q^T\xi(t)\dd t,\qquad J^k=J^k_1 , \end{equation} and the thresholds \begin{align*} \kJ(\alpha)&=\inf\{k:\ J^k(\xi)<0\ \text{for some }\xi\in\mathfrak X\},\\ \kappa^0_{\mathrm c}(\alpha)&=\inf\{k:\ J^k(\xi)<0\ \text{for some step function }\xi\in\mathfrak X\}. \end{align*} If $k<\kJ$, then $J^k(\xi)\ge0$ for every $\xi\in\mathfrak X$. Two elementary properties of the Cole--Hopf step $\cT_{m,s}$ of \eqref{II:fun:recursion} drive this section. Let $m\ge0$, $s>0$, and let $g$ be bounded above with $g(x)\ge-C(1+x^2)$. For $K>0$ and $T>0$, \begin{equation}\label{II:fun:onestep} K\,\cT_{m,s}g=\cT_{m/K,s}(Kg),\qquad \frac1T(\cT_{m,s}g)(\sqrt T\,y)=\cT_{Tm,s/T}\,g^\sharp(y),\quad g^\sharp(y)=\frac1Tg(\sqrt T\,y). \end{equation} The first identity is \eqref{II:fun:Tprops}. For the second, if $m>0$ the right side equals $(Tm)^{-1}\log\E\exp\{m\,g(\sqrt T\,y+\sqrt s\,G)\}$, which is the left side, and if $m=0$ both sides equal $T^{-1}\E g(\sqrt T\,y+\sqrt s\,G)$. By the first identity, multiplying the terminal datum by $K$ and dividing all coefficients by $K$ multiplies the solution of the recursion by $K$. By the second, rescaling time by $T$ and space by $\sqrt T$ maps a recursion on $[0,T]$ to a recursion on $[0,1]$. Together with the monotonicity in \eqref{II:fun:Tprops}, these identities are all that the comparison of $\cP_\kappa$ with $J^k$ requires. \begin{lemma}\label{II:fun:quad} Let $T>0$, $M>0$, and $K_0\ge0$. \begin{enumerate}[label=\textup{(\alph*)}] \item For step functions $\xi,\widetilde\xi\in\mathfrak X_T$ with $\xi,\widetilde\xi\le M$ and for $\abs k\le K_0$, \[ \begin{gathered} \abs{\Upsilon^k_{T,\widetilde\xi}(q,x)-\Upsilon^k_{T,\xi}(q,x)}\le C(1+x^2)\norm{\widetilde\xi-\xi}_{L^1(0,T)},\\ \abs{J^k_T(\widetilde\xi)-J^k_T(\xi)}\le C(1+\alpha)\norm{\widetilde\xi-\xi}_{L^1(0,T)}, \end{gathered} \] with $C=C(T,M,K_0)$. Hence $\Upsilon^k_{T,\xi}$ and $J^k_T(\xi)$ extend uniquely to all $\xi\in\mathfrak X_T$, continuously in $L^1$ on $\{\xi\le M\}$. \item For $\xi\in\mathfrak X$ and $k0$, $\rho>0$, and $x\in\R$, \begin{equation}\label{II:fun:termcmp} f_{\kappa+\rho}(x)-\delta\Big(2+\log_+\frac{\sqrt\delta}\rho\Big)\le\delta\log\Phi\Big(\frac{x-\kappa}{\sqrt\delta}\Big) \le f_\kappa(x), \end{equation} because $\delta t_+^2=(\kappa-x)_+^2$ and $\delta(t+\theta)_+^2=(\kappa+\rho-x)_+^2$. Multiplying the hard-wall datum by $\delta$ therefore turns it into the quadratic penalty, up to a shift of the margin and an additive error of order $\delta$. \begin{lemma}\label{II:fun:compare} Let $\kappa\in\R$ and $\gamma\in\cU$ with $Q>0$, and put $\xi_\gamma=\gamma/\delta$ on $[0,Q)$. Then \begin{equation}\label{II:fun:cmpupper} \cP_\kappa(\gamma)\le\frac{J^\kappa_Q(\xi_\gamma)}\delta-\frac12\log\frac1\delta , \end{equation} and, for every $\rho>0$, \begin{equation}\label{II:fun:cmprho} \cP_\kappa(\gamma)\ge\frac{J^{\kappa+\rho}_Q(\xi_\gamma)}\delta+\frac12\log\delta -\alpha\Big(2+\log_+\frac{\sqrt\delta}\rho\Big). \end{equation} In particular, \begin{equation}\label{II:fun:cmpshift} \cP_\kappa(\gamma)\ge\frac{J^{\kappa+\rho}_Q(\xi_\gamma)}\delta+\frac12\log\delta-2\alpha\qquad \text{for every }\rho\ge\sqrt\delta . \end{equation} Moreover $J^k_Q(\xi_\gamma)=Q\,J^{k/\sqrt Q}(\widetilde\xi_\gamma)$ for every $k\in\R$, where $\widetilde\xi_\gamma(t)=Q\gamma(Qt)/\delta$ for $t\in[0,1)$. \end{lemma} \begin{proof} The function $\xi_\gamma$ belongs to $\mathfrak X_Q$ and is bounded by $1/\delta$. Since $\Lambda_{\xi_\gamma}(q)=1+\delta^{-1}\int_q^Q\gamma=\lambda(q)/\delta$, the entropy terms satisfy exactly \begin{equation}\label{II:fun:entid} \frac12\int_0^Q\frac{\dd q}{\lambda(q)}=\frac1\delta\cdot\frac12\int_0^Q\frac{\dd q}{\Lambda_{\xi_\gamma}(q)} . \end{equation} It remains to prove \begin{equation}\label{II:fun:energycmp} \Upsilon^{\kappa+\rho}_{Q,\xi_\gamma}(0,0)-\delta\Big(2+\log_+\frac{\sqrt\delta}\rho\Big)\le\delta\,u(0,0) \le\Upsilon^\kappa_{Q,\xi_\gamma}(0,0); \end{equation} multiplying by $\alpha/\delta$ and adding \eqref{II:fun:entid} and $\frac12\log\delta$ gives \eqref{II:fun:cmpupper} and \eqref{II:fun:cmprho}, and \eqref{II:fun:cmpshift} is the case $\sqrt\delta/\rho\le1$ of \eqref{II:fun:cmprho}. Suppose first that $\gamma$ is a step order parameter. Then $\gamma\in\cU_Q$, and $u$ is obtained on $[0,Q]$ from the datum $f_Q(x)=\log\Phi((x-\kappa)/\sqrt\delta)$ at time $Q$ by the recursion \eqref{II:fun:steprec} with the blocks of $\gamma$. By the first identity in \eqref{II:fun:onestep} with $K=\delta$, applied to the steps from the last to the first, $\delta u(0,\cdot)$ is obtained from the datum $\delta f_Q$ at time $Q$ by the recursion with the coefficients $\gamma/\delta=\xi_\gamma$. The recursion with these coefficients is nondecreasing in its datum and commutes with the addition of constants \eqref{II:fun:Tprops}, and $\Upsilon^{k}_{Q,\xi_\gamma}$ is the same recursion started from $f_k$. Hence \eqref{II:fun:termcmp} gives \eqref{II:fun:energycmp}. For general $\gamma$, choose partitions of $[0,Q]$ containing $Q$ with mesh tending to zero, and let $\gamma_n(q)=\gamma(t_j)$ on $[t_j,t_{j+1})$ and $\gamma_n=1$ on $[Q,1]$. Then $\gamma_n$ is a step order parameter with $\gamma_n\le\gamma$ and the same top $Q$, because $\gamma<1$ on $[0,Q)$; moreover $\gamma_n\to\gamma$ in $L^1$, and $\xi_{\gamma_n}\to\xi_\gamma$ in $L^1(0,Q)$ with $\xi_{\gamma_n}\le1/\delta$. \Cref{II:lem:regularity}(f) and \Cref{II:fun:quad}(a) pass \eqref{II:fun:cmpupper} and \eqref{II:fun:cmprho} to the limit. The last assertion is \Cref{II:fun:quad}(c) with $T=Q$, since $Q\xi_\gamma(Qt)=\widetilde\xi_\gamma(t)$. \end{proof} We can now state the main result of this section, the zero-temperature characterization of $\kc$. \begin{theorem}\label{II:thm:kJ} For every $\alpha>0$, $\kc(\alpha)=\kJ(\alpha)=\kappa^0_{\mathrm c}(\alpha)$. \end{theorem} The two directions of \Cref{II:thm:kJ} play different roles in \Cref{II:sec:rs}. The inequality $\kc\le\kappa^0_{\mathrm c}$ says that a single negative value of $J^k$ at a step coefficient forces $\kc\le k$; this gives the upper bound on $\kc$ in \Cref{II:thm:cubic}. The inequality $\kJ\le\kc$ is the converse: if $J^k\ge0$ for all coefficients, then $\kc\ge\kJ\ge k$; this turns the dual certificate of \Cref{II:sec:rs} into a lower bound on $\kc$. \begin{proof} We divide the proof into three steps. \proofstep{1}{The inequality $\kc\le\kappa^0_{\mathrm c}$.} Let $\xi\in\mathfrak X$ be a step function with $\xi\le M$ and $J^k(\xi)<0$, and let $k'>k$. For $\delta\in(0,\frac12)$ with $2\delta M<1$ put $Q=1-\delta$ and $\gamma(q)=\delta Q^{-1}\xi(q/Q)$ for $qk$, so $k'/\sqrt Q\ge k$ for small $\delta$, and then $J^{k'/\sqrt Q}(\xi)\le J^k(\xi)<0$ by \Cref{II:fun:quad}(b). Hence $\cP_{k'}(\gamma)\to-\infty$, and $\cP_*(k')=-\infty$ for every $k'>k$. Thus $\kc\le k$, and taking the infimum over $k$ gives $\kc\le\kappa^0_{\mathrm c}$. \proofstep{2}{The inequality $\kJ\le\kc$.} Fix margins $\kcw_r$, by \eqref{II:fun:valueineq}. \item $\delta_r\to0$ as $r\uparrow1$. Indeed, for fixed $r_0<1$ the minimum $\cP_*^{r_0}(k)$ is finite; once $w_r<\cP_*^{r_0}(k)$, the minimizer $\gamma_r$ does not belong to $\cU_{r_0}$, that is, $Q_r>r_0$. \item The slope $b_r$ is large: by \eqref{II:fun:tangent} at the margin $\kc-1$, where $\cP_*$ is finite, $\cP_*(\kc-1)\le\cP_{\kc-1}(\gamma_r)\le w_r+\alpha(k-\kc+1)b_r$. \end{enumerate} By \eqref{II:fun:tangent} at the margin $k''$ and by (iv), with $\vartheta=(k''-k)/(k-\kc+1)>0$, \[ \cP_{k''}(\gamma_r)\le w_r-\alpha(k''-k)b_r\le w_r-\vartheta\big(\cP_*(\kc-1)-w_r\big). \] For $r$ close to one, $Q_r>0$ by (iii), and \eqref{II:fun:cmpshift} at the margin $k''$ with $\rho=\sqrt{\delta_r}$, followed by (ii), gives \[ \frac{J^{k''+\sqrt{\delta_r}}_{Q_r}(\xi_{\gamma_r})}{\delta_r}\le\cP_{k''}(\gamma_r)-\frac12\log\delta_r+2\alpha <-\vartheta\big(\cP_*(\kc-1)-w_r\big)+2\alpha . \] By (i) the right side tends to $-\infty$ as $r\uparrow1$. For $r$ close to one it is negative, and the last assertion of \Cref{II:fun:compare} gives $J^{(k''+\sqrt{\delta_r})/\sqrt{Q_r}}(\widetilde\xi_{\gamma_r})<0$ with $\widetilde\xi_{\gamma_r}\in\mathfrak X$. Hence $\kJ\le(k''+\sqrt{\delta_r})/\sqrt{Q_r}$, and by (iii) this margin tends to $k''$. Thus $\kJ\le k''$ for every $k''>\kc$, that is, $\kJ\le\kc$. \proofstep{3}{Conclusion.} By Steps 1 and 2 and \Cref{II:fun:quad}(d), $\kc\le\kappa^0_{\mathrm c}=\kJ\le\kc$. This proves the theorem. \end{proof} Since $J^{\kc}(\xi)\ge0$ for every $\xi\in\mathfrak X$, \Cref{II:fun:quad}(b) yields a quantitative lower bound on $J$ below the critical margin. Combined with \eqref{II:fun:cmpshift} it bounds $\cP_\kappa(\gamma)$ from below in terms of $\lambda(0)$ and $\delta$; this is the form used for the top-gap lower bound in \Cref{II:sec:apriori}. \begin{corollary}\label{II:fun:slack} For $k<\kc(\alpha)$ and every $\xi\in\mathfrak X$, \[ J^k(\xi)\ge\frac{A(k)}{\Lambda_\xi(0)},\qquad A(k)=\alpha\int_k^{\kc}m_1(t)\dd t>0 . \] Consequently, for $\gamma\in\cU$ with $Q>0$, $\rho\ge\sqrt\delta$, and $(\kappa+\rho)/\sqrt Q\le k<\kc$, \[ \cP_\kappa(\gamma)\ge\frac{Q\,A(k)}{\lambda(0)}+\frac12\log\delta-2\alpha . \] \end{corollary} \begin{proof} If $J^{\kc}(\xi)<0$ for some $\xi$, the continuity in \Cref{II:fun:quad}(b) gives $J^{k'}(\xi)<0$ for some $k'<\kc=\kJ$, contradicting the definition of $\kJ$. Hence $J^{\kc}(\xi)\ge0$, and the lower bound in \Cref{II:fun:quad}(b) with $k'=\kc$ gives the first claim. For the second, \eqref{II:fun:cmpshift}, the last assertion of \Cref{II:fun:compare}, the monotonicity in $k$, and $\Lambda_{\widetilde\xi_\gamma}(0)=\Lambda_{\xi_\gamma}(0)=\lambda(0)/\delta$ give $J^{\kappa+\rho}_Q(\xi_\gamma)=Q\,J^{(\kappa+\rho)/\sqrt Q}(\widetilde\xi_\gamma)\ge Q\,J^k(\widetilde\xi_\gamma)\ge Q\,A(k)\,\delta/\lambda(0)$. \end{proof} \section{Logarithmic divergence at the critical margin}\label{II:fun:sec:rates} By \Cref{II:fun:critical}, $\cP_*(\kc-\eta)\le\frac12\log\eta+O(1)$. For $\alpha>2$ the comparison with $J^k$ gives a lower bound of the same logarithmic order. The reason is that $\kc<0$ for $\alpha>2$: dividing a margin below $\kc$ by $\sqrt Q\le1$, as in the last assertion of \Cref{II:fun:compare}, moves it further below $\kc=\kJ$, where $J\ge0$. The first lemma records the resulting lower bound, which holds for every order parameter. \begin{lemma}\label{II:fun:below} Let $\alpha>2$. Let $\gamma\in\cU$, $\kappa\in\R$, and $\rho>0$ with $\kappa+\rho<\kc$. Then \[ \cP_\kappa(\gamma)\ge\frac12\log\delta-\alpha\Big(2+\log_+\frac{\sqrt\delta}\rho\Big). \] \end{lemma} \begin{proof} By \Cref{II:fun:finite}, $\kappa+\rho<\kc<0$. If $Q=0$, then $\gamma=\1_{[0,1]}$, $\delta=1$, and $\cP_\kappa(\gamma)=\alpha\log\Phi(-\kappa)\ge\alpha\log\frac12>-2\alpha$ because $-\kappa>\rho>0$. If $Q>0$, then $0<\sqrt Q\le1$ gives $(\kappa+\rho)/\sqrt Q\le\kappa+\rho<\kc=\kJ$ (\Cref{II:thm:kJ}), so $J^{(\kappa+\rho)/\sqrt Q}(\widetilde\xi_\gamma)\ge0$. By the last assertion of \Cref{II:fun:compare}, $J^{\kappa+\rho}_Q(\xi_\gamma)\ge0$, and \eqref{II:fun:cmprho} gives the claim. \end{proof} \begin{proposition}\label{II:fun:rates} Let $\alpha>2$. There are constants $C,\eta_0>0$, depending only on $\alpha$, such that for $0<\eta\le\eta_0$ \begin{equation}\label{II:fun:valuerates} \alpha\log\eta-C\le\cP_*(\kc-\eta)\le\frac12\log\eta+C , \end{equation} and every $\gamma\in\cU$ with $\cP_{\kc-\eta}(\gamma)\le\cP_*(\kc-\eta)+1$ satisfies $1-Q_\gamma\le C\eta^{1/(1+\alpha)}$. \end{proposition} \begin{proof} We divide the proof into four steps. First we bound $\cP_\kappa(\gamma)$ below in terms of the top of $\gamma$, for every order parameter. This bounds the tops of near-minimizers. Then we bound the mass and the slope of minimizers, and finally we bound the value below by distinguishing two cases for the top gap. Put $C_*=C_{\kc}$; since $C_\kappa$ is increasing in $\kappa$, $C_\kappa\le C_*$ for $\kappa\le\kc$. By \Cref{II:fun:critical} there is $\eta_1\in(0,1]$ such that $\cP_*(\kc-\eta)\le\frac12\log\eta+C_1$ for $0<\eta\le\eta_1$, with $C_1=\frac12\log(2\alpha C_*)+\frac12$. This is the upper bound in \eqref{II:fun:valuerates}. \proofstep{1}{A lower bound in terms of the top.} We show that for all $\gamma\in\cU$ and $\kappa\le\kc$, \begin{equation}\label{II:fun:topbound} \cP_\kappa(\gamma)\ge\frac{1+\alpha}2\log\delta-C_2,\qquad C_2=\alpha(2+\log2+C_*). \end{equation} \Cref{II:fun:below} at the margin $\kc-\delta$ with $\rho=\delta/2$ gives, since $\sqrt\delta/\rho=2/\sqrt\delta\ge1$, \[ \cP_{\kc-\delta}(\gamma)\ge\frac12\log\delta-2\alpha-\alpha\log\frac2{\sqrt\delta} =\frac{1+\alpha}2\log\delta-\alpha(2+\log2). \] If $\kappa\le\kc-\delta$, then $\cP_\kappa(\gamma)\ge\cP_{\kc-\delta}(\gamma)$ because $\cP_\cdot(\gamma)$ is decreasing. If $\kc-\delta<\kappa\le\kc$, then $\cP_\kappa(\gamma)\ge\cP_{\kc}(\gamma)$, and on $[\kc-\delta,\kc]$ the derivative of $\cP_\cdot(\gamma)$ is at least $-\alpha C_*/\lambda(0)\ge-\alpha C_*/\delta$ by \Cref{II:lem:slope}; integrating over this interval of length $\delta$ gives $\cP_{\kc}(\gamma)\ge\cP_{\kc-\delta}(\gamma)-\alpha C_*$. This proves \eqref{II:fun:topbound}. \proofstep{2}{Tops of near-minimizers.} Let $0<\eta\le\eta_1$, put $k=\kc-\eta$, and let $\gamma\in\cU$ satisfy $\cP_k(\gamma)\le\cP_*(k)+1$. By \eqref{II:fun:topbound} and the upper bound in \eqref{II:fun:valuerates}, $\frac{1+\alpha}2\log\delta-C_2\le\frac12\log\eta+C_1+1$, that is, \[ 1-Q_\gamma=\delta\le C_3\eta^{1/(1+\alpha)},\qquad C_3=\e^{2(C_1+C_2+1)/(1+\alpha)} . \] We fix $\eta_0\in(0,\eta_1]$ with $C_3\eta_0^{1/(1+\alpha)}\le\frac12$. From now on $0<\eta\le\eta_0$, $k=\kc-\eta$, and $\gamma$ is a minimizer of $\cP_k$ over $\cU$, which exists by \Cref{II:fun:existence}. We write $\delta,\lambda,b$ for its quantities at the margin $k$. By the bound just proved, $\delta\le\frac12$, that is, $Q\ge\frac12$. \proofstep{3}{The mass and the slope.} The minimizer $\gamma$ minimizes $\cP_k$ over $\cU_r$ for every $r\in(Q,1)$, so \eqref{II:fun:lambdatop} and $C_k\le C_*$ give $\lambda(0)\le C_4\sqrt\delta$ with $C_4=\max\{1,2\alpha C_*/s_\alpha\}$. By \eqref{II:fun:slope}, $\lambda(0)b(0,0)\ge\E(k-\sqrt Q\,G)_+$. Since $k\ge\kc-1$, $\kc-1<0$, and $\frac12\le Q\le1$, we have $(\kc-1)/\sqrt Q\ge\sqrt2(\kc-1)$, and therefore \[ \lambda(0)b(0,0)\ge\sqrt Q\,m_1\Big(\frac{\kc-1}{\sqrt Q}\Big)\ge\frac1{\sqrt2}\,m_1\big(\sqrt2(\kc-1)\big)=:c_4>0 . \] \proofstep{4}{The lower bound in \eqref{II:fun:valuerates}.} Put $s=\sqrt\delta\le1$. We distinguish two cases. \emph{Case A: $s\ge\eta/4$.} \Cref{II:fun:below} at the margin $k$ with $\rho=\eta/2$ gives $\cP_*(k)=\cP_k(\gamma)\ge\log s-2\alpha-\alpha\log_+(2s/\eta)$. If $2s\ge\eta$, the right side equals $\alpha\log\eta+(1-\alpha)\log s-\alpha(2+\log2)\ge\alpha\log\eta-\alpha(2+\log2)$, because $s\le1$ and $\alpha>1$. If $2s<\eta$, it equals $\log s-2\alpha\ge\log\eta-\log4-2\alpha\ge\alpha\log\eta-\log4-2\alpha$, because $\eta\le1$. \emph{Case B: $s<\eta/4$.} Here the margin shift in \Cref{II:fun:below} costs nothing, but $\frac12\log\delta$ may be very negative; the slope compensates. By \eqref{II:fun:tangent} and Step 3, \[ \cP_k(\gamma)\ge\cP_{k+\eta/2}(\gamma)+\frac{\alpha\eta}2\,b(0,0)\ge\cP_{k+\eta/2}(\gamma)+\frac{\alpha c_4\eta}{2\lambda(0)} . \] \Cref{II:fun:below} at the margin $k+\eta/2=\kc-\eta/2$ with $\rho=\eta/4$ gives $\cP_{k+\eta/2}(\gamma)\ge\log s-2\alpha$, because $\sqrt\delta/\rho=4s/\eta<1$. With $\lambda(0)\le C_4s$ we obtain, with $c_5=\alpha c_4/(2C_4)$, \[ \cP_*(k)\ge\log s+\frac{c_5\eta}s-2\alpha\ge\log\eta+1+\log c_5-2\alpha , \] since $s\mapsto\log s+c_5\eta/s$ is minimal at $s=c_5\eta$. As $\log\eta\ge\alpha\log\eta$, both cases give the lower bound in \eqref{II:fun:valuerates} with a constant depending only on $\alpha$. \end{proof} The exponent $1/(1+\alpha)$ for near-minimizers comes from the factor $\frac{1+\alpha}2$ in \eqref{II:fun:topbound}. The proposition does not identify the coefficient of $\log\eta$ in $\cP_*(\kc-\eta)$. For exact minimizers, Part~I, \Cref{I:fun:pin,I:fun:collapse-prop}, prove the bound on $1-Q_\gamma$ for every $\alpha>0$, directly from the value identity. \chapter{Replica-symmetric order parameters, the de Almeida--Thouless line, and the ordering of thresholds}\label{II:sec:rs} This chapter studies the replica-symmetric order parameters $\1_{[q,1]}$ and the three margins $\kRS$, $\kd$, and $\kc$ for $\alpha=2+e$ with $e>0$ small. The margin $\kRS$ is Gardner's prediction, and \Cref{II:fun:finite} gives $\kc\le\kRS$. \Cref{II:rs:sec-krs} expands $\kRS$ near $\alpha=2$, and \Cref{II:rs:sec-rs} computes the functional, the function $D$, and the replicon eigenvalue at a replica-symmetric order parameter. In \Cref{II:rs:sec-curve} an exact formula for the derivative of $D$ along the replica-symmetric family (\Cref{II:rs:lem-unique}) shows that, for $\alpha<\pi^2/4$, every margin below $\kRS$ has exactly one stationary replica-symmetric order parameter, and \Cref{II:prop:crossing} shows that its replicon eigenvalue changes sign at a single margin $\kd$, with $\kRS-\kd=\thd e^2+O(e^3\log(1/e))$. \Cref{II:rs:sec-J} proves a duality bound for the zero-temperature functional $J^k$ of \Cref{II:sec:functional} (\Cref{II:rs:thm-dual}). In \Cref{II:rs:sec-order} we construct an explicit dual certificate and an explicit negative value of $J^k$. This gives $\kRS-C_{\rm ord}e^3\le\kc\le\kRS-c_{\rm cub}e^3$ (\Cref{II:thm:ordering,II:thm:cubic}); on the interval $(\kd,\kc)$ no minimizer is replica symmetric (\Cref{II:cor:nonrs}). In the terms of \Cref{II:sec:intro-ideas}, the certified margin $k_*$ of \Cref{II:thm:ordering} enters Step~1 through the top-gap lower bound of \Cref{II:sec:apriori}; \Cref{II:thm:ordering}(iii) places every compact window $\Theta$ of rescaled margins below $\kc$ (\Cref{II:rs:rem-Theta}); and Steps~4 and~5 use the order parameter $\gamma_{\mathrm d}=\1_{[\qd,1]}$ of \Cref{II:prop:crossing} and the asymptotics of $1-\qd$. \section{Gardner's prediction near \texorpdfstring{$\alpha=2$}{alpha=2}}\label{II:rs:sec-krs} We use the Gaussian functions $R$, $\Rbar$, $V$, $m$, $m_1$, and $\varrho$ of \Cref{II:fun:sec:gauss} and the facts collected in \Cref{II:fun:gauss}. The next lemma expands Gardner's prediction $\kRS$ near $\alpha=2$. Its part (b) identifies the quantity $\alpha\Phi(\kRS)-1\approx e/4$, which drives both the de Almeida--Thouless crossing and the cubic separation below. \begin{lemma}\label{II:lem:krs} There is $e_{\rm g}>0$ such that for $00$, so $\kRS<0$. Since $m(-\frac12)<\frac12$, for small $e$ we have $m(\kRS)=\frac1{2+e}>m(-\frac12)$, that is, $\kRS>-\frac12$. By convexity of $m$ and \Cref{II:fun:gauss}(g), \[ 2m_1(-\tfrac12)|\kRS|\le m'(\kRS)|\kRS|\le m(0)-m(\kRS)=\frac e{2\alpha}\le m'(0)|\kRS|=2\phi_0|\kRS| , \] which gives the two-sided bound, since $\alpha\ge2$. Taylor's formula with $00$, \[ \rsF(k,\beta):=\beta^2\,\E R\Bigl(\frac{G-k}\beta\Bigr)^2,\qquad \rsH(k,\beta):=\E V\Bigl(\frac{G-k}\beta\Bigr)^2 . \] We call $q\in(0,1)$ \emph{RS-stationary} at $\kappa$ if $D_{\gamma_q}(q)=0$. At an RS-stationary $q$ the \emph{replicon eigenvalue} is \[ \Lambda_{\mathrm R}:=\frac{1-\alpha\rsH(k,\beta)}{4(1-q)^2}. \] When $\kappa$ admits exactly one RS-stationary point, we call it the \emph{stationary RS point} and denote it by $q^*(\kappa)$. \end{definition} For $\gamma_q$ the diffusion is a Brownian motion on $[0,q]$, so $Z^q$ is the normalized top position $Z=(X_Q-\kappa)/\sqrt\delta$ of $\gamma_q$. The quantity $\rsH$ is the Gaussian second moment of $V$; it should not be confused with the obstacle potential $\cH$. The next lemma identifies $\Lambda_{\mathrm R}$ with one half of the curvature of $\cH_{\gamma_q}$ at the atom, as in \eqref{II:eq:intro-replicon}. Its sign decides whether moving a small part of the mass of the atom to a nearby point lowers the functional. At an RS-stationary point, $\alpha\rsH(k,\beta)=1$ is the de Almeida--Thouless condition recalled in \Cref{II:sec:intro-problem}, and the stationarity condition $\alpha\rsF(k,\beta)=1$ is \cite[(43)]{FPSUZ2017}. \begin{lemma}\label{II:rs:lem-formulas} Let $\alpha>0$, $\kappa\in\R$, $q\in(0,1)$, $\gamma=\gamma_q$, and $r\in(q,1)$. \begin{enumerate}[label=\textup{(\alph*)}] \item $\displaystyle\PRS_\kappa(q)=\alpha\E\log\Phi\Bigl(\frac{\sqrt q\,G-\kappa}{\sqrt{1-q}}\Bigr)+\frac q{2(1-q)}+\frac12\log(1-q)$. The same formula holds at $q=0$. \item $\displaystyle D_\gamma(q)=\frac{q\bigl(\alpha\rsF(k,\beta)-1\bigr)}{(1-q)^2}$ and $S_\gamma(q)=\alpha\rsH(k,\beta)$. Moreover $D_\gamma\in C^1([0,r])$ and $D_\gamma'=(S_\gamma-1)/\lambda_\gamma^2$. \item $q\mapsto\PRS_\kappa(q)$ is $C^1$ on $(0,1)$ with $\frac{\dd}{\dd q}\PRS_\kappa(q)=-\frac12D_{\gamma_q}(q)$. \item If $q$ is RS-stationary, then $\cH_\gamma(s)-\cH_\gamma(q)=\Lambda_{\mathrm R}(s-q)^2+o((s-q)^2)$ as $s\to q$. \item If $q$ is RS-stationary and $\Lambda_{\mathrm R}<0$, then $q$ is a strict local maximum of $\cH_\gamma$. For every $q'\ne q$ sufficiently close to $q$ and all sufficiently small $t>0$, \[ \cP_\kappa\bigl((1-t)\gamma_q+t\,\gamma_{q'}\bigr)<\cP_\kappa(\gamma_q) . \] In particular $\gamma_q$ is neither a global minimizer of $\cP_\kappa$ nor a local minimizer for the $L^1$ distance. \end{enumerate} \end{lemma} \begin{proof} (a) Since $\gamma=0$ on $[0,q)$, \Cref{II:fun:colehopf}(a) with $m=0$ and \eqref{II:fun:top} give $u_\gamma(0,0;\kappa)=\E u_\gamma(q,B_q;\kappa)=\E\log\Phi(Z^q)$. Since $\lambda_\gamma=1-q$ on $[0,q]$ and $\lambda_\gamma(t)=1-t$ on $[q,1]$, \[ \CS(\gamma)=\frac12\int_0^q\Bigl(\frac1{1-q}-\frac1{1-t}\Bigr)\dd t=\frac q{2(1-q)}+\frac12\log(1-q). \] At $q=0$ both sides equal $\alpha\log\Phi(-\kappa)$. (b) By \eqref{II:fun:topderiv} with $z=(x-\kappa)/\sqrt{1-q}$ we have $b(q,x)=R(z)/\sqrt{1-q}$ and $c(q,x)=V(z)/(1-q)$, and $X_q=B_q$ has the law of $\sqrt q\,G$. Hence, as $\lambda_\gamma=1-q$ on $[0,q]$, \[ D_\gamma(q)=\frac{\alpha\E R(Z^q)^2}{1-q}-\frac q{(1-q)^2}=\frac{\alpha(1-q)\E R(Z^q)^2-q}{(1-q)^2}, \] and $\alpha(1-q)\E R(Z^q)^2=q\alpha\rsF(k,\beta)$ because $\beta^2=(1-q)/q$. Similarly $S_\gamma(q)=\alpha(1-q)^2\E c(q,X_q)^2=\alpha\E V(Z^q)^2$. The last claim is \Cref{II:lem:generator}(c)(i). (c) The derivative of $\frac q{2(1-q)}+\frac12\log(1-q)$ is $\frac q{2(1-q)^2}$. Next \[ \frac{\dd Z^q}{\dd q}=\frac{G}{2\sqrt{q(1-q)}}+\frac{Z^q}{2(1-q)}, \] and Gaussian integration by parts with $R'=-V$ gives $\E[R(Z^q)G]=-\sqrt{q/(1-q)}\,\E V(Z^q)$. Since $V(z)=zR(z)+R(z)^2$, \[ \frac{\dd}{\dd q}\E\log\Phi(Z^q)=\E\Bigl[R(Z^q)\frac{\dd Z^q}{\dd q}\Bigr]=\frac{\E[R(Z^q)Z^q-V(Z^q)]}{2(1-q)}=-\frac{\E R(Z^q)^2}{2(1-q)} . \] Differentiation under $\E$ is justified by $0q$ is identical with $\gamma_{q'}-\gamma_q=-\1_{[q,q')}$. Moreover $\|(1-t)\gamma_q+t\gamma_{q'}-\gamma_q\|_{L^1}=t|q'-q|\to0$. \end{proof} Part (e) is the analogue for $\cP_\kappa$ of the instability of the replica-symmetric solution beyond the de Almeida--Thouless line for SK-type models \cite{Toninelli2002}, \cite[Theorem~1.8]{JagannathTobasco2017}, recalled in \Cref{II:sec:intro}. \section{The stationary curve and the crossing}\label{II:rs:sec-curve} We now locate the RS-stationary points. By \Cref{II:rs:lem-formulas}(b), $q$ is RS-stationary exactly when $\alpha\rsF(k,\beta)=1$. Uniqueness of the stationary point comes from an exact formula for the derivative of $D_{\gamma_q}(q)$ in $q$ (\Cref{II:rs:lem-unique}); the crossing is then located by following the replicon eigenvalue along the stationary curve, parametrized by $\beta$. \begin{lemma}\label{II:rs:lem-F} Let $k\in\R$ and $\beta>0$, and write $Z:=(G-k)/\beta$. \begin{enumerate}[label=\textup{(\alph*)}] \item $\rsF(k,\beta)=m(k)+\beta^2\E\varrho(Z)$; hence $m(k)<\rsF(k,\beta)\le m(k)+M_\varrho\beta^2$. \item $|\rsF(k,\beta)-m(k)-2\beta^2\Phi(k)|\le C_F\beta^3$, with $C_F:=\phi_0\|\varrho-2\cdot\1_{(-\infty,0)}\|_{L^1}$. \item $\rsF$ is continuous on $\R\times(0,\infty)$ and strictly increasing in each variable. For fixed $\beta$, $\rsF(\cdot,\beta)$ increases from $0$ to $\infty$. \item For $\alpha>0$ and $\beta>0$ there is a unique $K_\alpha(\beta)$ with $\alpha\rsF(K_\alpha(\beta),\beta)=1$. The function $K_\alpha$ is continuous and strictly decreasing, and $m(K_\alpha(\beta))\in[\frac1\alpha-M_\varrho\beta^2,\frac1\alpha)$. In particular $K_\alpha(\beta)<\kRS(\alpha)$, and $K_\alpha(\beta)\uparrow\kRS(\alpha)$ as $\beta\downarrow0$. \end{enumerate} \end{lemma} \begin{proof} (a) Since $\beta Z_-=(k-G)_+$, we have $\beta^2\E Z_-^2=m(k)$ and $\beta^2\E R(Z)^2=m(k)+\beta^2\E\varrho(Z)$. Use $0<\varrho\le M_\varrho$. (b) $Z$ has density $z\mapsto\beta\phi(k+\beta z)\le\beta\phi_0$, and $\Prob(Z<0)=\Phi(k)$. Hence $|\E\varrho(Z)-2\Phi(k)|\le\beta\phi_0\|\varrho-2\cdot\1_{(-\infty,0)}\|_{L^1}$, which is finite by \Cref{II:fun:gauss}(f). (c) $\rsF(k,\beta)=\E f_{k,\beta}(G)^2$ with $f_{k,\beta}(g):=\beta R((g-k)/\beta)>0$. For fixed $g$, $f_{k,\beta}(g)$ is strictly increasing in $k$ because $R$ is decreasing. It is strictly increasing in $\beta$ because, with $z=(g-k)/\beta$, \[ \partial_\beta f_{k,\beta}(g)=R(z)-zR'(z)=R(z)\bigl(1+z\Rbar(z)\bigr)=R(z)\,\E T_z^2>0 \] by \Cref{II:fun:gauss}(a). Since $R(z)\le z_-+R(0)$ (\Cref{II:fun:gauss}(a)), we have $f_{k,\beta}(g)\le(k-g)_++\beta R(0)$, and dominated convergence gives continuity and $\rsF\to0$ as $k\to-\infty$, where $f_{k,\beta}\to0$ pointwise. Finally $\rsF\ge m(k)\to\infty$ as $k\to\infty$. (d) Existence and uniqueness follow from (c). If $\beta<\beta'$, then $\alpha\rsF(K_\alpha(\beta),\beta')>1$, so $K_\alpha(\beta')0$; the strict inequalities $\alpha\rsF(K_\alpha(\beta)-\varepsilon,\cdot)<1<\alpha\rsF(K_\alpha(\beta)+\varepsilon,\cdot)$ hold at $\beta$, hence near $\beta$, and there $|K_\alpha-K_\alpha(\beta)|<\varepsilon$. By (a), $m(K_\alpha(\beta))=\frac1\alpha-\beta^2\E\varrho(Z)\in[\frac1\alpha-M_\varrho\beta^2,\frac1\alpha)$. Since $m$ is continuous and strictly increasing and $\alpha m(\kRS)=1$, this gives $K_\alpha(\beta)<\kRS$ and $K_\alpha(\beta)\to\kRS$ as $\beta\to0$. \end{proof} The next lemma is the structural input for the uniqueness of the stationary RS point: along the RS family, every zero of $D$ is a strict down-crossing. It holds for every margin below $\kRS$, as soon as $\alpha<\pi^2/4$. \begin{lemma}\label{II:rs:lem-unique} Let $0<\alpha<\pi^2/4$ and $\kappa<\kRS(\alpha)$. For $q\in[0,1)$ put $\delta_q=1-q$, $Z^q=(\sqrt q\,G-\kappa)/\sqrt{\delta_q}$, \begin{equation}\label{II:rs:eq-DRS} D_{\rm RS}(q;\kappa):=\frac\alpha{\delta_q}\E R(Z^q)^2-\frac q{\delta_q^2},\qquad S_{\rm RS}(q;\kappa):=\alpha\E V(Z^q)^2 . \end{equation} Then $D_{\rm RS}(q;\kappa)=D_{\gamma_q}(q)$ and $S_{\rm RS}(q;\kappa)=S_{\gamma_q}(q)$. \begin{enumerate}[label=\textup{(\alph*)}] \item $q\mapsto D_{\rm RS}(q;\kappa)$ is $C^1$ on $[0,1)$, and \begin{equation}\label{II:rs:eq-DRSderiv} \delta_q^2\,\partial_qD_{\rm RS}=\alpha\E\bigl[V^2-2R^2(1-V)\bigr](Z^q)-1+2\delta_qD_{\rm RS}. \end{equation} \item There is exactly one $q^*(\kappa)\in(0,1)$ with $D_{\rm RS}(q^*(\kappa);\kappa)=0$. Moreover $D_{\rm RS}>0$ on $[0,q^*(\kappa))$ and $D_{\rm RS}<0$ on $(q^*(\kappa),1)$. \item $\kappa\mapsto q^*(\kappa)$ is continuous and strictly increasing on $(-\infty,\kRS)$. \end{enumerate} \end{lemma} \begin{proof} For $q>0$ the identities $D_{\rm RS}=D_{\gamma_q}(q)$ and $S_{\rm RS}=S_{\gamma_q}(q)$ are contained in the proof of \Cref{II:rs:lem-formulas}(b). For $q=0$ the order parameter $\gamma_0=\1_{[0,1]}$ has $u(q,x)=\log\Phi((x-\kappa)/\sqrt{1-q})$ on $[0,1)$ by \eqref{II:fun:top}, so $D_{\gamma_0}(0)=\alpha b(0,0)^2=\alpha R(-\kappa)^2$ and $S_{\gamma_0}(0)=\alpha V(-\kappa)^2$, as claimed. (a) Let $\bar b(q,x):=\partial_x\log\Phi\bigl((x-\kappa)/\sqrt{\delta_q}\bigr)=R\bigl((x-\kappa)/\sqrt{\delta_q}\bigr)/\sqrt{\delta_q}$, so that $D_{\rm RS}=\alpha\E\bar b(q,B_q)^2-q/\delta_q^2$ for a Brownian motion $B$. The function $\Phi((x-\kappa)/\sqrt{\delta_q})=\Prob(x+\sqrt{\delta_q}\,G\ge\kappa)$ solves the backward heat equation, so its logarithm $\bar u$ solves $\bar u_q=-\frac12(\bar u_{xx}+\bar u_x^2)$, and $\bar b=\bar u_x$ solves $\bar b_q=-\frac12(\bar b_{xx}+2\bar b\bar b_x)$; also $\bar b_x=-V((x-\kappa)/\sqrt{\delta_q})/\delta_q$. On $[0,r]\times\R$ with $r<1$ these functions are smooth, $\bar b$ grows at most linearly in $x$ (as $R(z)\le z_-+R(0)$), and $\bar b_x,\bar b_{xx}$ are bounded, because $V$ and $V'=R(1-V-\Rbar^2)$ are bounded by \Cref{II:fun:gauss}(a),(b),(e). By It\^o's formula, \[ \frac{\dd}{\dd q}\E\bar b(q,B_q)^2=\E\bigl[2\bar b\bar b_q+\bar b_x^2+\bar b\bar b_{xx}\bigr]=\E\bigl[\bar b_x^2-2\bar b^2\bar b_x\bigr]=\delta_q^{-2}\,\E\bigl[V^2+2R^2V\bigr](Z^q), \] which is continuous in $q$. Since $\partial_q(q/\delta_q^2)=\delta_q^{-2}+2q\delta_q^{-3}$ and $2q\delta_q^{-3}=2\delta_q^{-2}\bigl(\alpha\E R(Z^q)^2-\delta_qD_{\rm RS}\bigr)$ by the definition of $D_{\rm RS}$, \eqref{II:rs:eq-DRSderiv} follows. (b) Put $\mathfrak p:=V^2-2R^2(1-V)$. For $z\ge0$, $\mathfrak p\le V^2\le V(0)^2=4/\pi^2$, because $V$ is decreasing and $V(0)=R(0)^2=2/\pi$. For $z<0$, \eqref{II:fun:khintchine} gives $2R^2(1-V)\ge\frac23R^2\Rbar^2=\frac23V^2$, so $\mathfrak p\le\frac13V^2<\frac13<\frac4{\pi^2}$. Hence at every zero $q$ of $D_{\rm RS}(\cdot;\kappa)$, (a) gives $\delta_q^2\partial_qD_{\rm RS}\le4\alpha/\pi^2-1<0$. A $C^1$ function on an interval whose zeros are all strict down-crossings has at most one zero. Indeed, suppose $q_10$, while by \Cref{II:rs:lem-formulas}(b) and \Cref{II:rs:lem-F}(a), $\delta_q^2D_{\rm RS}=q(\alpha\rsF(\kappa/\sqrt q,\beta)-1)$ with $\beta^2=\delta_q/q$ and $m(\kappa/\sqrt q)<\rsF(\kappa/\sqrt q,\beta)\le m(\kappa/\sqrt q)+M_\varrho\beta^2$. Thus $\delta_q^2D_{\rm RS}\to\alpha m(\kappa)-1<0$ as $q\uparrow1$. The intermediate value theorem and uniqueness give the sign pattern. (c) For fixed $q$, $Z^q$ decreases in $\kappa$ and $R$ is positive and decreasing, so $D_{\rm RS}(q;\kappa)$ is strictly increasing in $\kappa$; it is continuous in $\kappa$ by dominated convergence. If $\kappa<\kappa'<\kRS$, then $D_{\rm RS}(q^*(\kappa);\kappa')>0$, so $q^*(\kappa)0$. The strict inequalities $D_{\rm RS}(q^*(\kappa)-\varepsilon;\cdot)>0>D_{\rm RS}(q^*(\kappa)+\varepsilon;\cdot)$ (the first omitted if $q^*(\kappa)\le\varepsilon$) hold at $\kappa$, hence near $\kappa$, and there $|q^*-q^*(\kappa)|<\varepsilon$ by (b). \end{proof} The stability function $\rsH$ is expanded next. Its first-order term in $\beta$ comes from the signed area $\int(V^2-\1_{\{z<0\}})=-c_0$ of \Cref{II:fun:gauss}(f). \begin{lemma}\label{II:rs:lem-H} There is a numerical constant $C_H<\infty$ such that, for $|k|\le1$ and $0<\beta\le\frac12$, \[ \begin{gathered} \bigl|\rsH(k,\beta)-\Phi(k)+c_0\beta\phi(k)\bigr|\le C_H\beta^2\log(1/\beta),\\ \bigl|\partial_\beta\rsH(k,\beta)+c_0\phi(k)\bigr|\le C_H\beta\log(1/\beta). \end{gathered} \] Moreover $\rsH(\cdot,\beta)$ is nondecreasing for every $\beta>0$. \end{lemma} \begin{proof} Put $\mathfrak v:=V^2-\1_{\{z<0\}}$, so that $|\mathfrak v(z)|\le C_{\rm g}/(1+z^2)$ and $\int\mathfrak v=-c_0$ by \Cref{II:fun:gauss}(f). Since $Z=(G-k)/\beta$ has density $\beta\phi(k+\beta z)$ and $\Prob(Z<0)=\Phi(k)$, \begin{align*} \rsH(k,\beta)&=\Phi(k)+\beta\int_\R\phi(k+\beta z)\mathfrak v(z)\dd z\\ &=\Phi(k)-c_0\beta\phi(k)+\beta\int_\R\bigl(\phi(k+\beta z)-\phi(k)\bigr)\mathfrak v(z)\dd z . \end{align*} Since $|\phi'|\le\phi_0$, $|\phi(k+\beta z)-\phi(k)|\le\phi_0\min\{1,\beta|z|\}$, and \begin{align*} \int_\R\frac{\min\{1,\beta|z|\}}{1+z^2}\dd z&\le\beta\int_{|z|\le1/\beta}\frac{|z|}{1+z^2}\dd z+\int_{|z|>1/\beta}\frac{\dd z}{z^2}\\ &=\beta\log\bigl(1+\beta^{-2}\bigr)+2\beta\le C\beta\log(1/\beta) \end{align*} for $\beta\le\frac12$. This proves the first bound. Differentiating under the integral, which is justified by the same majorants, \[ \partial_\beta\rsH=\int_\R\phi(k+\beta z)\mathfrak v(z)\dd z+\beta\int_\R z\,\phi'(k+\beta z)\mathfrak v(z)\dd z . \] The first integral is $-c_0\phi(k)+O(\beta\log(1/\beta))$ by the estimate above. Since $|z|/(1+z^2)\le\beta$ for $|z|>1/\beta$ and $\int_\R|\phi'(k+\beta z)|\dd z=\beta^{-1}\|\phi'\|_{L^1}=2\phi_0\beta^{-1}$, the second is at most \[ \beta C_{\rm g}\Bigl[\phi_0\int_{|z|\le1/\beta}\frac{|z|}{1+z^2}\dd z+\beta\int_{|z|>1/\beta}|\phi'(k+\beta z)|\dd z\Bigr]\le C\beta\log(1/\beta). \] Finally, $Z=(G-k)/\beta$ decreases in $k$ and $V$ is positive and decreasing, so $V(Z)^2$ increases in $k$. \end{proof} For $\alpha>0$ and $\beta>0$ put \[ \kappa_\alpha(\beta):=\frac{K_\alpha(\beta)}{\sqrt{1+\beta^2}},\qquad \varpi_\alpha(\beta):=\alpha\rsH\bigl(K_\alpha(\beta),\beta\bigr)-1 . \] For $q\in(0,1)$ write $\beta(q)=\sqrt{(1-q)/q}$, so that $\kappa/\sqrt q=\kappa\sqrt{1+\beta(q)^2}$. By \Cref{II:rs:lem-formulas}(b) and part (a) of the next lemma, $\varpi_\alpha(\beta(q))=S_{\gamma_q}(q)-1=-4(1-q)^2\Lambda_{\mathrm R}$ whenever $q$ is RS-stationary. \begin{lemma}\label{II:rs:lem-curve} \begin{enumerate}[label=\textup{(\alph*)}] \item For every $\alpha>0$, $\kappa\in\R$, and $q\in(0,1)$, $\operatorname{sign}D_{\gamma_q}(q)=\operatorname{sign}\bigl(\kappa-\kappa_\alpha(\beta(q))\bigr)$. Thus $q$ is RS-stationary at $\kappa$ exactly when $\kappa=\kappa_\alpha(\beta(q))$, and then $k=K_\alpha(\beta)$ and $\alpha\rsH(k,\beta)-1=\varpi_\alpha(\beta)$. \item There are numerical constants $\beta_0\in(0,\frac14]$, $e_{\rm a}>0$, and $C_{\rm c}<\infty$ such that for $00$ and $c_2>0$ such that for every $00$ for $q\in[0,q^*)$ and $D_{\gamma_q}(q)<0$ for $q\in(q^*,1)$; $q^*(\kappa)$ is the unique minimizer of $\PRS_\kappa$ on $[0,1)$; and $\kappa\mapsto q^*(\kappa)$ is continuous and strictly increasing. \item \emph{The crossing and the replicon eigenvalue.} There is exactly one $\kd=\kd(\alpha)\in(\kRS-c_2,\kRS)$ at which $\alpha\rsH=1$ at $q^*(\kd)$. Put $\qd:=q^*(\kd)$ and $\gamma_{\mathrm d}:=\1_{[\qd,1]}$. The replicon eigenvalue $\Lambda_{\mathrm R}(\kappa)$ of \Cref{II:def:rs} at $q^*(\kappa)$ satisfies \[ \Lambda_{\mathrm R}(\kappa)<0\quad(\kd<\kappa<\kRS),\qquad \Lambda_{\mathrm R}(\kd)=0,\qquad \Lambda_{\mathrm R}(\kappa)>0\quad(\kRS-c_2\le\kappa<\kd). \] Consequently, for $\kappa\in(\kd,\kRS)$ the stationary replica-symmetric order parameter $\gamma_{q^*(\kappa)}$ is unstable in the sense of \Cref{II:rs:lem-formulas}(e). \item \emph{Asymptotics.} As $e\downarrow0$, with $\beta_{\mathrm d}:=\beta(\qd)$, \[ \begin{gathered} \kRS-\kd=\thd e^2+O\bigl(e^3\log(1/e)\bigr),\qquad \thd=\frac{(2\pi)^{3/2}}{128\,c_0^2},\\ \beta_{\mathrm d}=\frac{e}{8c_0\phi_0}+O\bigl(e^2\log(1/e)\bigr), \end{gathered} \] and the rescaled top gap of $\gamma_{\mathrm d}$ satisfies \[ d_{\mathrm d}:=\frac{1-\qd}{e^2}=\frac1{64c_0^2\phi_0^2}+O\bigl(e\log(1/e)\bigr)=2\phi_0\thd+O\bigl(e\log(1/e)\bigr). \] \item \emph{The stationary point in the window.} For every compact $\Theta\subset(0,\infty)$ there are $e_\Theta>0$ and $C_\Theta<\infty$ such that, for $00$, and $1/(1+\beta^2)$ is RS-stationary at $\kappa_\alpha(\beta)$ by \Cref{II:rs:lem-curve}(a); so $\beta^*(\kappa_\alpha(\beta))=\beta$ by (i). In particular $\kappa_\alpha(\beta_0)\le\kRS-c_2$, hence $\beta^*(\kappa)\in(0,\beta_0]$ for $\kappa\in[\kRS-c_2,\kRS)$. By \Cref{II:rs:lem-curve}(b), $\varpi_\alpha$ is continuous and strictly decreasing on $(0,\beta_0]$, $\varpi_\alpha(0+)=\alpha\Phi(\kRS)-1>0$ by \Cref{II:lem:krs}(b), and \[ \varpi_\alpha(\beta_0)\le\alpha\Phi(\kRS)-1-\tfrac12c_0\phi_0\beta_0\le\alpha\phi_0|\kRS|-\tfrac12c_0\phi_0\beta_0<0 \] after decreasing $e_1$ (\Cref{II:lem:krs}(a),(b)). Hence $\varpi_\alpha$ has exactly one zero $\beta_{\mathrm d}\in(0,\beta_0)$, with $\varpi_\alpha>0$ on $(0,\beta_{\mathrm d})$ and $\varpi_\alpha<0$ on $(\beta_{\mathrm d},\beta_0]$. Put $\kd:=\kappa_\alpha(\beta_{\mathrm d})$, so that $\beta^*(\kd)=\beta_{\mathrm d}$. By (iii) below, $\beta_{\mathrm d}\le\beta_0/2$ after decreasing $e_1$, so $\kRS-\kd\le\frac{3\beta_{\mathrm d}^2}{4\phi_0}\kd$. This gives the signs and the uniqueness of $\kd$, and \Cref{II:rs:lem-formulas}(e) applies for $\kappa\in(\kd,\kRS)$. (iii) Letting $\beta'\downarrow0$ in \eqref{II:rs:eq-varpimono} and using \Cref{II:lem:krs}(b), $0=\varpi_\alpha(\beta_{\mathrm d})\le\alpha\phi_0|\kRS|-\frac12c_0\phi_0\beta_{\mathrm d}$, so $\beta_{\mathrm d}\le2\alpha|\kRS|/c_0\le Ce$ by \Cref{II:lem:krs}(a). As $x^2\log(1/x)$ is increasing on $(0,\frac12]$, $\beta_{\mathrm d}^2\log(1/\beta_{\mathrm d})\le Ce^2\log(1/e)$, and the expansion \eqref{II:rs:eq-curve} of $\varpi_\alpha$ at $\beta_{\mathrm d}$ gives $|\frac e4-2c_0\phi_0\beta_{\mathrm d}|\le Ce^2\log(1/e)$. Hence \[ \beta_{\mathrm d}=\frac{e}{8c_0\phi_0}+O\bigl(e^2\log(1/e)\bigr),\qquad \beta_{\mathrm d}^2=\frac{e^2}{64c_0^2\phi_0^2}+O\bigl(e^3\log(1/e)\bigr). \] By \eqref{II:rs:eq-curve} at $\beta=\beta_{\mathrm d}$, $\kRS-\kd=\beta_{\mathrm d}^2/(2\phi_0)+O(e^3)=e^2/(128c_0^2\phi_0^3)+O(e^3\log(1/e))$, and $1/(128\phi_0^3)=(2\pi)^{3/2}/128$. Finally $1-\qd=\beta_{\mathrm d}^2/(1+\beta_{\mathrm d}^2)$, so $d_{\mathrm d}=\beta_{\mathrm d}^2e^{-2}(1+O(e^2))$, and $2\phi_0\thd=1/(64c_0^2\phi_0^2)$. (iv) Let $e_\Theta\le e_1$ be so small that $e_\Theta^2\max\Theta\le c_2$; then $\kappa_e(\theta)\in[\kRS-c_2,\kRS)$, and $\beta^*:=\beta^*(\kappa_e(\theta))\in(0,\beta_0]$ by the proof of (ii). By \Cref{II:rs:lem-curve}(a) and \eqref{II:rs:eq-curve}, \[ e^2\theta=\kRS-\kappa_\alpha(\beta^*)=\frac{\beta^{*2}}{2\phi_0}+O\bigl(e\beta^{*2}+\beta^{*3}\bigr),\qquad \frac{\beta^{*2}}{4\phi_0}\le e^2\theta\le\frac{3\beta^{*2}}{4\phi_0} . \] Hence $\frac43\phi_0\theta e^2\le\beta^{*2}\le4\phi_0\theta e^2$, and then $\beta^{*2}=2\phi_0\theta e^2+O(e^3)$, with constants depending on $\Theta$. Since $1-q^*=\beta^{*2}/(1+\beta^{*2})$, the first claim follows. For the second, $\beta^*=e\sqrt{2\phi_0\theta}+O(e^2)$, $\beta^{*2}\log(1/\beta^*)\le Ce^2\log(1/e)$, and $1-\alpha\rsH=-\varpi_\alpha(\beta^*)$ by \Cref{II:rs:lem-curve}(a); insert these into \eqref{II:rs:eq-curve}. Finally $2\phi_0\thd=1/(64c_0^2\phi_0^2)$ gives $c_0t(\thd)=c_0\cdot2\phi_0\cdot\frac1{8c_0\phi_0}=\frac14$. \end{proof} At the stationary RS point, $\cH_{\gamma_{q^*}}(s)-\cH_{\gamma_{q^*}}(q^*)=\Lambda_{\mathrm R}(s-q^*)^2+o((s-q^*)^2)$ by \Cref{II:rs:lem-formulas}(d). The instability for $\kappa>\kd$ concerns mass split off near the atom. The longitudinal direction, in which the whole atom moves, is stable on both sides of $\kd$, since $q^*$ minimizes $\PRS_\kappa$. \Cref{II:prop:crossing} does not assert that $\gamma_{q^*}$ is a global minimizer for $\kappa\le\kd$; for $\kappa\in[\kd-ce^2,\kd]$ this is proved in \Cref{II:sec:frsb} (\Cref{II:thm:uniform}). \section{The zero-temperature functional and a dual certificate}\label{II:rs:sec-J} The critical margin is defined through the divergence of $\cP_*$, which is not directly computable. \Cref{II:thm:kJ} characterizes it through the zero-temperature functional $J^k$ of \eqref{II:fun:J}, in which the terminal datum $\log\Phi((x-\kappa)/\sqrt\delta)$ is replaced by its Gaussian-tail approximation $-(\kappa-x)_+^2/(2\delta)$ and the top gap is scaled out. We first record the consequences of \Cref{II:sec:functional} used below. We then prove a duality bound: a single function $g\ge k-x$ satisfying two Gaussian moment constraints forces $J^k\ge0$ for all coefficients at once, and hence $\kc\ge k$. For $T\in(0,1]$ let $\rsS_T$ be the set of nondecreasing, right-continuous step functions $\xi:[0,T)\to[0,\infty)$ with finitely many values. For $\xi\in\rsS_T$ and $k\in\R$, $\Upsilon=\Upsilon^k_{T,\xi}$ denotes the solution of \eqref{II:fun:quadpde}, given by the Cole--Hopf recursion run backwards from the terminal datum $f_k(x)=-\frac12(k-x)_+^2$ at time $T$ with coefficient $\xi$. As in \eqref{II:fun:J}, $\Lambda_\xi(q)=1+\int_q^T\xi$, $L_\xi=\Lambda_\xi(0)$, $J^k_T(\xi)=\alpha\Upsilon(0,0)+\frac12\int_0^T\Lambda_\xi(q)^{-1}\dd q$ and $J^k=J^k_1$. The coefficient $\xi$ may exceed one. For $\xi=0$, $\Upsilon(0,0)=-\frac12\E(k-\sqrt TG)_+^2$, so \begin{equation}\label{II:rs:eq-J0} J^k(0)=\tfrac12\bigl(1-\alpha m(k)\bigr), \end{equation} which vanishes exactly at $k=\kRS$. \begin{lemma}\label{II:rs:lem-J} Let $\alpha>0$ and $k\in\R$. \begin{enumerate}[label=\textup{(\alph*)}] \item \emph{Two criteria.} If $J^k(\xi)<0$ for some $\xi\in\rsS_1$, then $\kc\le k$. If $J^k(\xi)\ge0$ for every $\xi\in\rsS_1$, then $\kc\ge k$. \item \emph{The optimal process.} Let $\xi\in\rsS_1$ and $\Upsilon=\Upsilon^k_{1,\xi}$, let $B$ be a Brownian motion with natural filtration $(\cF_q)$, and let $X^*$ be the strong solution of $\dd X^*_q=\xi(q)\Upsilon_x(q,X^*_q)\dd q+\dd B_q$, $X^*_0=0$. Then $M_q:=\Upsilon_x(q,X^*_q)$ is a square-integrable $(\cF_q)$-martingale on $[0,1]$ with $M_1=(k-X^*_1)_+$, and \[ \Upsilon(0,0)=-\frac12\,\E\Bigl[M_1^2+\int_0^1\xi(q)M_q^2\dd q\Bigr]. \] \end{enumerate} \end{lemma} \begin{proof} (a) By \Cref{II:thm:kJ}, $\kc=\kappa^0_{\mathrm c}$, the infimum of the margins $k'$ for which $J^{k'}(\xi)<0$ for some nondecreasing step function $\xi$ on $[0,1)$. The first claim follows. For the second, $k'\mapsto J^{k'}(\xi)$ is nonincreasing by \Cref{II:fun:quad}(b), so $J^{k'}(\xi)\ge J^k(\xi)\ge0$ for all $k'\le k$ and $\xi\in\rsS_1$; hence $\kappa^0_{\mathrm c}\ge k$. (b) The terminal datum $f_k$ belongs to the class of terminal data of \Cref{II:fun:control}, with $f_k'(x)=(k-x)_+$, and $\Upsilon$ is the solution considered there. By \Cref{II:fun:control}, $X^*$ exists and $M$ is a square-integrable martingale with $M_1=f_k'(X^*_1)=(k-X^*_1)_+$; it is adapted to $(\cF_q)$ because $X^*$ is a strong solution. Equality holds in the control inequality \eqref{II:fun:controlineq} for the feedback control $\theta^*_q=\xi(q)M_q$, whose cost is $\theta^{*2}_q/\xi(q)=\xi(q)M_q^2$. Hence $\Upsilon(0,0)=\E f_k(X^*_1)-\frac12\E\int_0^1\xi M_q^2\dd q$, and $f_k(X^*_1)=-\frac12M_1^2$. \end{proof} We now prove the duality bound. A trial control in the variational representation of $\Upsilon(0,0)$ (\Cref{II:fun:control}) gives a lower bound on $J^k(\xi)$ for a fixed $\xi$. To obtain a nonnegative lower bound uniformly in $\xi$, we use a dual object: a function $g$ lying above $k-x$ that is small in $L^2$ of the Gaussian measure and has small slope in the same space. Neither condition involves $\xi$. \begin{theorem}\label{II:rs:thm-dual} Let $\alpha>0$ and $k\in\R$, and let $g:\R\to\R$ be Lipschitz with $g(x)\ge k-x$ for all $x\in\R$. Assume \begin{equation}\label{II:rs:eq-C12} \text{\textup{(C1)}}\quad \alpha\E g(G)^2\le1,\qquad\qquad \text{\textup{(C2)}}\quad \alpha\E g'(G)^2\le1 . \end{equation} Then $J^k(\xi)\ge0$ for every $\xi\in\rsS_1$. \end{theorem} \begin{proof} Fix $\xi\in\rsS_1$, write $\Lambda=\Lambda_\xi$ and $L=L_\xi$, and let $B$, $(\cF_t)$, $\Upsilon$, $X^*$, and $M$ be as in \Cref{II:rs:lem-J}(b). Put $\nu:=M_1=(k-X^*_1)_+\ge0$, so that $M_s=\E[\nu\mid\cF_s]$. We divide the proof into four steps. \proofstep{0}{Martingale representations.} Let $(P_t)$ denote the heat semigroup. Since $g$ is Lipschitz, $\partial_xP_tg=P_tg'$ for $t>0$ and $|P_tg'|\le\Lip(g)$. The martingale $\E[g(B_1)\mid\cF_r]=(P_{1-r}g)(B_r)$ has terminal value $g(B_1)$, and for $r'<1$ It\^o's formula, with $\partial_r(P_{1-r}g)=-\frac12\partial_{xx}P_{1-r}g$, gives \[ (P_{1-r'}g)(B_{r'})=\E g(B_1)+\int_0^{r'}(P_{1-r}g')(B_r)\dd B_r . \] Letting $r'\uparrow1$ in $L^2$, we obtain $g(B_1)=\E g(B_1)+\int_0^1\varsigma_r\dd B_r$ with $\varsigma_r:=(P_{1-r}g')(B_r)$, which is the Clark--Ocone formula for $g(B_1)$. By Jensen's inequality and (C2), \begin{equation}\label{II:rs:eq-varsigma} \E\varsigma_r^2\le\E(P_{1-r}g'^2)(B_r)=\E g'(G)^2\le\frac1\alpha\qquad(0\le r<1). \end{equation} Similarly write $\nu=\E\nu+\int_0^1n_r\dd B_r$ (by the martingale representation theorem; see \cite{KaratzasShreve1991}), so that $\E M_s^2=(\E\nu)^2+\int_0^s\E n_r^2\dd r$. \proofstep{1}{$J$ in terms of $\nu$.} Put \[ \mathcal Q:=\E\nu^2+\int_0^1\xi(s)\,\E M_s^2\dd s=(\E\nu)^2L+\int_0^1\Lambda(r)\,\E n_r^2\dd r , \] where the second expression follows from the formula for $\E M_s^2$, Fubini's theorem, and $\Lambda(r)=1+\int_r^1\xi$. By \Cref{II:rs:lem-J}(b), $J^k(\xi)=\frac12\int_0^1\Lambda^{-1}-\frac\alpha2\mathcal Q$. Moreover $\mathcal Q=\E[\nu(k-B_1)]$. Indeed $k-B_1=(k-X^*_1)+\int_0^1\xi M_s\dd s$ and $\nu(k-X^*_1)=\nu^2$, so \[ \E[\nu(k-B_1)]=\E\nu^2+\int_0^1\xi(s)\E[\nu M_s]\dd s=\E\nu^2+\int_0^1\xi(s)\E M_s^2\dd s=\mathcal Q . \] \proofstep{2}{Comparison with $g(B_1)$.} Since $\nu\ge0$ and $g(B_1)\ge k-B_1$, the It\^o isometry and the Cauchy--Schwarz inequality, applied to $(\sqrt L\,\E\nu)(\E g(B_1)/\sqrt L)$ and to $(\sqrt{\Lambda(r)}\,n_r)(\varsigma_r/\sqrt{\Lambda(r)})$, give \[ \begin{gathered} \mathcal Q\le\E[\nu g(B_1)]=\E\nu\,\E g(B_1)+\int_0^1\E[n_r\varsigma_r]\dd r\le\mathcal Q^{1/2}\mathcal R^{1/2},\\ \mathcal R:=\frac{(\E g(B_1))^2}{L}+\int_0^1\frac{\E\varsigma_r^2}{\Lambda(r)}\dd r . \end{gathered} \] Hence $\mathcal Q\le\mathcal R$ and $J^k(\xi)\ge\frac12\bigl[\int_0^1\Lambda^{-1}-\alpha\mathcal R\bigr]$. \proofstep{3}{The sign.} The function $\Lambda^{-1}$ is nondecreasing with $\Lambda(0)^{-1}=1/L$, and $\int_0^1\E\varsigma_r^2\dd r=\Var g(B_1)$ by the It\^o isometry. Hence \[ \begin{aligned} \int_0^1\frac{\dd r}{\Lambda(r)}-\alpha\mathcal R&=\frac1L\bigl[1-\alpha\Var g(B_1)-\alpha(\E g(B_1))^2\bigr]+\int_0^1\Bigl(\frac1{\Lambda(r)}-\frac1L\Bigr)\bigl(1-\alpha\E\varsigma_r^2\bigr)\dd r\\ &\ge\frac1L\bigl(1-\alpha\E g(G)^2\bigr)\ \ge0, \end{aligned} \] by \eqref{II:rs:eq-varsigma} for the integral and (C1) for the last inequality. This proves the theorem. \end{proof} \section{Ordering of the thresholds}\label{II:rs:sec-order} We now construct an explicit certificate. A natural candidate is $g(x)=\beta R((x-k)/\beta)$, the terminal value of $\lambda b$ for a replica-symmetric order parameter. It satisfies $g(x)\ge(k-x)_+$, $\E g(G)^2=\rsF(k,\beta)$ and $\E g'(G)^2=\rsH(k,\beta)$, so (C1) and (C2) read $\alpha\rsF\le1$ and $\alpha\rsH\le1$. Along the stationary curve $\alpha\rsF=1$ they hold only for $\beta\ge\beta_{\mathrm d}$ (on the range of \Cref{II:rs:lem-curve}), and the certified margins $K_\alpha(\beta)\le K_\alpha(\beta_{\mathrm d})=\kd\sqrt{1+\beta_{\mathrm d}^2}$ lie below $\kd$. The loss comes from the left tail: $R(z)^2-z^2\to2$ as $z\to-\infty$, which costs $2\beta^2\Phi(k)$ in (C1) (\Cref{II:rs:lem-F}(b)). A certificate that equals $k-x$ exactly outside a window of width $O(e)$ around $k$ avoids this cost. Its price in (C1) is only cubic in the width, while it still lowers the slope enough to satisfy (C2). \begin{theorem}\label{II:thm:ordering} There are numerical constants $e_2>0$ and $C_{\rm ord}<\infty$ such that for $00$ such that, for $00$ and $\theta_+:=\max\Theta$. Up to \Cref{II:flat:ss-conv-cross}, $\Theta$ is arbitrary, and all constants depend on $\Theta$ only through $\theta_\pm$. From \Cref{II:flat:ss-conv-cross} on, we assume in addition that $\Theta$ contains $\thd$ in its interior; \Cref{II:sec:frsb} takes $\Theta=[\frac12\thd,2\thd]$. \end{remark} We now prove that the replica-symmetric formula overestimates $\kc$ at order $e^3$. The negative witness is a coefficient carried by a top layer of width $\varepsilon\asymp e^2$ and total mass one. At $k=\kRS$ the value $J^{\kRS}(0)$ vanishes by \eqref{II:rs:eq-J0}; the layer gains an amount linear in $\varepsilon$ with coefficient proportional to $\alpha\Phi(\kRS)-1\approx e/4$, at a cost of order $\varepsilon^{3/2}$, and the choice $\varepsilon\asymp e^2$ balances the two. The value of $J^k$ at a one-block coefficient involves exponential moments of the form $\E\exp(-t(a-G)_+^2)$, which also enter the upper bound on the capacity in \cite[Definition~2 and Theorem~4.2]{MontanariZhongZhou2024}. The function $J_\ell(k)$ of Part~I, \Cref{I:sec:gardner}, is $J^k(\xi_\varepsilon)$ at $\varepsilon=\ell$, for the one-block coefficient $\xi_\varepsilon=\varepsilon^{-1}\1_{[1-\varepsilon,1)}$ of Step~1 below. \begin{theorem}\label{II:thm:cubic} There are numerical constants $e_3>0$ and $c_{\rm cub}>0$ such that for $00]\le\Phi(z)$, so $\mathfrak r(z)\le\Phi(z)/\mathfrak a(z)$. The right side is at most $\sqrt2\,\e^{1/4}$ on $[-1,0)$, and at most $\sqrt2\phi_0\e^{-z^2/4}/|z|$ for $z\le-1$ by \Cref{II:fun:gauss}(b). Thus $\mathfrak r\le\bar{\mathfrak r}:=(\Phi/\mathfrak a)\1_{(-\infty,0)}\in L^1(\R)$. Using $\E(Z_\varepsilon)_-^2=m(k;p)/\varepsilon$ with $m(k;p):=\E(k-\sqrt pG)_+^2$, so that $m(k;1)=m(k)$, and $\Prob(Z_\varepsilon<0)=\Phi(k/\sqrt p)$, \[ J^k(\xi_\varepsilon)=\frac{p-\alpha m(k;p)}4+\frac{\varepsilon\log2}2\bigl(1-\alpha\Phi(k/\sqrt p)\bigr)+\alpha\varepsilon\,\E\mathfrak r(Z_\varepsilon). \] \proofstep{3}{Expansion.} Let $|k|\le1$ and $\alpha\le3$. The density of $Z_\varepsilon$ is $z\mapsto\sqrt\varepsilon\,\phi_p(k+\sqrt\varepsilon z)\le\sqrt{2\varepsilon}\,\phi_0$, where $\phi_p$ is the $N(0,p)$ density, so $\alpha\varepsilon\E\mathfrak r(Z_\varepsilon)\le3\sqrt2\phi_0\|\bar{\mathfrak r}\|_{L^1}\varepsilon^{3/2}$. By Gaussian integration by parts, $\partial_tm(k;t)=\Phi(k/\sqrt t)$. For $t\in[\frac12,1]$, $|\partial_t\Phi(k/\sqrt t)|=|k|\phi(k/\sqrt t)/(2t^{3/2})\le\sqrt2\phi_0|k|\le|k|$, so $|\Phi(k/\sqrt t)-\Phi(k)|\le|k|(1-t)$. Since $m(k)-m(k;p)=\int_p^1\Phi(k/\sqrt t)\dd t$, this gives $|m(k;p)-m(k)+\varepsilon\Phi(k)|\le\varepsilon^2$ and $|\Phi(k/\sqrt p)-\Phi(k)|\le\varepsilon$. Hence, with $p=1-\varepsilon$, \[ \begin{aligned} \frac{p-\alpha m(k;p)}4&\le\frac{1-\alpha m(k)}4+\frac\varepsilon4\bigl(\alpha\Phi(k)-1\bigr)+\frac{\alpha\varepsilon^2}4,\\ \frac{\varepsilon\log2}2\bigl(1-\alpha\Phi(k/\sqrt p)\bigr)&\le-\frac{\varepsilon\log2}2\bigl(\alpha\Phi(k)-1\bigr)+\frac{\alpha\log2}2\varepsilon^2 . \end{aligned} \] Adding these bounds and the bound on $\alpha\varepsilon\E\mathfrak r(Z_\varepsilon)$, and using $\varepsilon^2\le\varepsilon^{3/2}$ and $\alpha\le3$, we obtain, with $\mathsf d:=\log2-\frac12$, \begin{equation}\label{II:rs:eq-cubexp} J^k(\xi_\varepsilon)\le\frac{1-\alpha m(k)}4-\frac\varepsilon2\,\mathsf d\,\bigl(\alpha\Phi(k)-1\bigr)+C_J\varepsilon^{3/2}, \end{equation} where $C_J:=\frac34+\frac32\log2+3\sqrt2\phi_0\|\bar{\mathfrak r}\|_{L^1}$ is a numerical constant. Here $\mathsf d>0$ because $\log2=\int_1^2t^{-1}\dd t>\frac12$. \proofstep{4}{Choice of $\varepsilon$.} At $k=\kRS$, $1-\alpha m(\kRS)=0$ and, by \Cref{II:lem:krs}(b), $\alpha\Phi(\kRS)-1\ge c_\Phi e$ with $c_\Phi:=m_1(-\frac12)/(4\phi_0)$. Let $t_*:=(\mathsf dc_\Phi/(4C_J))^2$ and $\varepsilon:=t_*e^2\le\frac12$. Then \eqref{II:rs:eq-cubexp} gives \[ J^{\kRS}(\xi_\varepsilon)\le-\frac{\mathsf dc_\Phi}2t_*e^3+C_Jt_*^{3/2}e^3=-\frac{\mathsf dc_\Phi}4t_*e^3=:-c_4e^3 . \] Let $c_{\rm cub}:=c_4/3$. By \Cref{II:fun:quad}(b), and since $m_1(t)\le m_1(0)=\phi_0$ for $t\le0$ and $\alpha\phi_0\le3\phi_0\le\frac32$, \[ J^{\kRS-c_{\rm cub}e^3}(\xi_\varepsilon)\le J^{\kRS}(\xi_\varepsilon)+\alpha\phi_0c_{\rm cub}e^3\le-c_4e^3+\tfrac12c_4e^3<0 . \] By \Cref{II:rs:lem-J}(a), $\kc\le\kRS-c_{\rm cub}e^3$. \end{proof} The two cubic bounds are statements about the zero-temperature functional $J$ only: \Cref{II:thm:ordering}(i) certifies $J^{k_*}\ge0$ and \Cref{II:thm:cubic} exhibits $J^{\kRS-c_{\rm cub}e^3}(\xi_\varepsilon)<0$; they are transferred to $\kc$ through the characterization $\kc=\kappa^0_{\mathrm c}$ of \Cref{II:thm:kJ}. \begin{corollary}\label{II:cor:nonrs} Let $00$, $\kappa\in\R$, and let $\gamma\in\cU$ have top $Q=1-\delta$ and bottom $q_0=\min\supp\mu_\gamma$; write $\lambda=\lambda_\gamma$ and $u=u_\gamma(\cdot,\cdot;\kappa)$. The \emph{level} and the \emph{rescaled space variable} are \[ s=\frac{1-q}{e^2},\qquad q(s):=1-e^2s,\qquad y=\frac{x-\kappa}e . \] The \emph{rescaled profile} and its primitive are \[ \zeta(s):=\gamma(1-e^2s)\quad(s\in[0,e^{-2}]),\qquad \ell(s):=\int_0^s\zeta(t)\dd t, \] so that $\lambda(q(s))=e^2\ell(s)$. The \emph{top gap}, the \emph{depth}, and the \emph{mass} are \[ d:=\frac\delta{e^2},\qquad s_0:=\frac{1-q_0}{e^2},\qquad \ell_0:=\ell(s_0)=\frac{\lambda(0)}{e^2}. \] On $(0,e^{-2}]\times\R$ we put \[ U(s,y):=u\bigl(q(s),\kappa+ey\bigr),\qquad v(s,y):=-\ell(s)U_{yy}(s,y),\qquad \Psi(s,y):=\ell(s)U_y(s,y)+y, \] \[ F(s,y):=s\bigl(-U_{yy}(s,y)\bigr)-U_y(s,y)\Psi(s,y). \] The endpoint $s=0$ corresponds to the singular hard wall $q=1$ and is not in the domain of these smooth functions; the profile $\zeta$ and its primitive $\ell$ are nevertheless defined at $s=0$. For $s\ge d$ also put \[ \Gamma(s):=\frac12\log d+\frac12\int_d^s\frac{\dd t}{\ell(t)},\qquad \widetilde\Gamma:=\Gamma(s_0)-\frac12\log\ell_0 . \] Finally $Y_s:=(X_{q(s)}-\kappa)/e$ for $s\in[d,e^{-2}]$, and $Y_0:=Y_{s_0}$. We call $\gamma$ (or $\zeta$) a \emph{step profile} if it takes finitely many values. \end{definition} Levels $s\in[d,e^{-2}]$ correspond to $q\in[0,Q]$. In rescaled variables $v=\lambda c$ becomes $v=-\ell U_{yy}$, and the harmonic function $\psi=\lambda b+x-\kappa$ of \eqref{II:fun:psi} becomes $\Psi=\psi/e$. The rescaled profile $\zeta$ is nonincreasing and left-continuous, $\zeta=1$ on $[0,d]$, $\zeta<1$ on $(d,e^{-2}]$ (because $Q\in\supp\mu_\gamma$), $\zeta>0$ on $[0,s_0)$ (because $q_0\in\supp\mu_\gamma$) and $\zeta=0$ on $(s_0,e^{-2}]$; also $\zeta(d+)=\gamma(Q-)$. Hence $\ell$ is concave and nondecreasing, $\ell(s)=s$ on $[0,d]$, $\ell\le s$, and $d\le\ell_0\le s_0$. Since $(\log\ell)'=\zeta/\ell\le1/\ell$ on $[d,s_0]$, \begin{equation}\label{II:ap:eq-Gammatilde} \widetilde\Gamma=\frac12\log\frac d{\ell_0}+\frac12\int_d^{s_0}\frac{\dd t}{\ell(t)}\ \ge\ 0 . \end{equation} \begin{lemma}\label{II:ap:lem-rescaled} Let $\gamma\in\cU$, $\kappa\in\R$, and $e>0$. On $[d,e^{-2}]\times\R$: \begin{enumerate}[label=\textup{(\alph*)}] \item $U\le0$, $U_y>0$, $00$, and $\Psi_y=1-v$; $\Psi(s,\cdot)$ is convex and increasing, and $\ell U_y$ is decreasing in $y$, with $(\ell U_y)_y=-v$. At the top level, \[ U(d,y)=\log\Phi(y/\sqrt d),\qquad U_y(d,y)=R(y/\sqrt d)/\sqrt d,\qquad \Psi(d,y)=\sqrt d\,\Rbar(y/\sqrt d). \] On every open interval of levels on which $\zeta\equiv m$, $U$ is smooth and $U_s=\frac12\bigl(U_{yy}+mU_y^2\bigr)$. \item \emph{The sandwich.} \begin{equation}\label{II:ap:eq-P} v\le U_y\Psi\le\frac s{\ell(s)}\,v ; \end{equation} equivalently, $F\ge0$ and $v\le U_y\Psi$. \item For $y<0$: $U_y(s,y)\ge|y|/\ell(s)$ and $\Psi(s,y)\le\min\{\sqrt s,s/|y|\}$. For $y>0$: $y<\Psi(s,y)\le y+\min\{\sqrt s,s/y\}$. \item Let $q=q(s)$, $x=\kappa+ey$, and let $X^{q,x}$ be the diffusion \eqref{II:fun:X} started from $x$ at time $q$. Then \[ \Psi(s,y)=\E\Bigl[\sqrt d\,\Rbar\Bigl(\frac{X^{q,x}_Q-\kappa}{e\sqrt d}\Bigr)\Bigr]\ \ge\ \max\Bigl\{\E\bigl(y+\sqrt{s-d}\,G\bigr)_+,\ \sqrt d\,\Rbar\bigl(y/\sqrt d\bigr)\Bigr\}. \] \end{enumerate} \end{lemma} \begin{proof} Under the change of variables, $b=U_y/e$, $c=-U_{yy}/e^2$, $\lambda=e^2\ell$, $\lambda c=v$, $\psi=e\Psi$, $\partial_x=e^{-1}\partial_y$, and $(1-q)/\lambda=s/\ell$. (a) The signs are \Cref{II:lem:regularity}(b),(c) and \Cref{II:lem:generator}(b), and $(\ell U_y)_y=\ell U_{yy}=-v$. The top formulas follow from \eqref{II:fun:top}, $R'=-V$, and $\Rbar=z+R$. Since $\partial_s=-e^2\partial_q$, the equation $u_q=-\frac12(u_{xx}+\gamma u_x^2)$ becomes $U_s=\frac12(U_{yy}+\zeta U_y^2)$ (\Cref{II:lem:regularity}(a)). (b) The sandwich $v\le b\psi\le\frac{1-q}\lambda v$ of \Cref{II:lem:sandwich}(a) becomes \eqref{II:ap:eq-P}. Since $s(-U_{yy})=\frac s\ell v$, the upper half is $F\ge0$. (c) Let $y<0$. Since $\Psi>0$, $\ell U_y=\Psi-y>|y|$. By \eqref{II:ap:eq-P} and $v<1$, $\Psi\le s/(\ell U_y)0$. Then $\Psi>y$ because $U_y>0$, so $\ell U_y\le s/\Psiz$, $\Rbar>0$, and $\Rbar$ is increasing, $\sqrt d\,\Rbar(\Xi/\sqrt d)\ge\Xi_+\ge(y+(B_Q-B_q)/e)_+$, which gives the first lower bound. Since $\Rbar$ is convex and increasing (\Cref{II:fun:gauss}(a)) and $\E\Xi\ge y$, Jensen's inequality gives the second. \end{proof} \section{The value formula and the exact decomposition}\label{II:ap:sec-decomp} Below its bottom $q_0$ the order parameter vanishes, so the diffusion $X$ is a Brownian motion there and $u$ solves the heat equation. This makes the input of the rescaled problem an explicit Gaussian. \begin{lemma}\label{II:lem:valueformula} Let $\alpha>0$, let $\gamma\in\cU$ with $q_0>0$, let $\kappa\in\R$ and $e>0$. Then $Y_0=(B_{q_0}-\kappa)/e$ has the Gaussian law $N(-\kappa/e,\,q_0/e^2)$, with density \[ p_0(y)=\frac e{\sqrt{q_0}}\,\phi\Bigl(\frac{ey+\kappa}{\sqrt{q_0}}\Bigr)\le\frac{e\phi_0}{\sqrt{q_0}}, \] and \begin{equation}\label{II:ap:eq-value} \cP_\kappa(\gamma)=\alpha\E U(s_0,Y_0)+\Gamma(s_0)+\frac{q_0}{2e^2\ell_0}+\log e . \end{equation} \end{lemma} \begin{proof} Since $\gamma=0$ on $[0,q_0)$, the drift of $X$ vanishes there, so $X_{q_0}=B_{q_0}$; and by \Cref{II:fun:colehopf}(a) with $m=0$, $u(0,0)=\E u(q_0,B_{q_0})=\E U(s_0,Y_0)$. The entropy is $\CS(\gamma)=\frac12\int_0^Q\lambda^{-1}+\frac12\log\delta$. On $[0,q_0]$ we have $\lambda=\lambda(0)=e^2\ell_0$, so $\frac12\int_0^{q_0}\lambda^{-1}=\frac{q_0}{2e^2\ell_0}$. Substituting $q=1-e^2t$ on $[q_0,Q]$ gives $\frac12\int_{q_0}^Q\lambda^{-1}\dd q=\frac12\int_d^{s_0}\ell^{-1}\dd t$. Finally $\frac12\log\delta=\log e+\frac12\log d$. \end{proof} The proof uses only that $\gamma=0$ on $[0,q_0)$. Hence \Cref{II:lem:valueformula}, and with it \Cref{II:prop:decomp} below, remains valid verbatim if $q_0$ is replaced by any level $q_1\in(0,q_0]$, with $s_0$ replaced by $s_1=(1-q_1)/e^2$ and $Y_0$ by $(B_{q_1}-\kappa)/e$; the mass is unchanged, $\ell(s_1)=\ell_0$, because $\zeta=0$ on $(s_0,s_1]$, and \eqref{II:ap:eq-Gammatilde} continues to hold with $s_1$ in place of $s_0$. The next lemma records the tails of $U$. In the far left, the terminal datum $\log\Phi(y/\sqrt d)$ behaves like $-y^2/(2d)-\log|y|$. Below the top the quadratic coefficient becomes $1/(2\ell(s))$, the logarithm keeps coefficient one, and the constant term records the entropy $\Gamma$. This is \Cref{II:lem:regularity}(b),(d),(e) in rescaled variables. \begin{lemma}\label{II:ap:lem-asym} Let $\gamma\in\cU$, let $\kappa\in\R$ and $e>0$, and let $r\in(Q,1)$. There are $C,Y_1<\infty$, depending only on $e$, $|\kappa|$, and $r$, such that for $s\in[d,e^{-2}]$: \begin{enumerate}[label=\textup{(\alph*)}] \item $0\ge U(s,y)\ge-C(1+y^2)$ for all $y\in\R$, and $U(s,y)\ge-C\,\e^{-y^2/(8s)}$ for $y\ge0$; \item with $\omega(s):=\log\ell(s)-\Gamma(s)-\log\sqrt{2\pi}$, \[ \Bigl|U(s,y)+\frac{y^2}{2\ell(s)}+\log|y|-\omega(s)\Bigr|\le\frac{C}{y^2}\qquad(y\le-Y_1). \] \end{enumerate} \end{lemma} \begin{proof} Put $q=1-e^2s\in[0,Q]$ and $x=\kappa+ey$, and apply \Cref{II:lem:regularity}(b),(d),(e) with the cutoff $r$ and $K_0=|\kappa|$. Part (a) follows from $u\le0$, the quadratic lower bound of \Cref{II:lem:regularity}(b), \Cref{II:lem:regularity}(d), and $(x-\kappa)^2/(8(1-q))=y^2/(8s)$. For (b), we have $(x-\kappa)^2/(2\lambda(q))=y^2/(2\ell(s))$ and $\log(\kappa-x)=\log e+\log|y|$. Since $\lambda=e^2\ell$, $\delta=e^2d$, and the substitution $t=1-e^2t'$ gives $\int_q^Q\lambda(t)^{-1}\dd t=\int_d^s\ell(t')^{-1}\dd t'$, the constant \eqref{II:fun:omega} of \Cref{II:lem:regularity}(e) equals \[ \log\bigl(e^2\ell(s)\bigr)-\tfrac12\log\bigl(e^2d\bigr)-\tfrac12\int_d^s\frac{\dd t'}{\ell(t')}-\log\sqrt{2\pi}=\log e+\omega(s). \] The terms $\log e$ cancel, and the error $C/(\kappa-x)^2$ equals $Ce^{-2}/y^2$. \end{proof} We now split $\cP_\kappa(\gamma)$. The explicit left asymptotics of \Cref{II:ap:lem-asym} can be integrated against the Gaussian law exactly. What remains is a function $W$ that is integrable on $\R$. The estimate $|\alpha\E W(Y_0)|\le\alpha e\phi_0q_0^{-1/2}\|W\|_{L^1}$ makes this contribution $O(e)$ on classes where $q_0$ is bounded away from zero and $\|W\|_{L^1}$ is uniformly bounded. \begin{proposition}\label{II:prop:decomp} Let $\alpha>0$, let $\gamma\in\cU$ with $q_0>0$, let $\kappa\in\R$ and $e>0$. Define \[ \begin{gathered} k_0:=\frac\kappa{\sqrt{q_0}},\qquad \rho:=\alpha\Phi(k_0),\qquad c_e(q_0):=\frac{q_0\bigl(1-\alpha m(k_0)\bigr)}{2e^2},\\ \omega(s_0):=\log\ell_0-\Gamma(s_0)-\log\sqrt{2\pi}, \end{gathered} \] \[ K_e(q_0):=\log e-\alpha\E\bigl[\log|Y_0|;\,Y_0\le-1\bigr]-\rho\log\sqrt{2\pi}, \] \[ W(y):=U(s_0,y)+\frac{y_-^2}{2\ell_0}-\omega(s_0)\1_{\{y\le0\}}+\log|y|\,\1_{\{y\le-1\}}, \] and the profile-independent constant $K_e:=\log e-\alpha\E[\log|Y^*|;Y^*\le-1]-\alpha\Phi(\kappa)\log\sqrt{2\pi}$, where $Y^*:=(G-\kappa)/e$; this is $K_e(q_0)$ at $q_0=1$. \begin{enumerate}[label=\textup{(\alph*)}] \item $\E|W(Y_0)|<\infty$ and \begin{equation}\label{II:ap:eq-decomp} \cP_\kappa(\gamma)=K_e(q_0)+\frac{c_e(q_0)}{\ell_0}+\frac{1+\rho}2\log\ell_0+(1-\rho)\widetilde\Gamma+\alpha\E W(Y_0). \end{equation} Here $K_e(q_0)$, $c_e(q_0)$, and $\rho$ depend on $\gamma$ only through $q_0$, and $\widetilde\Gamma\ge0$. \item $W$ is bounded on compact sets, $W(y)=O(y^{-2})$ as $y\to-\infty$, and $W(y)=U(s_0,y)\in[-C_\gamma\e^{-y^2/(8s_0)},0]$ for $y>0$. In particular $W\in L^1(\R)$. The constants here depend on $\gamma$ (through a cutoff $r\in(Q,1)$) and on $e$; no bound on $\|W\|_{L^1}$ uniform in $e$ is claimed at this stage. \end{enumerate} \end{proposition} \begin{proof} (a) Put $\mathcal E(y):=-\frac{y_-^2}{2\ell_0}+\omega(s_0)\1_{\{y\le0\}}-\log|y|\1_{\{y\le-1\}}$, so that $U(s_0,\cdot)=\mathcal E+W$. Both $U(s_0,\cdot)$ (by \Cref{II:ap:lem-asym}(a)) and $\mathcal E$ have at most quadratic growth, so $W(Y_0)$ is integrable. Since $e(Y_0)_-=(\kappa-B_{q_0})_+$ and $\Prob(Y_0\le0)=\Phi(k_0)$, \[ \alpha\E\mathcal E(Y_0)=-\frac{\alpha q_0m(k_0)}{2e^2\ell_0}+\rho\,\omega(s_0)-\alpha\E\bigl[\log|Y_0|;Y_0\le-1\bigr]. \] Add $\Gamma(s_0)+\frac{q_0}{2e^2\ell_0}+\log e$ as in \eqref{II:ap:eq-value}. The first term and $\frac{q_0}{2e^2\ell_0}$ combine to $c_e(q_0)/\ell_0$. Using $\omega(s_0)=\log\ell_0-\Gamma(s_0)-\log\sqrt{2\pi}$ and $\Gamma(s_0)=\widetilde\Gamma+\frac12\log\ell_0$, \begin{align*} \rho\,\omega(s_0)+\Gamma(s_0)&=\rho\log\ell_0+(1-\rho)\Bigl(\widetilde\Gamma+\frac12\log\ell_0\Bigr)-\rho\log\sqrt{2\pi}\\ &=\frac{1+\rho}2\log\ell_0+(1-\rho)\widetilde\Gamma-\rho\log\sqrt{2\pi}. \end{align*} Collecting the remaining terms gives $K_e(q_0)$. Nonnegativity of $\widetilde\Gamma$ is \eqref{II:ap:eq-Gammatilde}. (b) $U(s_0,\cdot)$ is continuous (\Cref{II:lem:regularity}(a)), and $\mathcal E$ is bounded on compact sets. The left tail is \Cref{II:ap:lem-asym}(b) at $s=s_0$, and the right tail is \Cref{II:ap:lem-asym}(a). \end{proof} The coefficients in \eqref{II:ap:eq-decomp} depend on $\gamma$ only through $q_0=1-e^2s_0$. The next lemma expands them. The main term of $c_e$ is $2\phi_0\theta$, coming from $\kRS-\kappa=e^2\theta$ and $\alpha m'(\kRS)\approx4\phi_0$; the correction $es_0/8$ comes from the shift of the effective margin $\kappa/\sqrt{q_0}$ with depth. \begin{lemma}\label{II:lem:level0} There are $e_{\rm L}>0$ and $C_{\rm L}<\infty$, depending only on $\Theta$, such that for $00$, $\kappa\in\R$, and $00$ and $C_{\rm RS}<\infty$, depending only on $\Theta$, such that for $00$ because $\Psi>0$ (\Cref{II:ap:lem-rescaled}(a)). Two further facts are used in (b) and (c). By \Cref{II:lem:valueformula}, $Y_0$ has density $p_0\le e\phi_0/\sqrt{q_0}$. By \Cref{II:lem:level0}(a),(c), $-\frac e2\le\kappa<0$ and \begin{equation}\label{II:ap:eq-ce} 2c_e(q_0)\le5\phi_0\theta_++\tfrac52\phi_0\,e\,s_0 . \end{equation} (b) Suppose that $s_0>\max\{2d,C_2\}$; we show $s_0\le C_1e^{-2/3}$. Put $\sigma_0:=\sqrt{s_0-d}$ and $y_1:=\frac12\phi_0\sigma_0$. Since $s_0>2d$, $\sigma_0^2\ge s_0/2$. For $y\ge0$, \Cref{II:ap:lem-rescaled}(d) and the monotonicity of $y\mapsto\E(y+\sigma_0G)_+$ give $\Psi(s_0,y)\ge\E(\sigma_0G)_+=\phi_0\sigma_0$; hence $g(y)=(\Psi-y)^2\ge\frac14\phi_0^2\sigma_0^2$ for $0\le y\le y_1$. Put $x_1:=ey_1=\frac12e\phi_0\sigma_0$. Then $x_1\le\frac12\phi_0<1$ because $e^2s_0=1-q_0<1$. Also $c_G<\phi_0^3$ gives $C_2>(5/\phi_0^2)^2>16\pi$, so $\sigma_0^2\ge s_0/2>8\pi=4/\phi_0^2$ and $x_1\ge e\ge2|\kappa|$. As $\kappa<0<\kappa+x_1\le1$ and $q_0\le1$, \begin{align*} \Prob(0\le Y_0\le y_1)&=\Prob\bigl(\kappa\le\sqrt{q_0}\,G\le\kappa+x_1\bigr)\ge\Prob(0\le G\le\kappa+x_1)\\ &\ge\phi(1)(\kappa+x_1)\ge\tfrac12\phi(1)x_1 . \end{align*} By (a), $g\ge0$ and $\alpha\ge2$, \[ 2c_e(q_0)\ge\alpha\E\bigl[g(Y_0);0\le Y_0\le y_1\bigr]\ge2\cdot\frac{\phi_0^2\sigma_0^2}4\cdot\frac{\phi(1)\phi_0e\sigma_0}4=\frac{\phi(1)\phi_0^3}8\,e\,\sigma_0^3\ge c_G\,e\,s_0^{3/2}. \] Since $s_0>C_2$, we have $\frac52\phi_0es_0<\frac12c_Ges_0^{3/2}$, so \eqref{II:ap:eq-ce} gives $\frac12c_Ges_0^{3/2}<5\phi_0\theta_+$, that is, $s_0\frac38$. Let $y<0$. The lower half of the sandwich \eqref{II:ap:eq-P} at level $s_0$ gives $(\Psi+|y|)\Psi=\ell_0U_y\Psi\ge\ell_0v$, so \[ g(y)=2(\Psi+|y|)\Psi-\Psi^2\ge2\ell_0-2\ell_0(1-v)-\Psi^2 . \] By \Cref{II:ap:lem-rescaled}(a),(c), $\Psi(s_0,\cdot)$ is increasing with $\Psi_y=1-v$ and $\Psi(s_0,y)\le\min\{\sqrt{s_0},s_0/|y|\}$ for $y<0$. Hence \[ \begin{gathered} \int_{-\infty}^0(1-v)(s_0,y)\dd y=\Psi(s_0,0)\le\sqrt{s_0},\\ \int_{-\infty}^0\Psi(s_0,y)^2\dd y\le\int_0^\infty\min\Bigl\{s_0,\frac{s_0^2}{u^2}\Bigr\}\dd u=2s_0^{3/2}. \end{gathered} \] Using $g\ge0$ on $[0,\infty)$, $\Prob(Y_0<0)=\Phi(k_0)$, and $p_0\le\sqrt2\phi_0e$, \[ 2c_e(q_0)=\alpha\E g(Y_0)\ge\alpha\Bigl[2\ell_0\bigl(\Phi(k_0)-\sqrt2\phi_0e\sqrt{s_0}\bigr)-2\sqrt2\phi_0e\,s_0^{3/2}\Bigr]\ge\ell_0-6\sqrt2\phi_0e\,s_0^{3/2}, \] where we used $\Phi(k_0)-\sqrt2\phi_0e\sqrt{s_0}\ge\frac14$ and $2\le\alpha\le3$. Combining with \eqref{II:ap:eq-ce}, and using $\frac52\phi_0<1$ and $6\sqrt2\phi_0<4$, gives (c). \end{proof} Only the stationarity $D(q_0)=0$ of the bottom enters \Cref{II:lem:bottom}, not global minimality. The two parts use disjoint regions of the Gaussian input. On $\{Y_0<0\}$ the sandwich makes $g\ge2\ell_0v-\Psi^2$, which forces $\ell_0\lesssim c_e(q_0)$. On $\{0\le Y_0\lesssim\sqrt{s_0}\}$, heat smoothing over the levels $[d,s_0]$ makes $g$ of order $s_0$ on a set of probability of order $e\sqrt{s_0}$; this gain of order $es_0^{3/2}$ cannot be balanced by $2c_e(q_0)=O(1+es_0)$ at large depth. \section{A priori bounds for global minimizers}\label{II:ap:sec-bounds} The top gap is bounded below by comparison with the certified margin $k_*\le\kc$ of \Cref{II:thm:ordering}. At a distance $\eta\asymp e^2$ below $k_*$, the slack bound of \Cref{II:fun:slack} makes $\cP_\kappa(\gamma)$ large when $\delta\ll\eta^2$, unless $\lambda(0)\gtrsim\eta$; weak top stability (\Cref{II:lem:weaktop}) forces $\lambda(0)\lesssim\sqrt\delta$ for minimizers; and the value inequality of \Cref{II:lem:value} closes the argument. \begin{theorem}\label{II:thm:gapD} There are $e_{\rm D}>0$, depending only on $\Theta$, and $c_{\rm D}>0$, depending only on $\theta_-$, such that for $0\eta^2/9$, then $\delta\ge\theta_-^2e^4/36$. Assume from now on that $\delta\le\eta^2/9$, so that $Q\ge\frac{35}{36}$. We divide the rest of the proof into three steps. \proofstep{1}{A lower bound on $\lambda(0)$.} Apply \Cref{II:fun:slack} with the margin shift $\eta/3\ge\sqrt\delta$ and with $k:=k_*-\eta/3<\kc$. Its hypothesis holds: $\kappa+\eta/3=k_*-\frac23\eta<0$ and $Q\le1$, so $(\kappa+\eta/3)/\sqrt Q\le\kappa+\eta/3\le k$. Since $k_*\le\kc$ and $m_1>0$ is increasing, with $k\ge\kappa\ge-1$, the constant of \Cref{II:fun:slack} satisfies $A(k)\ge\alpha\int_k^{k_*}m_1(t)\dd t\ge\alpha m_1(-1)\eta/3$. Hence \[ \cP_\kappa(\gamma)\ge\frac{Q\,\alpha\,m_1(-1)\,\eta}{3\lambda(0)}+\frac12\log\delta-2\alpha . \] Together with $\cP_\kappa(\gamma)<\frac12\log\delta$ (\Cref{II:lem:value}) and $Q\ge\frac12$, this gives $\lambda(0)>m_1(-1)\eta/12$. \proofstep{2}{An upper bound on $\lambda(0)$.} A global minimizer minimizes $\cP_\kappa$ over $\cU_r$ for every $r\in(Q,1)$ (\Cref{II:prop:kkt}(c)), so \eqref{II:fun:lambdatop} of \Cref{II:lem:weaktop} gives $\lambda(0)\le\max\{1,2\alpha C_\kappa/s_\alpha\}\sqrt\delta\le C_\lambda\sqrt\delta$ with the numerical constant $C_\lambda:=288$. Indeed $\alpha\le3$, $C_\kappa\le1$ for $\kappa\le0$, and $s_\alpha=\min\{\frac12,\frac1{16\alpha}\}=\frac1{16\alpha}$ for $\alpha\ge2$, so that $2\alpha C_\kappa/s_\alpha\le32\alpha^2\le288$. \proofstep{3}{Conclusion.} By Steps 1 and 2, $\sqrt\delta\ge m_1(-1)\eta/(12C_\lambda)\ge m_1(-1)\theta_-e^2/(24C_\lambda)$. Both cases give $\delta\ge c_{\rm D}e^4$ with $c_{\rm D}:=(m_1(-1)\theta_-/(24C_\lambda))^2$, which is smaller than $\theta_-^2/36$ because $m_1(-1)<10$ and $\Dtop,L,C_0<\infty$, depending only on $\Theta$, such that for $00 ; \end{equation} it is positive by the Mills-ratio bound $\Phi(-3)<\phi(3)/3$ of \Cref{II:fun:gauss}(b). \begin{theorem}\label{II:thm:depth} There is $e_{\rm S}>0$, depending only on $\Theta$, such that for $00$, $\kappa\in\R$, and let $\gamma$ be a KKT point of $\cP_\kappa$ in the sense of \Cref{II:prop:kkt}(c); for instance, let $\gamma$ be a global minimizer. Then $\lambda D$ is absolutely continuous on $[0,r]$ for every $r<1$, and \begin{equation}\label{II:dep:IF} (\lambda D)'=\chi(0)-\frac w\lambda\qquad\text{almost everywhere on }[0,1). \end{equation} Consequently $\int_{q_0}^Qw/\lambda\dd q=\chi(0)(Q-q_0)$. If $\alpha=2+e$ and $\widetilde\chi:=e^2\chi(0)$, then \begin{equation}\label{II:dep:chitilde} \widetilde\chi=\frac{\alpha\E v(s_0,Y_0)-1}{\ell_0} \end{equation} and \begin{equation}\label{II:dep:IF-int} \int_d^{s_0}\frac{w(s)}{\ell(s)}\dd s=\widetilde\chi\,(s_0-d). \end{equation} \end{lemma} \begin{proof} The function $\lambda$ is Lipschitz with $\lambda'=-\gamma$ almost everywhere, and $D\in C^1([0,1))$ with $D'=(S-1)/\lambda^2$ by \Cref{II:lem:generator}(c)(i). Hence $\lambda D$ is absolutely continuous on $[0,r]$, and almost everywhere \[ (\lambda D)'=-\gamma D+\frac{S-1}{\lambda}=-\gamma D+\frac{\lambda\chi-w}{\lambda}=-\gamma D+\chi(0)+\gamma D-\frac w\lambda , \] by the identity $S-1=\lambda\chi-w$ of \Cref{II:lem:generator}(c)(v) and by $\chi=\chi(0)+\gamma D$ from \Cref{II:prop:kkt}(d). The points $q_0$ and $Q$ lie in $\supp\mu_\gamma$, so $D(q_0)=D(Q)=0$ by \Cref{II:prop:kkt}(c), and integrating \eqref{II:dep:IF} over $[q_0,Q]$ gives the integrated identity. By \Cref{II:prop:kkt}(c),(d), $\chi$ is constant on $[0,q_0]$, $\lambda=\lambda(0)=e^2\ell_0$ there, and $X_{q_0}=B_{q_0}$; hence \[ \chi(0)=\chi(q_0)=\frac{\alpha\E v(q_0,X_{q_0})-1}{\lambda(0)}, \] and $v(q_0,X_{q_0})=v(s_0,Y_0)$ gives \eqref{II:dep:chitilde}. The substitution $q=1-e^2s$, $\lambda=e^2\ell$ has $\dd q=-e^2\dd s$ and maps $q_0,Q$ to $s_0,d$; reversing the limits gives \eqref{II:dep:IF-int}. \end{proof} The identity uses $D=0$ at every support point, through $\chi=\chi(0)+\gamma D$; the endpoint conditions $D(q_0)=D(Q)=0$ alone would not suffice. \begin{remark}\label{II:dep:rem-weighted} The same argument gives a family of weighted balance laws, which we do not use below. Let $\varphi:[0,1]\to\R$ be continuous and monotone. The measure $\dd(\varphi\circ\gamma)$ is carried by $\supp\mu_\gamma$, because $\gamma$ is constant on each component of the complement, and $\lambda D$ vanishes on $\supp\mu_\gamma$. Integration by parts therefore gives \[ \int_{q_0}^Q\varphi(\gamma(q))\Bigl(\chi(0)-\frac{w(q)}{\lambda(q)}\Bigr)\dd q=-\int_{(q_0,Q]}\lambda D\dd(\varphi\circ\gamma)=0 . \] The weight $\varphi\equiv1$ is \Cref{II:lem:if}. The weight $\varphi(t)=1-t$, together with the representation of $b\psi-v$ in \Cref{II:lem:sandwich}(a), the identity $c(1-v)=v(1-v)/\lambda$, and the Markov property of $X$, gives the identity $s_0/\ell_0-1=\alpha\E F(s_0,Y_0)$ at the bottom. It involves only the solution at the level $s_0$ and the Gaussian law of $Y_0$, and it expresses stationarity under dilations. For $c$ close to one let $\gamma_c\in\cU$ have rescaled profile $\zeta(\cdot/c)$. Its rescaled solution at level $cs$ is $U(s,y/\sqrt c)$, by the Cole--Hopf recursion of \Cref{II:fun:colehopf}(a), Brownian scaling, and approximation in $L^1$ (\Cref{II:lem:regularity}(f)). Differentiating \eqref{II:ap:eq-value} for $\gamma_c$ at $c=1$ gives \[ \frac{\dd}{\dd c}\cP_\kappa(\gamma_c)\Big|_{c=1}=\frac12\Bigl(\alpha\E F(s_0,Y_0)-\frac{s_0}{\ell_0}+1\Bigr)+\frac12\lambda(0)D(q_0). \] Thus, when $D(q_0)=0$, stationarity under dilations of the rescaled profile carries no information beyond the weighted balance laws. The dilation $\gamma\mapsto\gamma_c$ pushes $\mu_\gamma$ forward by the affine map $q\mapsto1-c(1-q)$, so it is a transport variation of the kind used by Talagrand \cite[Lemma~3.7]{TalagrandParisiMeasures}, and by Auffinger and W.-K.\ Chen to derive the conditions on the support of Parisi measures \cite[Section~3]{AuffingerChenProperties}. \end{remark} \section{Signed-area cancellation}\label{II:dep:sec-cancel} The bottom scalar $\widetilde\chi$ of \eqref{II:dep:chitilde} is a Gaussian average of $v(s_0,\cdot)$. At every level, $v$ passes from one to zero around $y=0$, and the excess of $\Psi$ over its asymptote measures the signed area between $v$ and the step function $\1_{\{y<0\}}$. At a level $s\in[d,e^{-2}]$ put \begin{equation}\label{II:dep:excess} \mathfrak o_s(y):=\Psi(s,y)-y_+=\begin{cases}\Psi(s,y),&y\le0,\\ \ell(s)U_y(s,y),&y>0,\end{cases}, \end{equation} the two expressions agreeing at $y=0$ because $\Psi=\ell U_y+y$. \begin{lemma}\label{II:lem:cancel} Let $e>0$, $\kappa\in\R$, $\gamma\in\cU$, and $s\in[d,e^{-2}]$, and write $v=v(s,\cdot)$. The function $\mathfrak o_s$ is Lipschitz, tends to zero at $\pm\infty$, and satisfies \[ \mathfrak o_s'=\1_{\{y<0\}}-v\quad\text{almost everywhere},\qquad 0<\mathfrak o_s\le\sqrt s,\qquad |y|\,\mathfrak o_s(y)\le s\quad(y\in\R). \] In particular $v-\1_{\{y<0\}}$ is integrable, and \begin{equation}\label{II:dep:Z} \int_{-\infty}^0\bigl(1-v(y)\bigr)\dd y=\Psi(s,0)=\int_0^\infty v(y)\dd y, \qquad \int_\R\bigl(v(y)-\1_{\{y<0\}}\bigr)\dd y=0. \end{equation} For every Gaussian random variable $Y'$ with mean $\bar y$ and variance $\sigma^2>0$, \begin{equation}\label{II:dep:gauss-cancel} \E v(Y')-\Prob(Y'<0) =-\frac{\E[\mathfrak o_s(Y')(Y'-\bar y)]}{\sigma^2},\qquad \bigl|\E v(Y')-\Prob(Y'<0)\bigr|\le\frac{s+|\bar y|\sqrt s}{\sigma^2}. \end{equation} \end{lemma} \begin{proof} By \Cref{II:ap:lem-rescaled}(a), $\Psi_y=1-v$ and $(\ell U_y)_y=-v$, with $00$ we have $0<\ell U_y(s,y)=\Psi(s,y)-y\le\min\{\sqrt s,s/y\}$. These are the two bounds, and they give the zero limits at $\pm\infty$. Integrating $\mathfrak o_s'$ over each half-line gives \eqref{II:dep:Z}. If $p_{Y'}$ is the density of $Y'$, then $p_{Y'}'=-(y-\bar y)p_{Y'}/\sigma^2$ and $v-\1_{\{y<0\}}=-\mathfrak o_s'$. Integration by parts, in which the boundary terms vanish because $\mathfrak o_s$ is bounded and $p_{Y'}$ decays, gives the identity in \eqref{II:dep:gauss-cancel}. The bounds $|y|\mathfrak o_s\le s$ and $\mathfrak o_s\le\sqrt s$ give $|\E[\mathfrak o_s(Y')(Y'-\bar y)]|\le s+|\bar y|\sqrt s$. \end{proof} The cancellation \eqref{II:dep:Z} is essential below. The cruder bound $\|p_{s_0}\|_\infty\int_\R|v-\1_{\{y<0\}}|\dd y=2\|p_{s_0}\|_\infty\Psi(s_0,0)$ is of order $e\sqrt{s_0}$, which is the same order as the width contribution in \eqref{II:dep:IF-int}. \section{Proof of the bounded-depth theorem}\label{II:dep:sec-proof} \begin{proof}[Proof of \Cref{II:thm:depth}] Fix a global minimizer $\gamma$. By \Cref{II:thm:upper} and \Cref{II:lem:level0}(a), for $e$ small enough, depending only on $\Theta$, \[ q_0\ge\tfrac12,\qquad s_0\le C_0e^{-2/3},\qquad \ell_0\le L, \qquad -\tfrac e2\le\kappa<0. \] If $\gamma$ is replica symmetric, then $s_0=d=\ell_0\le L\le S_*$. Assume from now on that $s_0>d$. We divide the rest of the proof into three steps. \proofstep{1}{The bottom scalar.} By \Cref{II:lem:valueformula}, $Y_0$ is Gaussian with mean $-\kappa/e\in(0,\frac12]$ and variance $q_0/e^2\ge1/(2e^2)$, and $\Prob(Y_0<0)=\Phi(\kappa/\sqrt{q_0})$. Applying \eqref{II:dep:gauss-cancel} at the level $s_0$ and using $\alpha\le3$ gives \[ \bigl|\alpha\E v(s_0,Y_0)-\alpha\Phi(\kappa/\sqrt{q_0})\bigr|\le6e^2\bigl(s_0+\tfrac12\sqrt{s_0}\bigr)\le Ce^2(1+s_0). \] Combined with the expansion $\alpha\Phi(\kappa/\sqrt{q_0})-1=\frac e4+O(e^2+e^3s_0)$ of \Cref{II:lem:level0}(b), \eqref{II:dep:chitilde} yields \begin{equation}\label{II:dep:chi-bound} \widetilde\chi\ell_0=\frac e4+O\bigl(e^2(1+s_0)\bigr)\le\frac e3 \end{equation} for small $e$: since $s_0\le C_0e^{-2/3}$, the error is $O(e^2+e^{4/3})=o(e)$, uniformly in the minimizer. \proofstep{2}{A fixed positive window.} Let $s\in[d,s_0]$, $q=1-e^2s$, and $x=\kappa+ey$. By \Cref{II:ap:lem-rescaled}(d) and the pathwise comparison \eqref{II:fun:XgeB}, since $\Rbar$ is increasing, \[ \Psi(s,y)\ge\E\Bigl[\sqrt d\,\Rbar\Bigl(\frac{y+\sqrt{s-d}\,G}{\sqrt d}\Bigr)\Bigr]. \] For every $z$, $\sqrt d\,\Rbar(z/\sqrt d)\ge\sqrt d\,\Phi(z/\sqrt d)\Rbar(z/\sqrt d)=\sqrt d\,m_1(z/\sqrt d)=\E(z+\sqrt d\,G')_+$ by \Cref{II:fun:gauss}(g), where $G'$ is a standard Gaussian independent of $G$. Adding the two independent Gaussian variances, \begin{equation}\label{II:dep:heat} \Psi(s,y)\ge\E\bigl(y+\sqrt s\,G\bigr)_+=\sqrt s\,m_1\bigl(y/\sqrt s\bigr)\qquad(y\in\R). \end{equation} Since $v(s,\cdot)$ is decreasing and $\int_0^\infty v(s,y)\dd y=\Psi(s,0)\le\sqrt s$ by \eqref{II:dep:Z} and \Cref{II:ap:lem-rescaled}(c), \[ v(s,2\sqrt s)\le\frac1{2\sqrt s}\int_0^{2\sqrt s}v(s,y)\dd y\le\frac12. \] The upper half of the sandwich \eqref{II:ap:eq-P} reads $sv\ge\ell U_y\Psi=(\Psi-y)\Psi$. At $y=3\sqrt s$ we have $\Psi\ge3\sqrt s$ by \Cref{II:ap:lem-rescaled}(c), and $\Psi-3\sqrt s\ge\sqrt s\,(m_1(3)-3)=\sqrt s\,m_1(-3)$ by \eqref{II:dep:heat}, because $m_1(k)-m_1(-k)=\E(k-G)=k$. Hence \[ v(s,3\sqrt s)\ge3m_1(-3). \] Since $v$ is decreasing and $3m_1(-3)\le\frac12$, throughout the interval $[2\sqrt s,3\sqrt s]$ \[ v(s,y)\bigl(1-v(s,y)\bigr)\ge3m_1(-3)\bigl(1-3m_1(-3)\bigr)\ge\tfrac32m_1(-3). \] On this interval the density floor \eqref{II:fun:floor}, applied to the density $p^X_q$ of $X_q$, gives \[ p_s(y)=e\,p^X_q(\kappa+ey)\ge e\,\Phi(y/\sqrt s)\,\phi_q(\kappa+ey)\ge\frac{e\phi_0}{4}. \] Indeed $\Phi(y/\sqrt s)\ge\Phi(2)\ge\frac12$; and $q\ge q_0\ge\frac12$ gives $\phi_q(x)\ge\phi_0\e^{-x^2}$, while $|\kappa+ey|\le\frac e2+3e\sqrt{s_0}\le\frac e2+3\sqrt{C_0}e^{2/3}$, whose square is at most $\log2$ for small $e$, so that $\phi_q(\kappa+ey)\ge\phi_0/2$. Integrating over the interval, which has length $\sqrt s$, and using $\alpha\ge2$ proves \begin{equation}\label{II:dep:width} w(s)\ge2\cdot\tfrac32m_1(-3)\cdot\frac{e\phi_0}4\cdot\sqrt s=c_we\sqrt s\qquad(d\le s\le s_0). \end{equation} \proofstep{3}{The scalar inequality.} Since $\ell$ is nondecreasing, $\ell(s)\le\ell_0$ on $[d,s_0]$. Inserting \eqref{II:dep:chi-bound} and \eqref{II:dep:width} into \eqref{II:dep:IF-int} gives \[ \frac{c_we}{\ell_0}\cdot\frac23\bigl(s_0^{3/2}-d^{3/2}\bigr) \le\int_d^{s_0}\frac{w(s)}{\ell(s)}\dd s=\widetilde\chi(s_0-d)\le\frac e{3\ell_0}(s_0-d). \] Cancel $e/(3\ell_0)$ and use $s_0^{3/2}-d^{3/2}-\sqrt{s_0}(s_0-d)=d(\sqrt{s_0}-\sqrt d)\ge0$. Since $s_0>d$, division by $s_0-d$ gives $2c_w\sqrt{s_0}\le1$. This proves (i), and together with the replica-symmetric case it proves (ii). \end{proof} The factors $e$ and $\ell_0$ cancel in the last display, so the proof uses no lower bound on $d$ or $\ell_0$; it also uses no bound on the number of steps of $\gamma$ and no approximation by stationary step profiles. The obstacle inequality $S\le1$ is not used either: the balance law uses only $D=0$ on the support. The coarse depth bound of \Cref{II:thm:upper} is used only to make the bottom error in \eqref{II:dep:chi-bound} of order $o(e)$ and to keep the window $[2\sqrt s,3\sqrt s]$ in the region where the density floor is of order $e$. The conclusion says that the entire rescaled profile of a global minimizer is supported in $[0,S_*]$. \begin{remark}\label{II:rem:nearmin} For fixed $e$ and $\theta$, let $\gamma$ be a global minimizer and choose any $S\in(s_0,e^{-2})$. Put $q_1=1-e^2S$ and, for $0<\varepsilon<1$, let $\gamma_\varepsilon=\max\{\gamma,\varepsilon\1_{[q_1,1]}\}$. Then $\gamma_\varepsilon\in\cU$ has rescaled depth $S$, it lies in the same class $\cU_r$ as $\gamma$ for every $r\in(Q,1)$, and \[ \|\gamma_\varepsilon-\gamma\|_{L^1}\le\varepsilon e^2(S-d)\longrightarrow0\qquad(\varepsilon\downarrow0). \] By \Cref{II:lem:regularity}(f), $\cP_\kappa(\gamma_\varepsilon)\to\cP_\kappa(\gamma)=\cP_*(\kappa)$. Thus profiles of every admissible depth have values arbitrarily close to the minimum, and near-minimality of the value alone cannot bound the depth, even for fixed $e$; exact stationarity is what excludes a tail whose mass tends to zero. \end{remark} \chapter{Compactness of minimizers}\label{II:sec:compact} This chapter proves that, uniformly for $\theta\in\Theta$ and small $e$, the rescaled profile of every global minimizer lies in a fixed compact class. \Cref{II:thm:depth} bounds the depth by $s_0\le S_*$. What remains are lower bounds on the rescaled mass $\ell_0$ and on the rescaled top gap $d$, with constants that do not depend on $e$ or $\theta$. This is Step~3 of \Cref{II:sec:intro-ideas}. The main tool is the jamming identity of \Cref{II:cpt:lem-jam}, which is exact for every $\alpha>0$, every $\kappa\in\R$, and every order parameter $\gamma$ with $q_0>0$. It writes $\cP_\kappa(\gamma)$ as a replica-symmetric expression $\Pi(\ell_0;q_0)$ in the mass and the bottom, plus an integral over the \emph{jammed} levels, where $\zeta<1$, with weight $(1-\zeta)/(2\ell)$ and an explicit \emph{jamming coefficient} $\mathcal J$. Near $\alpha=2$ this coefficient is bounded below in \Cref{II:cpt:lem-J}: a jammed top is expensive. The proof of \Cref{II:thm:T1} has three steps. First, a profile of bounded depth and small mass has a value above the replica-symmetric value, so every global minimizer has $\ell_0\ge\ellmin$ (\Cref{II:prop:massLB}). Second, the first-order conditions of \Cref{II:prop:kkt} bound the mass $\gamma(Q-)$ just below the top atom of every global minimizer by a fixed $m_*<1$ (\Cref{II:thm:topjump}); the long estimates for this step are proved in \Cref{II:app:compact}. Third, a profile of bounded depth, with mass at least $\ellmin$, top jump $1-\gamma(Q-)\ge1-m_*$, and a small rescaled top gap, has a larger value than the replica-symmetric order parameter of the same mass (\Cref{II:cpt:prop-gap}). Hence $d>d_1$ for every global minimizer. The value comparisons in the first and third steps are inequalities for all profiles; only their conclusions use minimality. The third step uses the bounds of the first two as hypotheses, and the first uses the bounded depth. The constants are therefore chosen in the order \begin{equation}\label{II:cpt:eq-order} \Theta\ \longrightarrow\ (\Dtop,L,C_0,c_{\mathrm D},C'_{\mathrm{RS}})\ \longrightarrow\ S_*\ \longrightarrow\ \ellmin\ \longrightarrow\ m_*\ \longrightarrow\ d_1 , \end{equation} and each is independent of $e$ and $\theta$. The constant $m_*$ is in fact numerical (\Cref{II:thm:topjump}); it appears after $\ellmin$ in \eqref{II:cpt:eq-order} because $d_1$ depends on both. For $00$, $m_*\in(0,1)$, $d_1>0$, and $e_*>0$, depending only on $\Theta$ and chosen in the order \eqref{II:cpt:eq-order}, such that the following holds for $0d_1 . \] In particular the rescaled profile of $\gamma$ belongs to $\cK(d_1,S_*)$. \end{theorem} The proof is given in \Cref{II:cpt:ssec-T1}, after the three steps. The relation of the second step to earlier work is described in \Cref{II:cpt:ssec-top}. \section{Conventions and inputs}\label{II:cpt:ssec-conv} Throughout this chapter, except in \Cref{II:cpt:ssec-nonstep,II:cpt:ssec-jam}, $\alpha=2+e$ and $\kappa=\kappa_e(\theta)=\kRS(2+e)-e^2\theta$ with $\theta\in\Theta$, where $\Theta\subset(0,\infty)$ is a compact interval as in \Cref{II:rs:rem-Theta}; as there, $\theta_-=\min\Theta$ and $\theta_+=\max\Theta$. Unless stated otherwise, $c$ and $C$ denote positive constants that depend only on $\Theta$ and on the constants named in the statement at hand. They never depend on $e$, on $\theta$, or on the profile, and their values may change from line to line. We use the following inputs from \Cref{II:sec:apriori,II:sec:depth}. Let $e_{\mathrm I}>0$ be smaller than the thresholds of \Cref{II:thm:upper,II:thm:gapD,II:thm:depth}, of \Cref{II:lem:level0} and of \Cref{II:lem:rsvalue}(b). For $00$ (\Cref{II:thm:gapD}); \item\label{II:cpt:in-depth} $s_0\le S_*$ (\Cref{II:thm:depth}); \item\label{II:cpt:in-rs} $\cP_\kappa(\gamma)=\cP_*(\kappa)\le K_e+C'_{\mathrm{RS}}$, where $K_e$ is the profile-independent constant of \Cref{II:prop:decomp} and $C'_{\mathrm{RS}}$ is the constant of \Cref{II:lem:rsvalue}(b). \end{enumerate} Moreover $-e/2\le\kappa<0$ by \Cref{II:lem:level0}(a). The resulting bound $|\kappa|\le e$ does not involve the profile, and we use it for every profile. The constant $L$ chosen in the proof of \Cref{II:thm:upper} satisfies $L\ge5$, and $S_*\ge L$ by \Cref{II:thm:depth}; in particular $L\ge1$ and $S_*\ge1$. \paragraph{Rescaled variables.} We use the rescaled variables $s,y,\zeta,\ell,d,s_0,\ell_0,U,v,\Psi,\Gamma,\widetilde\Gamma,Y_0$ of \Cref{II:def:rescaled} and the properties of $U$, $v$, and $\Psi$ listed in \Cref{II:ap:lem-rescaled}. They are defined for every $e>0$, $\kappa\in\R$, and $\gamma\in\cU$. Recall that $\zeta$ is nonincreasing and left-continuous, that $\zeta=1$ on $[0,d]$ and $\zeta<1$ on $(d,\infty)$, that $\zeta>0$ on $[0,s_0)$ and $\zeta=0$ on $(s_0,\infty)$, and that $\zeta(d+)=\gamma(Q-)$. Hence $\ell$ is concave and nondecreasing with $\ell(0)=0$ and $\ell(d)=d$, so that $\ell\ge d$ on $[d,\infty)$ and \begin{equation}\label{II:cpt:eq-concave} \frac{\ell_0}{s_0}\,t\ \le\ \ell(t)\ \le\ t\qquad(0\le t\le s_0). \end{equation} Since $(\log\ell)'=\zeta/\ell$ on $[d,s_0]$ and $\Gamma(s)=\frac12\log d+\frac12\int_d^s\ell^{-1}$, \begin{equation}\label{II:cpt:eq-Gtilde-int} \widetilde\Gamma=\frac12\log\frac d{\ell_0}+\frac12\int_d^{s_0}\frac{\dd t}{\ell(t)}=\frac12\int_d^{s_0}\frac{1-\zeta(t)}{\ell(t)}\dd t . \end{equation} If $q_0>0$, then $Y_0$ is Gaussian with mean $-\kappa/e$ and variance $q_0/e^2$ (\Cref{II:lem:valueformula}); we denote its density by $p_0$. \paragraph{Coefficients.} The quantities $K_e(q_0)$, $c_e(q_0)$, and $\rho=\alpha\Phi(\kappa/\sqrt{q_0})$ of \Cref{II:prop:decomp} depend on the profile only through $q_0$. For $q\in(0,1]$ we write $K_e(q)$, $c_e(q)$, and $\rho_q:=\alpha\Phi(\kappa/\sqrt q)$ for the same formulas with $q$ in place of $q_0$. Let $S\ge1$. By \Cref{II:lem:level0}(b) there are $e_{\mathrm K}(S)\in(0,(2S)^{-1/2}]$ and $C'_S<\infty$, depending only on $S$ and $\Theta$, such that for $00$, $\kappa\in\R$, and $\gamma\in\cU$ with $q_0>0$, with rescaled data $(d,s_0,\ell_0)$. Let $T\subset[d,s_0]$ be a finite set containing $d$ and $s_0$. For $n\ge1$ let $\mathcal P_n$ be the partition of $[d,s_0]$ generated by $T$ and the points $d+k2^{-n}(s_0-d)$, $0\le k\le2^n$. Define $\zeta_n=1$ on $[0,d]$, $\zeta_n=0$ on $(s_0,\infty)$, and, on each cell $(t',t'']$ of $\mathcal P_n$, $\zeta_n\equiv(t''-t')^{-1}\int_{t'}^{t''}\zeta$. Let $\gamma_n(q)=\zeta_n((1-q)/e^2)$ for $q\ge q_0$ and $\gamma_n(q)=0$ for $qd$. (a) Averages of a nonincreasing function over consecutive intervals are nonincreasing, the first average is at most $\zeta(d+)$, and all averages are positive because $\zeta>0$ on $[0,s_0)$. The first average is smaller than one because $\zeta<1$ on $(d,\infty)$. Thus $\zeta_n$ is a nonincreasing left-continuous step function with $\zeta_n=1$ exactly on $[0,d]$ and $\zeta_n>0$ exactly on $[0,s_0]$. Hence $\gamma_n\in\cU$ has the same $q_0$, $d$, and $s_0$. Integrals over cells are preserved by construction. If $t\in T$ and $(t',t'']\subset[0,t]$ is a cell, then $\zeta\ge\zeta(t)$ on $(t',t'']$, so the average over this cell is at least $\zeta(t)$. (b) Both profiles equal one on $[1-e^2d,1]$. For every continuity point $s\in(d,s_0)$ of $\zeta$, the cell containing $s$ shrinks to $s$, so $\zeta_n(s)\to\zeta(s)$. A monotone function has countably many discontinuities and $0\le\zeta_n\le1$, so $\zeta_n\to\zeta$ in $L^1$. The identity for $\|\gamma_n-\gamma\|_{L^1}$ is the change of variables $q=1-e^2s$. (c) $|\ell_n(t)-\ell(t)|\le\|\zeta_n-\zeta\|_{L^1}$. For $t\ge d$ both $\ell_n$ and $\ell$ are at least $d$ on $[d,t]$, so $|\Gamma_n(t)-\Gamma(t)|\le(t-d)\sup|\ell_n-\ell|/(2d^2)$. (d) By (b) and \Cref{II:lem:regularity}(f), $\cP_\kappa$ is continuous on $\cU_r$ in $L^1$, and $u_{\gamma_n}(q,x;\kappa)\to u_\gamma(q,x;\kappa)$ for each $q\le r$ and $x\in\R$. \end{proof} \section{The jamming identity}\label{II:cpt:ssec-jam} In this section $\alpha>0$, $\kappa\in\R$, and $e>0$ are arbitrary, and the relation $\alpha=2+e$ is not used; $e$ only fixes the units of the rescaled variables. The next lemma compares the solution $U$, at every level, with the replica-symmetric solution of the same mass. For $\ell>0$ and $q\in(0,1]$ put \begin{equation}\label{II:cpt:eq-Pi} Y^{(q)}:=\frac{\sqrt q\,G-\kappa}{e},\qquad \Pi(\ell;q):=\alpha\E\log\Phi\bigl(Y^{(q)}/\sqrt\ell\bigr)+\frac{q}{2e^2\ell}+\frac12\log(e^2\ell) . \end{equation} This is the replica-symmetric expression with the mass $\ell$ and the bottom $q$ decoupled. If $e^2\ell<1$, then $\Pi(\ell;1-e^2\ell)=\PRS_\kappa(1-e^2\ell)$ is the value of the replica-symmetric order parameter $\1_{[1-e^2\ell,1]}$ of rescaled mass $\ell$, by \Cref{II:lem:rsvalue}(a). If $\gamma\in\cU$ has $q_0>0$, then $Y^{(q_0)}$ has the law of $Y_0$. The dependence on the bottom is explicit: \begin{equation}\label{II:cpt:eq-dPi} \partial_q\Pi(\ell;q)=\frac{1-\alpha\E V\bigl(Y^{(q)}/\sqrt\ell\bigr)}{2e^2\ell}\qquad(00$, $\kappa\in\R$, and $e>0$, let $\gamma\in\cU$ with $q_0>0$, and put $\mathsf w:=U-\mathsf g$ on $[d,s_0]\times\R$. For $(s,y)\in[d,s_0]\times\R$ let $\widehat Y=\widehat Y^{s,y}$ be the solution, indexed by the decreasing level $t$, of \begin{equation}\label{II:cpt:eq-Yhat} \widehat Y_t=y+\int_t^s\tfrac12\zeta(r)\bigl(U_y+\mathsf g_y\bigr)(r,\widehat Y_r)\dd r+B_{s-t},\qquad d\le t\le s , \end{equation} where $(B_\tau)_{\tau\ge0}$ is a standard Brownian motion. This equation has a unique strong solution, and the following hold. \begin{enumerate}[label=(\alph*)] \item For $s\in[d,s_0]$ and $y\in\R$, \[ \mathsf w(s,y)=\E\int_d^s\frac{1-\zeta(t)}{2\ell(t)}\Bigl(1-V\bigl(\widehat Y^{s,y}_t/\sqrt{\ell(t)}\bigr)\Bigr)\dd t\ \ge0 . \] \item $\widehat Y_t\ge y+B_{s-t}$ and $\widehat Y_t/\ell(t)\ge y/\ell(s)+N_t$ for $t\in[d,s]$, where $N_t:=\int_0^{s-t}\dd B_\tau/\ell(s-\tau)$ is a centered Gaussian variable. \item Define the \emph{jamming coefficient} \begin{equation}\label{II:cpt:eq-J} \mathcal J(t):=\alpha\int_\R p_0(y)\,\E\Bigl[1-V\bigl(\widehat Y^{s_0,y}_t/\sqrt{\ell(t)}\bigr)\Bigr]\dd y-(\alpha-1)\qquad(d\le t\le s_0). \end{equation} Then \begin{equation}\label{II:cpt:eq-jam} \cP_\kappa(\gamma)=\Pi(\ell_0;q_0)+\int_d^{s_0}\frac{1-\zeta(t)}{2\ell(t)}\,\mathcal J(t)\dd t . \end{equation} \end{enumerate} \end{lemma} \begin{proof} We divide the proof into five steps. Part (a) is first proved for step profiles and then extended to all profiles by \Cref{II:lem:nonstep} (Steps 2 and 3); part (b) is proved directly (Step 4), and part (c) follows from part (a) (Step 5). \proofstep{1}{Well-posedness and the defect of $\mathsf g$.} Write $z=y/\sqrt{\ell(s)}$. Then $\mathsf g_y=R(z)/\sqrt\ell$ and $\mathsf g_{yy}=-V(z)/\ell$, while $U_{yy}=-v/\ell$ with $00$, and let $\gamma_n$ be the step profiles of \Cref{II:lem:nonstep} with $T=\{d,s_0\}$. They have the same $q_0$, $d$, and $s_0$ as $\gamma$. We mark the objects of $\gamma_n$ by an index $n$. Fix $(s,y)$ and drive $\widehat Y^n=\widehat Y^{n,s,y}$ and $\widehat Y=\widehat Y^{s,y}$ by the same Brownian motion. Since $\widehat b_n$ is Lipschitz in $y$ with constant $1/d$, \[ |\widehat Y^n_t-\widehat Y_t|\le\frac1d\int_t^s|\widehat Y^n_r-\widehat Y_r|\dd r+\int_d^s|\widehat b_n-\widehat b|(r,\widehat Y_r)\dd r , \] and Gronwall's inequality gives $\sup_{t\in[d,s]}|\widehat Y^n_t-\widehat Y_t|\le\e^{(s-d)/d}\int_d^s|\widehat b_n-\widehat b|(r,\widehat Y_r)\dd r$. Now \[ |\widehat b_n-\widehat b|\le\tfrac12|\zeta_n-\zeta|\bigl(|U_y|+|\mathsf g_y|\bigr)+\tfrac12|U_{n,y}-U_y|+\tfrac12|\mathsf g_{n,y}-\mathsf g_y| . \] All $\gamma_n$ and $\gamma$ lie in the common class $\cU_r$ of \Cref{II:lem:nonstep}(b). By \Cref{II:lem:regularity}(b),(f), in rescaled form, $|U_y|+|\mathsf g_y|\le C(1+|y|)$ and $|U_{n,y}-U_y|\le C\|\gamma_n-\gamma\|_{L^1}(1+|y|)$ on $[d,s_0]\times\R$. Since $R$ is Lipschitz ($|R'|=V\le1$) and $\ell_n,\ell\ge d$ on $[d,s_0]$, also $|\mathsf g_{n,y}-\mathsf g_y|\le C(1+|y|)\sup|\ell_n-\ell|$. Here $C$ depends on $\gamma$, $e$, and $\kappa$, but not on $n$. Since $\widehat Y$ has moments of every order, \[ \E\sup_{t\in[d,s]}|\widehat Y^n_t-\widehat Y_t|\le C\bigl(\|\zeta_n-\zeta\|_{L^1}+\|\gamma_n-\gamma\|_{L^1}+\sup|\ell_n-\ell|\bigr)\longrightarrow0 \] by \Cref{II:lem:nonstep}(b),(c). As $V$ is bounded and continuous and $\ell_n\to\ell$ uniformly, $\E V(\widehat Y^n_t/\sqrt{\ell_n(t)})\to\E V(\widehat Y_t/\sqrt{\ell(t)})$ for every $t$. The weights satisfy $|(1-\zeta_n)/\ell_n-(1-\zeta)/\ell|\le|\zeta_n-\zeta|/d+|\ell_n-\ell|/d^2$, so they converge in $L^1(d,s_0)$, and they are bounded by $1/d$. By dominated convergence the right side of (a) for $\gamma_n$ converges to the right side for $\gamma$, while the left side $\mathsf w_n(s,y)=U_n(s,y)-\mathsf g_n(s,y)$ converges to $\mathsf w(s,y)$ by \Cref{II:lem:nonstep}(c),(d). This proves (a) for $\gamma$. \proofstep{4}{Part (b).} The drift $\widehat b$ is nonnegative because $U_y>0$ and $R>0$; this gives the first inequality. For the second, $\ell$ is absolutely continuous with $\ell'=\zeta$ almost everywhere, so It\^o's formula for the product of $\widehat Y_{s-\tau}$ with the absolutely continuous function $1/\ell(s-\tau)$ gives \[ \begin{aligned} \dd\Bigl(\frac{\widehat Y}{\ell}\Bigr)(s-\tau) &=\Bigl[\frac{\ell\,\widehat b+\zeta\widehat Y}{\ell^2}\Bigr](s-\tau)\dd\tau+\frac{\dd B_\tau}{\ell(s-\tau)},\\ \ell\,\widehat b+\zeta\widehat Y &=\tfrac12\zeta\bigl(\ell U_y+\widehat Y\bigr)+\tfrac12\zeta\bigl(\ell \mathsf g_y+\widehat Y\bigr)=\tfrac12\zeta\Psi+\tfrac12\zeta\sqrt\ell\,\Rbar(z), \end{aligned} \] with $z=\widehat Y/\sqrt\ell$. This is nonnegative because $\Psi>0$ (\Cref{II:ap:lem-rescaled}(a)) and $\Rbar>0$ (\Cref{II:fun:gauss}(a)). Integrating over $\tau\in[0,s-t]$ gives the claim. The variable $N_t$ is centered Gaussian with variance $\int_t^s\ell(r)^{-2}\dd r$. \proofstep{5}{Part (c).} By \Cref{II:lem:valueformula}, \[ \cP_\kappa(\gamma)=\alpha\E U(s_0,Y_0)+\Gamma(s_0)+\frac{q_0}{2e^2\ell_0}+\log e . \] Insert $U(s_0,\cdot)=\mathsf g(s_0,\cdot)+\mathsf w(s_0,\cdot)$. Since $Y_0$ has the law of $Y^{(q_0)}$ and $\Gamma(s_0)=\widetilde\Gamma+\frac12\log\ell_0$, the terms not involving $\mathsf w$ equal \[ \alpha\E\log\Phi\bigl(Y_0/\sqrt{\ell_0}\bigr)+\frac\alpha2\log\ell_0-(\alpha-1)\Gamma(s_0)+\frac{q_0}{2e^2\ell_0}+\log e=\Pi(\ell_0;q_0)-(\alpha-1)\widetilde\Gamma . \] By (a) with $s=s_0$, the integrand there being nonnegative, and Tonelli's theorem, \[ \alpha\int_\R p_0(y)\,\mathsf w(s_0,y)\dd y=\int_d^{s_0}\frac{1-\zeta(t)}{2\ell(t)}\bigl(\mathcal J(t)+\alpha-1\bigr)\dd t . \] Both sides are finite, since $0\le \mathsf w(s_0,\cdot)\le\widetilde\Gamma$ by (a) and \eqref{II:cpt:eq-Gtilde-int}. Together with \eqref{II:cpt:eq-Gtilde-int} for $(\alpha-1)\widetilde\Gamma$, this gives \eqref{II:cpt:eq-jam}. \end{proof} The proof of part (a) follows the argument of \cite[Lemma~14]{JagannathTobasco2016}: the difference $\mathsf w$ solves a linear equation with the averaged drift of \eqref{II:cpt:eq-Yhat}, and the Feynman--Kac formula represents $\mathsf w$ through the source term of that equation. Jagannath and Tobasco use it to compare the Parisi solutions of two order parameters; the history of the resulting Lipschitz bound is recalled after \Cref{II:lem:regularity}. Here $\mathsf g$ solves the rescaled Parisi equation only up to the defect \eqref{II:cpt:eq-defect}, which is proportional to $1-\zeta$ and is the source term, so the identity isolates the contribution of the jammed levels. The identity \eqref{II:cpt:eq-jam} is exact and holds for every order parameter with $q_0>0$; its proof uses nothing specific to $\alpha$ near two. In the original variables the weight $\frac{1-\zeta(t)}{2\ell(t)}\dd t$ is $\frac{1-\gamma(q)}{2\lambda(q)}\dd q$ on $[q_0,Q]$, so only the levels where $\gamma<1$ contribute. Each of them contributes with the sign of $\mathcal J$. Since $0<1-V<1$, the jamming coefficient takes values in $(1-\alpha,1)$, and \eqref{II:cpt:eq-jam} together with \eqref{II:cpt:eq-Gtilde-int} gives the two-sided bound \[ \Pi(\ell_0;q_0)-(\alpha-1)\widetilde\Gamma\ \le\ \cP_\kappa(\gamma)\ \le\ \Pi(\ell_0;q_0)+\widetilde\Gamma . \] Part (a) says that $U+\Gamma\ge\log\Phi(y/\sqrt\ell)+\frac12\log\ell$ at every level: the Cole--Hopf steps never lower $U+\Gamma$ below the replica-symmetric solution of the same mass. Below we use \eqref{II:cpt:eq-jam}, together with a Lipschitz bound for $\Pi$ in $q$ (\Cref{II:cpt:lem-Pi}(c)). \begin{remark}\label{II:cpt:rem-samemass} The first term $\Pi(\ell_0;q_0)$ depends on the mass and the bottom separately. The replica-symmetric order parameter $\1_{[1-\lambda(0),1]}$ of the same mass has the bottom $1-e^2\ell_0\ge q_0$, and the difference $e^2(s_0-\ell_0)=e^2\int_d^{s_0}(1-\zeta)$ is again carried by the jammed levels. Integrating \eqref{II:cpt:eq-dPi} along $\hat q(t):=1-e^2(\ell_0+t-\ell(t))$, which decreases from $1-e^2\ell_0$ at $t=d$ to $q_0$ at $t=s_0$ with $\hat q'=-e^2(1-\zeta)$, turns \eqref{II:cpt:eq-jam} into \begin{equation}\label{II:cpt:eq-samemass} \cP_\kappa(\gamma)=\PRS_\kappa\bigl(1-\lambda(0)\bigr)+\int_d^{s_0}\frac{1-\zeta(t)}{2\ell(t)}\Bigl[\mathcal J(t)-\frac{\ell(t)}{\ell_0}\Bigl(1-\alpha\E V\bigl(Y^{(\hat q(t))}/\sqrt{\ell_0}\bigr)\Bigr)\Bigr]\dd t . \end{equation} Thus the value of every order parameter is the replica-symmetric value of the same mass plus an integral over its jammed levels. This form is not used below. \end{remark} \section{The jamming coefficient near \texorpdfstring{$\alpha=2$}{alpha=2}}\label{II:cpt:ssec-J} From now on $\alpha=2+e$ again, so that $\alpha-1=1+e$ and $2\le\alpha\le3$ for $e\le1$. We first record the Gaussian input and the properties of $\Pi$ that we need. \begin{lemma}\label{II:cpt:lem-Y0} Let $00$ let $W_{\mathrm{RS},\ell}$ be the remainder of \Cref{II:lem:rsvalue}(a), with $d$ replaced by $\ell$: \begin{equation}\label{II:cpt:eq-WRS} W_{\mathrm{RS},\ell}(y):=\log\Phi\bigl(y/\sqrt\ell\bigr)+\frac{y_-^2}{2\ell}-\Bigl(\frac12\log\ell-\log\sqrt{2\pi}\Bigr)\1_{\{y\le0\}}+\1_{\{y\le-1\}}\log|y| . \end{equation} \begin{lemma}\label{II:cpt:lem-Pi} Let $00$, and $q,q'\in[\frac12,1]$. \begin{enumerate}[label=(\alph*)] \item $\Pi(\ell;q)=K_e(q)+c_e(q)/\ell+\frac{1+\rho_q}2\log\ell+\alpha\E W_{\mathrm{RS},\ell}(Y^{(q)})$. \item If $0<\ell\le L'$, then $W_{\mathrm{RS},\ell}\ge-\log(1+L')$ on $(-\infty,0]$, and $\alpha\E W_{\mathrm{RS},\ell}(Y^{(q)})\ge-3\log(1+L')-3\sqrt{L'}\,e$. \item $|\Pi(\ell;q)-\Pi(\ell;q')|\le3(1+\sqrt\ell)\,|q-q'|/(2e\ell)$. \end{enumerate} \end{lemma} \begin{proof} (a) Write $Y=Y^{(q)}$ and $k=\kappa/\sqrt q$. Since $eY_-=(\kappa-\sqrt q\,G)_+$, we have $\E Y_-^2=q\,m(k)/e^2$; also $\Prob(Y\le0)=\Phi(k)$. Hence, by \eqref{II:cpt:eq-WRS}, \begin{align*} \alpha\E\log\Phi(Y/\sqrt\ell)={}&\alpha\E W_{\mathrm{RS},\ell}(Y)-\frac{\alpha q\,m(k)}{2e^2\ell}\\ &+\rho_q\Bigl(\frac12\log\ell-\log\sqrt{2\pi}\Bigr)-\alpha\E\bigl[\log|Y|;Y\le-1\bigr]. \end{align*} Adding $\frac q{2e^2\ell}+\log e+\frac12\log\ell$ gives $\Pi(\ell;q)$ on the left. On the right, $\frac{q(1-\alpha m(k))}{2e^2\ell}=c_e(q)/\ell$ and $\log e-\alpha\E[\log|Y|;Y\le-1]-\rho_q\log\sqrt{2\pi}=K_e(q)$ by the definitions in \Cref{II:prop:decomp}, which proves (a). (b) Put $z=y/\sqrt\ell$. Since $\log\Phi(z)+z^2/2+\log\sqrt{2\pi}=-\log R(z)$, for $y\le0$ we have $W_{\mathrm{RS},\ell}(y)=-\log R(z)-\frac12\log\ell+\1_{\{y\le-1\}}\log|y|$. We use the Mills-ratio bounds $t0$ of \Cref{II:fun:gauss}(b). If $y\le-1$, then $t:=|z|\ge1/\sqrt\ell$ and $W_{\mathrm{RS},\ell}(y)=-\log(R(-t)/t)\ge-\log(1+t^{-2})\ge-\log(1+\ell)$. If $-10$, $W_{\mathrm{RS},\ell}(y)=\log\Phi(z)\ge-2\Phi(-z)$ by \Cref{II:fun:gauss}(c), and $\int_0^\infty2\Phi(-y/\sqrt\ell)\dd y=2\phi_0\sqrt\ell$. With $\alpha\le3$ and the density bound of \Cref{II:cpt:lem-Y0}, \[ \alpha\E W_{\mathrm{RS},\ell}(Y^{(q)})\ge-3\log(1+L')-\frac{3e}{\sqrt\pi}\,2\phi_0\sqrt{L'}, \] and $6\phi_0/\sqrt\pi<3$. (c) By \eqref{II:cpt:eq-dPi}, it suffices to show $|1-\alpha\E V(Y^{(q)}/\sqrt\ell)|\le3e(1+\sqrt\ell)$. By \Cref{II:cpt:lem-Y0} and \eqref{II:cpt:eq-rho}, $|1-\alpha\Prob(Y^{(q)}<0)|=|1-\rho_q|\le3e$. By the density bound of \Cref{II:cpt:lem-Y0} and $\alpha\le3$, \[ \alpha\bigl|\E\bigl[V(Y^{(q)}/\sqrt\ell)-\1_{\{Y^{(q)}<0\}}\bigr]\bigr|\le\frac{3e}{\sqrt\pi}\sqrt\ell\int_\R\bigl|V(z)-\1_{\{z<0\}}\bigr|\dd z=\frac{3e}{\sqrt\pi}\sqrt\ell\cdot2R(0)\le3e\sqrt\ell . \] Here $\int_{-\infty}^0(1-V)=\Rbar(0)=R(0)$ and $\int_0^\infty V=R(0)=2\phi_0$, because $\Rbar'=1-V$, $R'=-V$, and $\Rbar(-\infty)=R(\infty)=0$ (\Cref{II:fun:gauss}(a),(b)); and $12\phi_0/\sqrt\pi<3$. \end{proof} In \eqref{II:cpt:eq-jam} every jammed level contributes with the sign of $\mathcal J$. The next lemma shows that $\mathcal J$ is at least $-Ce$ at every level and is large and positive near the top. The source of the gain is \Cref{II:cpt:lem-jam}(b): the process $\widehat Y/\ell$ has nonnegative drift, so the diffusion started at $y\in[-\ell_0/\sqrt{\ell(t)},0]$ reaches level $t$ within $\sqrt{\ell(t)}$ of the wall, or beyond it, with probability at least $\frac12$. The Gaussian input puts mass of order $e\,\ell_0/\sqrt{\ell(t)}$ on this interval. \begin{lemma}\label{II:cpt:lem-J} Let $S\ge1$, let $00,\qquad C_S:=3+3\sqrt S . \] For every $\gamma\in\cU$ with $q_0>0$ and $s_0\le S$, and every $t\in[d,s_0]$, \[ \mathcal J(t)\ \ge\ c_{\mathcal J}\,e\,\min\Bigl\{\frac{\ell_0}{\sqrt{\ell(t)}},\frac1{2e}\Bigr\}-C_Se . \] \end{lemma} \begin{proof} We have $q_0=1-e^2s_0\in[\frac12,1]$, so \Cref{II:cpt:lem-Y0} applies to $p_0$. Write $\widehat Y^y:=\widehat Y^{s_0,y}$, $\ell=\ell(t)$, and $A:=\ell_0/\sqrt\ell$. Since $1-V>0$, we may restrict the integral in \eqref{II:cpt:eq-J} to $\{y>0\}\cup[-A,0]$. We treat the two regions separately and then combine them. \proofstep{1}{The region $y>0$.} By \Cref{II:cpt:lem-jam}(b) and since $V$ is decreasing, $V(\widehat Y^y_t/\sqrt\ell)\le V((y+B_{s_0-t})/\sqrt\ell)$ almost surely, and $B_{s_0-t}$ has the law of $\sigma G$ with $\sigma^2=s_0-t\le S$. For fixed $G$, since $00)-\frac{3e}{\sqrt\pi}\,3\phi_0\sqrt S\\ &\ge\alpha\bigl(1-\Phi(\kappa/\sqrt{q_0})\bigr)-3\sqrt S\,e , \end{align*} because $9\phi_0/\sqrt\pi<3$. \proofstep{2}{The region $-A\le y\le0$.} By \Cref{II:cpt:lem-jam}(b) with $s=s_0$, $\widehat Y^y_t\ge\ell\,y/\ell_0+\ell N_t\ge-\sqrt\ell+\ell N_t$. The centered Gaussian variable $N_t$ satisfies $\Prob(N_t\ge0)\ge\frac12$, with equality unless its variance vanishes, which happens at $t=s_0$, where $\Prob(N_t\ge0)=1$. On the event $\{N_t\ge0\}$ we get $\widehat Y^y_t/\sqrt\ell\ge-1$, and hence $1-V(\widehat Y^y_t/\sqrt\ell)\ge1-V(-1)$ because $V$ is decreasing. By \Cref{II:cpt:lem-Y0}, $\Prob(-A\le Y_0\le0)\ge c_pe\min\{A,1/(2e)\}$. Since $\alpha/2\ge1$, \[ \alpha\int_{-A}^0p_0(y)\,\E\bigl[1-V(\widehat Y^y_t/\sqrt\ell)\bigr]\dd y\ge c_{\mathcal J}\,e\min\{A,1/(2e)\}. \] \proofstep{3}{Conclusion.} Since $\alpha-(\alpha-1)=1$, we have $\alpha(1-\Phi(\kappa/\sqrt{q_0}))-(\alpha-1)=1-\rho\ge-3e$ by \eqref{II:cpt:eq-rho}. Adding the two regions and subtracting $\alpha-1$ gives the claim. \end{proof} \begin{corollary}\label{II:cpt:cor-jam} Let $S\ge1$, and let $c_{\mathcal J},C_S$ be as in \Cref{II:cpt:lem-J}. Put $e_{\mathrm J}=e_{\mathrm J}(S,\Theta):=\min\{e_{\mathrm I},\frac14,(2S)^{-1/2},c_{\mathcal J}/(4C_S)\}$. For $00$ and $s_0\le S$ satisfies \begin{equation}\label{II:cpt:eq-jamLB} \cP_\kappa(\gamma)\ \ge\ \Pi(\ell_0;q_0)+\frac{c_{\mathcal J}e}4\int_d^{t_1}\frac{1-\zeta(t)}{\ell(t)}\min\Bigl\{\frac{\ell_0}{\sqrt{\ell(t)}},\frac1{2e}\Bigr\}\dd t-C_Se\,\frac{s_0}{2\ell_0}\log\frac{s_0}{t_1}, \end{equation} where $t_1:=\min\{s_0,(c_{\mathcal J}\ell_0/(2C_S))^2\}$, and the integral is zero if $t_1\le d$. \end{corollary} \begin{proof} The choice of $e_{\mathrm J}$ gives $|\kappa|\le e\le\frac14$, $e^2S\le\frac12$, and $c_{\mathcal J}/(2e)\ge2C_S$, so \Cref{II:cpt:lem-J} applies. We use \eqref{II:cpt:eq-jam}. For $t\in[d,t_1]$ we have $\ell(t)\le t\le t_1$, so $c_{\mathcal J}\ell_0/\sqrt{\ell(t)}\ge2C_S$. Together with $c_{\mathcal J}/(2e)\ge2C_S$, \Cref{II:cpt:lem-J} gives $\mathcal J(t)\ge\frac12c_{\mathcal J}e\min\{\ell_0/\sqrt{\ell(t)},1/(2e)\}$ there. For $t\in[\max\{d,t_1\},s_0]$ we use $\mathcal J(t)\ge-C_Se$, $1-\zeta\le1$, and $\ell(t)\ge\ell_0t/s_0$ from \eqref{II:cpt:eq-concave}; the integral of $s_0/(2\ell_0t)$ over $[\max\{d,t_1\},s_0]$ is at most $\frac{s_0}{2\ell_0}\log\frac{s_0}{t_1}$. \end{proof} \section{The mass lower bound}\label{II:cpt:ssec-mass} We now show that the rescaled mass of a global minimizer of bounded depth cannot be small. The mechanism is visible in \Cref{II:cpt:lem-Pi}(a): the coefficient $c_e(q_0)\approx2\phi_0\theta$ of $1/\ell_0$ is positive, so $c_e(q_0)/\ell_0$ dominates $\log\ell_0$ as $\ell_0\downarrow0$. By \eqref{II:cpt:eq-jamLB}, the jammed levels lower $\cP_\kappa(\gamma)$ below $\Pi(\ell_0;q_0)$ by at most $O(e\log(1/e)/\ell_0)$. \begin{lemma}\label{II:cpt:lem-allLB} Let $S\ge1$ and $L\ge1$. There are $C_{\mathrm A},C'_{\mathrm A}<\infty$ and $e_{\mathrm A}>0$, depending only on $S$, $L$, $c_{\mathrm D}$, and $\Theta$, with the following property. Let $00$ have rescaled data satisfying \[ s_0\le S,\qquad \ell_0\le L,\qquad d\ge c_{\mathrm D}e^2 . \] Then \begin{equation}\label{II:cpt:eq-allLB} \cP_\kappa(\gamma)\ \ge\ K_e+\frac{2\phi_0\theta-C_{\mathrm A}e\log(1/e)}{\ell_0}+\frac{1+\rho}2\log\ell_0-C'_{\mathrm A} . \end{equation} \end{lemma} \begin{proof} Let $e_{\mathrm A}:=\min\{e_{\mathrm J}(S,\Theta),e_{\mathrm K}(S),1/\e\}$, so that \eqref{II:cpt:eq-jamLB} and \eqref{II:cpt:eq-coef} apply and $\log(1/e)\ge1$. Drop the nonnegative middle term in \eqref{II:cpt:eq-jamLB}. If $t_10$ such that for $00$ (\Cref{II:prop:kkt}(c)), \ref{II:cpt:in-upper} and \ref{II:cpt:in-gap} show that $\gamma$ satisfies the hypotheses of \Cref{II:cpt:lem-allLB}. Combining \eqref{II:cpt:eq-allLB} with \ref{II:cpt:in-rs}, \[ \frac{\phi_0\theta_-}{\ell_0}+\frac{1+\rho}2\log\ell_0\le C'_{\mathrm{RS}}+C'_{\mathrm A}=:C''_{\mathrm A} . \] Suppose $\ell_0<1$. Since $0<(1+\rho)/2\le2$, the second term is at least $2\log\ell_0$. Let $x_0\in(0,1)$ be such that $2x\log(1/x)\le\phi_0\theta_-/2$ for $00$; only the smallness of $\kappa$ and of the mass $\lambda(0)$ enters. For $\gamma\in\cU$ let $X$ be its Parisi diffusion. We use the scalar functions $\cN$ and $\widetilde B$ of \eqref{II:fun:scalars} and the \emph{breakpoint ratio} $\widetilde x$ of \eqref{II:fun:Sigma}, with a subscript $\gamma$ when the dependence on the order parameter matters: \begin{equation}\label{II:cpt:eq-xt} \widetilde B_\gamma(q)=\E\bigl[v^2(1-v)\bigr](q,X_q)>0,\qquad \cN_\gamma(q)=\E\bigl[v_x(q,X_q)^2\bigr],\qquad \widetilde x_\gamma(q)=\frac{\lambda(q)\cN_\gamma(q)}{2\,\widetilde B_\gamma(q)} . \end{equation} The breakpoint ratio governs the derivative of $S_\gamma(q):=\alpha\E v(q,X_q)^2$, the function $S$ of \eqref{II:fun:scalars}, for which $D'=(S_\gamma-1)/\lambda^2$. By \Cref{II:lem:generator}(c)(iv), $S_\gamma$ has one-sided derivatives at every $q\in[0,1)$, \begin{equation}\label{II:cpt:eq-Sprime} S_\gamma'(q+)=\frac{2\alpha\widetilde B_\gamma(q)}{\lambda(q)}\bigl(\widetilde x_\gamma(q)-\gamma(q)\bigr),\qquad S_\gamma'(q-)=\frac{2\alpha\widetilde B_\gamma(q)}{\lambda(q)}\bigl(\widetilde x_\gamma(q)-\gamma(q-)\bigr), \end{equation} and the functions $\widetilde B_\gamma$ and $\widetilde x_\gamma$ are continuous (\Cref{II:lem:generator}(c)). These are the analogues of the one-sided derivatives in \cite[Proposition~3]{AuffingerChenProperties}, which also differ only through the value of the distribution function at $q$ or at $q-$. In the physics literature, $\widetilde x$ is the formula for the order parameter on an interval of the support in the full replica symmetry breaking solution of the perceptron \cite[(36)]{FPSUZ2017}. The argument of this section follows the strategy of Auffinger and W.-K.\ Chen in the proof of \cite[Theorem~4]{AuffingerChenProperties}. There, continuity of a Parisi measure at the top of its support and these one-sided derivatives identify the mass below the top with a ratio of two expectations, and Jensen's inequality shows that this ratio is not one under a condition on the mixture. The same strategy is used for the Sherrington--Kirkpatrick model at every $\beta>1$ in \cite[Proposition~9.3]{LopattoFRSB}, where the log-concavity of a transformed density and a covariance inequality exclude the corresponding relation. The argument here differs from these works as follows. The functional has a hard wall: at the top of the support the solution is $\log\Phi((x-\kappa)/\sqrt\delta)$, in place of $\log\cosh x$ up to a constant. When the top is isolated in the support, $\gamma(Q-)<1$ is automatic but a bound uniform in the minimizer is not; in this case the gap principle (\Cref{II:lem:gap}) produces a level $x^*$ below the top with $\widetilde x_\gamma(x^*)=\gamma(Q-)$. The conclusion is a bound $\gamma(Q-)\le m_*$, not only an atom at the top, and $m_*$ is uniform: \Cref{II:cpt:prop-topcheb} holds for all order parameters, and $m_*$ does not depend on $e$, $\theta$, or the minimizer. Finally, the monotone weight comes from the tilted-density representation of \Cref{II:lem:tilted}, not from log-concavity; Auffinger and W.-K.\ Chen express the corresponding averages through a Brownian motion and an exponential weight \cite[(18)]{AuffingerChenProperties}. As in \cite[Proposition~9.3]{LopattoFRSB}, the ratio is then bounded by a covariance inequality, here Chebyshev's association inequality combined with the monotonicity of $(V')^2/(V^2(1-V))$ (\Cref{II:appB:lem-mlr}); Auffinger and W.-K.\ Chen use Jensen's inequality at this point. We first translate the first-order conditions at the top into a statement about $\widetilde x_\gamma$. \begin{lemma}\label{II:cpt:lem-topKKT} Let $\gamma$ be a global minimizer of $\cP_\kappa$ with top $Q$ and $\mu_\gamma\ne\delta_Q$. \begin{enumerate}[label=(\alph*)] \item If $Q$ is a limit of points of $\supp\mu_\gamma\cap[0,Q)$, then $\gamma(Q-)=\widetilde x_\gamma(Q)$. \item Otherwise let $p:=\max(\supp\mu_\gamma\cap[0,Q))$. Then $\gamma\equiv m:=\gamma(Q-)$ on $[p,Q)$, $00$, $\gamma(Q-)=\widetilde x_\gamma(Q)$. (b) Since $Q$ is isolated from the left in $\supp\mu_\gamma$ and $\mu_\gamma\ne\delta_Q$, the point $p$ exists and $(p,Q)\cap\supp\mu_\gamma=\emptyset$. The gap principle, \Cref{II:lem:gap}(a) with $p'=Q$, gives $\gamma\equiv m=\gamma(p)$ on $[p,Q)$ with $00$ there is $\bar\eta\in(0,1/4]$ such that the following holds for every $\kappa\in[-\bar\eta,\bar\eta]$ and every $\gamma\in\cU$ with top $Q$. \begin{enumerate}[label=(\alph*)] \item If $\delta=1-Q\le\bar\eta$, then $\widetilde x_\gamma(Q)\le x_\Phi+\eps$. \item If $\gamma\equiv m$ on $[p,Q)$ for some $p0$ depending only on $\Theta$ such that for $0m_*$. Then $1-\bar\eta\le1-\eta_0m_*$. Hence $\gamma(Q-)\le m_*$. \end{proof} The theorem is global: it assumes neither closeness of $\gamma$ to a replica-symmetric order parameter nor any bound on $d$. The smallness needed in \Cref{II:cpt:prop-topcheb} comes only from $\lambda(p)\le\lambda(0)=e^2\ell_0\le e^2L$ and $|\kappa|\le e$. The modulus in \Cref{II:cpt:prop-topcheb} is small only when $\kappa$, $\lambda(p)$, and $1-m$ are all small, which is why $m_*$ is chosen through $\bar\eta$ and is not explicit. The bound uses stationarity and fails for arbitrary order parameters: for $\gamma=m\1_{[p,Q)}+\1_{[Q,1]}$ the jump $1-\gamma(Q-)=1-m$ can be arbitrarily small. \section{The top-gap bound}\label{II:cpt:ssec-gap} If the top gap $d$ is small and $\gamma(Q-)\le m_*<1$, then just below the wall the profile is jammed: on $(d,2d]$ one has $1-\zeta\ge1-m_*$ and $\ell\le2d$. By \Cref{II:cpt:lem-J} the jamming coefficient there is of order $e\min\{\ell_0/\sqrt d,1/e\}$, which dominates every other term of \eqref{II:cpt:eq-jamLB} once $d$ is small. The competitor is the replica-symmetric order parameter of the same mass. \begin{proposition}\label{II:cpt:prop-gap} Let $S\ge1$, $\ellmin\in(0,1]$ and $m_*\in(0,1)$. There are $d_1>0$ and $e_{\mathrm E}>0$, depending only on $S$, $\ellmin$, $m_*$, and $\Theta$, such that the following holds for $00$ have rescaled data satisfying \[ s_0\le S,\qquad \ell_0\ge\ellmin,\qquad \gamma(Q-)=\zeta(d+)\le m_*,\qquad d\le d_1 , \] and let $\gamma^\natural:=\1_{[1-e^2\ell_0,1]}$ be the replica-symmetric order parameter of the same mass. Then $\cP_\kappa(\gamma)>\cP_\kappa(\gamma^\natural)$. \end{proposition} \begin{proof} Let $c_{\mathcal J},C_S$ be as in \Cref{II:cpt:lem-J} and put \[ \begin{gathered} t_{\min}:=\min\bigl\{\ellmin,(c_{\mathcal J}\ellmin/(2C_S))^2\bigr\},\qquad C_\sharp:=\frac{3(1+\sqrt S)S}{2\ellmin}+\frac{C_SS}{2\ellmin}\log\frac S{t_{\min}},\\ \Lambda_\sharp:=\frac{8C_\sharp}{c_{\mathcal J}(1-m_*)\log2},\qquad d_1:=\min\Bigl\{\frac{t_{\min}}2,\ \frac{\ellmin^2}{2\Lambda_\sharp^2}\Bigr\},\qquad e_{\mathrm E}:=\min\Bigl\{e_{\mathrm J}(S,\Theta),\ \frac1{2\Lambda_\sharp}\Bigr\}. \end{gathered} \] Let $00 . \] This proves the proposition. \end{proof} The proposition is an inequality for all order parameters, and it is combined with minimality only in the proof of \Cref{II:thm:T1}. It needs no lower bound on $d$ in terms of $e$: when $d\lesssim e^2$ the minimum in the gain is $1/(2e)$, and the gain is of order one. The bounds $\ell_0\ge\ellmin$ and $s_0\le S$ enter only through $t_1\ge t_{\min}$, the Lipschitz bound of \Cref{II:cpt:lem-Pi}(c) and the factor $\ell_0/\sqrt\ell$ in the gain. \section{Proof that minimizers lie in a compact class}\label{II:cpt:ssec-T1} \begin{proof}[Proof of \Cref{II:thm:T1}] We fix the constants in the order \eqref{II:cpt:eq-order}. The window $\Theta$ fixes $\Dtop$, $L$ and $C_0$ (\Cref{II:thm:upper}), $c_{\mathrm D}$ (\Cref{II:thm:gapD}), $C'_{\mathrm{RS}}$ (\Cref{II:lem:rsvalue}(b)), and then $S_*$ (\Cref{II:thm:depth}). Let $\ellmin:=\ellmin(S_*,\Theta)$ and $e_{\mathrm M}$ be given by \Cref{II:prop:massLB} with $S=S_*$, and let $m_*$ and $e_{\mathrm T}$ be given by \Cref{II:thm:topjump}. Let $d_1$ and $e_{\mathrm E}$ be given by \Cref{II:cpt:prop-gap} with $(S,\ellmin,m_*)=(S_*,\ellmin,m_*)$. Put $e_*:=\min\{e_{\mathrm I},e_{\mathrm M},e_{\mathrm T},e_{\mathrm E}\}$. Let $00$ by \Cref{II:prop:kkt}(c). By \Cref{II:thm:depth}, $s_0\le S_*$. By \Cref{II:prop:massLB}, $\ell_0\ge\ellmin$. By \Cref{II:thm:topjump}, $\gamma(Q-)\le m_*$. Suppose that $d\le d_1$. Then $\gamma$ satisfies all hypotheses of \Cref{II:cpt:prop-gap}, so $\cP_\kappa(\gamma^\natural)<\cP_\kappa(\gamma)=\cP_*(\kappa)$, which contradicts the definition of $\cP_*(\kappa)$. Hence $d>d_1$, and the rescaled profile of $\gamma$ lies in $\cK(d_1,S_*)$. Each constant depends only on the constants preceding it in \eqref{II:cpt:eq-order}, and none depends on $e$, $\theta$, or the minimizer. This proves the theorem. \end{proof} \chapter{The flat limit}\label{II:sec:flat} \Cref{II:sec:depth,II:sec:compact} show that, for $\alpha=2+e$ with $e$ small and $\theta\in\Theta$, the rescaled profile of every global minimizer lies in a fixed compact class $\cK(d_1,S_*)$ (\Cref{II:thm:T1}). This chapter converts that compactness into closeness to the replica-symmetric crossing order parameter $\gamma_{\mathrm d}=\1_{[\qd,1]}$. This is Step~4 of \Cref{II:sec:intro-ideas}. \Cref{II:flat:ss-def} defines the flat functional $F_t=A+tI$ (\Cref{II:def:flat}). In \Cref{II:flat:ss-exp} we expand $\cP_\kappa(\gamma)$ to order $e$, uniformly on compact classes (\Cref{II:thm:compactexp}). \Cref{II:flat:ss-inverse,II:flat:ss-mono,II:flat:ss-conv} pass to the inverse profile $T=\ell^{-1}$ and prove that $I$ is convex along every segment from the replica-symmetric profile to a normalized step profile. \Cref{II:flat:ss-fs} computes the first variation of $F_{\td}$ at $\1_{[0,1]}$, and \Cref{II:flat:ss-uniq} combines the two into the cubic modulus of \Cref{II:thm:flatuniq}, for every $t\ge\td=1/(4c_0)$. In \Cref{II:flat:ss-conv-cross} the expansion and the modulus give a stability estimate (\Cref{II:flat:cor-nearmin}), uniform in $\theta$ near $\thd$, from which \Cref{II:sec:frsb} starts. Throughout this chapter, $\alpha=2+e$, $\Theta$ is a compact window as in \Cref{II:rs:rem-Theta}, with endpoints $\theta_-\le\theta_+$, and $\kappa=\kappa_e(\theta)=\kRS(2+e)-e^2\theta$ with $\theta\in\Theta$. We use the rescaled variables of \Cref{II:def:rescaled}, and we write $L_e=\log(1/e)$. Constants $c,C$ depend only on $\Theta$ and on the parameters named in each statement. The longer proofs are in \Cref{II:app:flat}. \section{Compact classes and the flat functional}\label{II:flat:ss-def} A \emph{profile} is a nonincreasing function $\zeta\colon[0,\infty)\to[0,1]$. Profiles that agree almost everywhere are identified, and we use the distance of $L^1(0,\infty)$. The class $\cK(d_1,S)$, $00$ with $e^2S<1$, and let $\gamma\in\cU$ be the right-continuous function with $\gamma(q)=\zeta((1-q)/e^2)$ for $q\ge1-e^2S$ and $\gamma=0$ below. Then $\zeta_\gamma=\zeta$, the bottom of the support of $\mu_\gamma$ is $1-e^2s_0>0$, and $\gamma$ has rescaled top gap $d$ and depth $s_0$. Apply \Cref{II:lem:nonstep}(a),(b) with the finite set $\{d,s_0\}$. The rescaled profiles $\zeta_n$ of the step profiles $\gamma_n$ do not depend on $e$; they are nonincreasing step functions, equal to one on $[0,d]$ and to zero on $(s_0,\infty)$, and they preserve the mass. Thus \begin{equation}\label{II:flat:eq-steps} \zeta_n\in\cK(d,s_0)\subset\cK(d_1,S),\qquad \int_0^\infty\zeta_n=\int_0^\infty\zeta,\qquad \zeta_n\to\zeta\ \text{ in }L^1 . \end{equation} In particular $\zeta_n\in\cK^1(d_1,S)$ if $\zeta\in\cK^1(d_1,S)$. We call $(\zeta_n)$ the \emph{step approximations} of $\zeta$. For $\zeta\in\cK(d_1,S)$, let $U=U_\zeta$ be the solution on $(0,S]\times\R$ of \begin{equation}\label{II:flat:eq-pde} U_s=\tfrac12\bigl(U_{yy}+\zeta\,U_y^2\bigr)\quad(d_10 . \] The flat functional renormalizes $U(S,\cdot)$ by its behavior as $y\to-\infty$. By \Cref{II:flat:lem-tails} below this behavior is $-y^2/(2\ell_0)-\log|y|+\omega(S)$, where, for $s\ge d_1$, \begin{equation}\label{II:flat:eq-Gamma} \Gamma(s)=\tfrac12\log d_1+\tfrac12\int_{d_1}^s\frac{\dd t}{\ell(t)},\qquad \omega(s)=\log\ell(s)-\Gamma(s)-\log\sqrt{2\pi} . \end{equation} Since $\ell(t)=t$ on $[d_1,d]$, $\Gamma(s)$ agrees with $\frac12\log d+\frac12\int_d^s\dd t/\ell(t)$, the function $\Gamma$ of \Cref{II:def:rescaled}. \begin{definition}\label{II:def:flat} Let $\zeta\in\cK(d_1,S)$. Put \begin{align*} W_\zeta(y)&=U_\zeta(S,y)+\frac{y_-^2}{2\ell_0}-\omega(S)\1_{\{y\le0\}} +\log|y|\,\1_{\{y\le-1\}},\\ \widetilde A(\zeta)&=\frac18\Bigl[\int_{d_1}^S\Bigl(\frac1{\ell_0}-\frac1{\ell(s)}\Bigr)\dd s +\frac{d_1}{\ell_0}-\log d_1\Bigr],\qquad \cP^{(1)}(\zeta)=\widetilde A(\zeta)+2\phi_0\int_\R W_\zeta(y)\dd y . \end{align*} For a normalized profile $\zeta\in\cK^1(d_1,S)$ we write $A(\zeta)=\widetilde A(\zeta)$, $I(\zeta)=\int_\R W_\zeta\dd y-1$ and, for $t\in\R$, \[ F_t(\zeta)=A(\zeta)+t\,I(\zeta). \] We call $F_t$ the \emph{flat functional}. Its \emph{crossing parameter} is $\td=1/(4c_0)$, where $c_0=\int_\R V(1-V)$. \end{definition} \Cref{II:flat:lem-tails} shows that $W_\zeta\in L^1(\R)$, and \Cref{II:flat:cor-welldef} shows that $\cP^{(1)}(\zeta)$ does not depend on the pair $(d_1,S)$ used to compute it. The two terms have different roles. The term $\widetilde A$ collects the order-$e$ part of the entropy and of the bottom Gaussian term. On normalized profiles it is maximized by $\1_{[0,1]}$ (\Cref{II:flat:cor-Amax}), so it favors replica symmetry breaking. The term $\int W_\zeta$ is the order-$e$ part of the energy $\alpha\E U(s_0,Y_0)$. On bounded sets the law of the bottom position $Y_0$, multiplied by $\alpha$, has density $2e\phi_0(1+o(1))$ by \Cref{II:lem:valueformula}, which is why Lebesgue measure appears. \section{The compact-class expansion}\label{II:flat:ss-exp} Let $K_e=\log e-\alpha\E[\log|Y^*|;\,Y^*\le-1]-\alpha\Phi(\kappa)\log\sqrt{2\pi}$, with $Y^*=(G-\kappa)/e$, be the profile-independent constant of \Cref{II:prop:decomp}. For $l>0$ put \begin{equation}\label{II:flat:eq-Ke} \Psi_\theta(l)=\frac{2\phi_0\theta}{l}+\log l,\qquad g_\theta(l)=\Bigl(\phi_0-\frac1{8\phi_0}\Bigr)\frac\theta l+\frac14\log l . \end{equation} The scalar function $\Psi_\theta$ of the mass always carries the subscript $\theta$; it is unrelated to the spatial function $\Psi(s,y)=\ell U_y+y$. The following uniform description of $U_\zeta$ as $y\to\pm\infty$ is the analytic input of the expansion. Parts (a) and (d) are the asymptotics of \Cref{II:lem:regularity} read through the dilation; parts (b) and (c) come from Feynman--Kac representations along the rescaled diffusion and a transport estimate for it. The proof is in \Cref{II:appC:ss-tails}. \begin{lemma}\label{II:flat:lem-tails} Fix $00$ such that for $00$ on $[0,s_0)$, so $\ell$ is strictly increasing on $[0,s_0]$. Let $T\colon[0,1]\to[0,s_0]$ be its inverse. Then $T(l)=l$ on $[0,d_1]$, $T(1)=s_0$, and $T'(l)=1/\zeta(T(l))\ge1$ for almost every $l$. Since $\zeta$ is nonincreasing, $\ell$ is concave, so its increasing inverse $T$ is convex. The inverse $T$ is continuous on $[0,1]$. Convexity together with continuity at both endpoints implies that $T$ is absolutely continuous on $[0,1]$. Conversely, every continuous convex $T$ on $[0,1]$ with $T(0)=0$, $T'=1$ on $[0,d_1]$ and $T(1)<\infty$ is the inverse of a unique $\zeta\in\cK^1(d_1,T(1))$, namely the derivative of $T^{-1}$ extended by zero. Normalized step profiles correspond to piecewise linear $T$. Bounded support is used here: a normalized profile with unbounded support has $T(1)=\infty$. We write $\sigma=T'$, a nondecreasing function. In the variable $l$, $\sigma$ plays the role of a diffusivity. The replica-symmetric profile $\1_{[0,1]}$ corresponds to $T=\mathrm{id}$ and $\sigma\equiv1$. Neither $A$ nor $I$ depends on the pair $(d_1,S)$ with $\zeta\in\cK^1(d_1,S)$. Indeed, $\ell(s)=s$ on $[d_1,d]$ and $\ell=\ell_0$ on $[s_0,S]$, so \[ 8\widetilde A(\zeta)=\int_d^{s_0}\Bigl(\frac1{\ell_0}-\frac1{\ell(s)}\Bigr)\dd s+\frac d{\ell_0}-\log d , \] which involves neither $d_1$ nor $S$; and $2\phi_0(I+1)=\cP^{(1)}-\widetilde A$ does not depend on them by \Cref{II:flat:cor-welldef}. We therefore compute $A$ and $I$ with the cutoff $S=s_0=T(1)$ whenever we work in the variable $l$. \begin{proposition}\label{II:flat:prop-inverse} Let $\zeta\in\cK^1(d_1,S)$, with inverse profile $T$ and $\sigma=T'$. Then \begin{equation}\label{II:flat:eq-Aaffine} 8A(\zeta)=1-\int_0^1(\sigma(l)-1)\frac{1-l}{l}\dd l,\qquad \Gamma(s_0)=\frac12\int_0^1\frac{\sigma(l)-1}{l}\dd l . \end{equation} Moreover, $\mathcal U(l,y)=U_\zeta(T(l),y)$ satisfies $\mathcal U(l,y)=\log\Phi(y/\sqrt l)$ for $l\le d_1$, $\mathcal U(1,\cdot)=U_\zeta(s_0,\cdot)$, and, for almost every $l\in(d_1,1)$, \begin{equation}\label{II:flat:eq-lpde} \mathcal U_l=\tfrac12\mathcal U_y^2+\tfrac{\sigma(l)}2\,\mathcal U_{yy} . \end{equation} \end{proposition} \begin{proof} Substitute $s=T(l)$, $\dd s=\sigma\dd l$, and $\ell(T(l))=l$. Since $(1-1/l)\sigma=-(\sigma-1)(1-l)/l+(1-1/l)$, \[ \int_{d_1}^{s_0}\Bigl(1-\frac1{\ell(s)}\Bigr)\dd s =\int_{d_1}^1\Bigl(1-\frac1l\Bigr)\sigma\dd l =-\int_{d_1}^1(\sigma-1)\frac{1-l}{l}\dd l+1-d_1+\log d_1 . \] Adding $d_1-\log d_1$ and using $\sigma=1$ on $[0,d_1]$ gives the formula for $8A$. Similarly, $\Gamma(s_0)=\frac12\log d_1+\frac12\int_{d_1}^1\sigma\dd l/l=\frac12\int_{d_1}^1(\sigma-1)\dd l/l$. Finally, $\mathcal U_l=\sigma\,U_s$ and $\sigma\,\zeta(T(l))=1$ give \eqref{II:flat:eq-lpde}. \end{proof} From now on, in the variable $l$, we write $U(l,y)$ for $\mathcal U(l,y)$. Since $\ell_0=1$, we have $\omega(s_0)=-\Gamma(s_0)-\log\sqrt{2\pi}$, and \begin{equation}\label{II:flat:eq-Il} \begin{split} I(\zeta)&=\int_\R\Bigl[U(1,y)+\frac{y_-^2}2+\bigl(\Gamma_\sigma+\log\sqrt{2\pi}\bigr)\1_{\{y\le0\}} +\log|y|\,\1_{\{y\le-1\}}\Bigr]\dd y-1,\\ \Gamma_\sigma&=\frac12\int_0^1\frac{\sigma-1}{l}\dd l . \end{split} \end{equation} Only $U(1,\cdot)$ depends nonlinearly on $\sigma$; the counterterms are affine in $\sigma$. \begin{corollary}\label{II:flat:cor-Amax} For every normalized profile $\zeta$ with bounded support and $\zeta=1$ near $0$, $A(\zeta)\le1/8$, with equality if and only if $\zeta=\1_{[0,1]}$. \end{corollary} \begin{proof} In \eqref{II:flat:eq-Aaffine}, $\sigma-1\ge0$ and $(1-l)/l>0$ on $(0,1)$. The integral vanishes if and only if $\sigma=1$ almost everywhere, that is, $T=\mathrm{id}$. \end{proof} \Cref{II:flat:cor-Amax} explains why the flat problem is not trivial: moving away from $\1_{[0,1]}$ always decreases $A$, and the question is whether $tI$ increases enough to compensate. For $t$ near $0$ it does not. \section{Monotonicity of the flat density}\label{II:flat:ss-mono} We now fix a normalized step profile, that is, a piecewise linear convex $T$. Its breakpoints $d_1\le l_00$. On a block, $F:=\e^{U/\sigma_j}$ solves the heat equation $F_l=\frac{\sigma_j}2F_{yy}$, by \eqref{II:flat:eq-lpde}. Hence, with $P_\tau f(y)=\E f(y+\sqrt\tau\,G)$, \begin{equation}\label{II:flat:eq-colehopf} U(l,\cdot)=\sigma_j\log P_{\sigma_j(l-l_{j-1})}\bigl[\e^{U(l_{j-1},\cdot)/\sigma_j}\bigr], \qquad l\in(l_{j-1},l_j]. \end{equation} We take \eqref{II:flat:eq-colehopf}, with $U(l,y)=\log\Phi(y/\sqrt l)$ for $l\le l_0$, as the definition of $U$ for a general positive step function $\sigma$. Write $b=U_y$ and $c=-U_{yy}$, and let \[ \mathcal G=b\,\partial_y+\frac\sigma2\,\partial_{yy} \] be the operator that appears in the linearization of \eqref{II:flat:eq-lpde}. The following signs are the rescaled counterparts of \Cref{II:lem:regularity}(c); for $\sigma$ that comes from a profile, the upper bound $c\le1/l$ is the rescaled form of the bound $c<1/\lambda$ there. What the lemma adds is that the signs, including $c\le1/l$, hold for every positive step function $\sigma$. The bound $c\le1/l$ controls the flat density below. For the Sherrington--Kirkpatrick datum $\log\cosh$, the analogous bounds on the first two derivatives are \cite[Proposition~2]{AuffingerChenProperties}, \cite[Proposition~2]{AuffingerChen2015Uniqueness}, and \cite[Lemma~16]{JagannathTobasco2016}, and the sign of the third derivative on the positive half-line is \cite[(7.6)]{LopattoFRSB}. \begin{lemma}\label{II:flat:lem-signs} Let $\sigma>0$ be a step function on a grid of $[l_0,1]$ with $\sigma=1$ on $(0,l_0]$, and let $U$ be defined by \eqref{II:flat:eq-colehopf}. For every $l\in(0,1]$, $U(l,\cdot)\le0$ is smooth, $b>0$, and \[ 00$, $c=V(y/\sqrt l)/l\in(0,1/l)$ and $c_y=V'(y/\sqrt l)\,l^{-3/2}<0$ by \Cref{II:fun:gauss}(a). We argue by induction over the blocks. Let $l\in(l_{j-1},l_j]$, $\sigma=\sigma_j$, $\tau=\sigma(l-l_{j-1})$, and $f=\e^{U(l_{j-1},\cdot)/\sigma}$. We verify the four claims in turn. \proofstep{1}{$U\le0$ and $b>0$.} The function $f$ takes values in $(0,1]$ and is strictly increasing, and $P_\tau$ preserves both properties. \proofstep{2}{$c\ge0$.} The function $f$ is log-concave because $U(l_{j-1},\cdot)$ is concave. The Gaussian convolution $P_\tau f$ is a marginal of the log-concave function $(w,y)\mapsto f(w)\phi_\tau(y-w)$, where $\phi_\tau$ is the $N(0,\tau)$ density, and the marginal of a log-concave function is log-concave by Pr\'ekopa's theorem \cite[Theorem~6]{Prekopa1973}. \proofstep{3}{$c\le1/l$.} By induction, $-(\log f)''=c(l_{j-1},\cdot)/\sigma\le K_0:=1/(\sigma l_{j-1})$, so $f(w)=\e^{-K_0w^2/2}g(w)$ with $g$ log-convex. Completing the square, \[ P_\tau f(y)=\frac{\e^{-K_0y^2/(2(1+K_0\tau))}}{\sqrt{1+K_0\tau}}\; \E\,g\Bigl(\frac{y}{1+K_0\tau}+\sqrt{\frac{\tau}{1+K_0\tau}}\,G\Bigr). \] A mixture of log-convex functions is log-convex, so the expectation is log-convex in $y$. Hence $-(\log P_\tau f)''\le K_0/(1+K_0\tau)=1/(\sigma l)$, that is, $c\le1/l$. \proofstep{4}{$c>0$ and $c_y\le0$.} Differentiating \eqref{II:flat:eq-lpde} twice and three times in $y$ gives, on each block, \[ c_l=\mathcal Gc-c\cdot c,\qquad g_l=\mathcal Gg-3c\,g\quad(g=c_y), \] and $c,g$ are continuous in $l$ across the breakpoints. The potentials $c$ and $3c$ are nonnegative and bounded by $3/l_0$, the drift $b$ is Lipschitz in $y$ (because $|b_y|=c\le1/l_0$), and $c,g$ have polynomial growth. Let $Y$ solve \[ Y_\lambda=y+\int_\lambda^lb(\lambda',Y_{\lambda'})\dd\lambda'+\int_\lambda^l\sqrt{\sigma(\lambda')}\,\dd\overline B_{\lambda'}\qquad(l_0\le\lambda\le l), \] where $\overline B$ is a Brownian motion in the reversed time $l-\lambda$. This is the diffusion with generator $\mathcal G$ run from $(l,y)$ downward in $l$, and it is well posed because $b$ is Lipschitz in $y$. By the Feynman--Kac formula along $Y$ (see \cite{KaratzasShreve1991}), \[ c(l,y)=\E\Bigl[c(l_0,Y_{l_0})\,\e^{-\int_{l_0}^lc(\lambda,Y_\lambda)\dd\lambda}\Bigr]>0,\qquad g(l,y)=\E\Bigl[g(l_0,Y_{l_0})\,\e^{-\int_{l_0}^l3c(\lambda,Y_\lambda)\dd\lambda}\Bigr]\le0, \] because $c(l_0,\cdot)>0$ and $g(l_0,\cdot)<0$. \end{proof} The diffusion in the Feynman--Kac formula is the rescaled Parisi diffusion of \Cref{II:def:rescaled}, written in the variable $l$ (see \eqref{II:appC:eq-Y}). Started at the initial level $l=1$ (the bottom of the support), it defines the density that governs the second variation. \begin{definition}\label{II:flat:def-density} Let $\sigma$ be as in \Cref{II:flat:lem-signs}. For $y\in\R$ let $(Y^y_l)_{00$ be a step function on a grid of $[l_0,1]$ with $\sigma=1$ on $(0,l_0]$. For every $l\in[l_0,1]$ the function $y\mapsto p_l(y)$ is nonincreasing. In particular $y\mapsto p_l(y)c(l,y)$ is nonincreasing. \end{proposition} \begin{proof} We argue by induction from the initial level $l=1$ downward. At $l=1$, $p_1\equiv1$. Suppose that $p_{l_j}$ is nonincreasing, and let $l\in[l_{j-1},l_j)$. Write $\sigma=\sigma_j$, $\Delta=\sigma(l_j-l)$, and $f_\lambda=U(\lambda,\cdot)/\sigma=\log F(\lambda,\cdot)$. By \Cref{II:flat:lem-signs}, at every level $\lambda$ the function $f_\lambda'=b/\sigma$ is positive, decreasing ($f_\lambda''=-c/\sigma<0$), and convex ($f_\lambda'''=-c_y/\sigma\ge0$). Fix $y$ and let $\mathbb Q_y(\dd x)\propto\phi_\Delta(x-y)\,p_{l_j}(x)F(l_j,x)^{-1}\dd x$. Differentiating \eqref{II:flat:eq-doob} and integrating by parts in $x$ (the boundary terms vanish by the bounds after \eqref{II:flat:eq-doob}), \[ (\log p_l)'(y)=f_l'(y)+\E_{\mathbb Q_y}\Bigl[\frac{X-y}\Delta\Bigr] =f_l'(y)+\E_{\mathbb Q_y}\bigl[(\log p_{l_j})'(X)-f'_{l_j}(X)\bigr]. \] Put $\bar r=\E_{\mathbb Q_y}f_{l_j}'(X)$. Since $(\log p_{l_j})'\le0$, the two expressions for the same expectation give \begin{equation}\label{II:flat:eq-mono1} (\log p_l)'(y)\le f'_l(y)-\bar r,\qquad \E_{\mathbb Q_y}X\le y-\Delta\bar r . \end{equation} \proofstep{1}{A lower bound for $\bar r$.} By Jensen's inequality ($f'_{l_j}$ is convex), then monotonicity ($f'_{l_j}$ is decreasing), and \eqref{II:flat:eq-mono1}, \[ \bar r\ge f'_{l_j}\bigl(\E_{\mathbb Q_y}X\bigr)\ge f'_{l_j}(y-\Delta\bar r). \] \proofstep{2}{A fixed point.} The function $\mathcal R(r)=f'_{l_j}(y-\Delta r)-r$ is continuous, and by $c\le1/l_j$, \[ \mathcal R'(r)=\frac\Delta\sigma\,c(l_j,y-\Delta r)-1=(l_j-l)\,c(l_j,y-\Delta r)-1\le\frac{l_j-l}{l_j}-1=-\frac l{l_j}<0 . \] Hence $\mathcal R$ has a unique zero $r^*$. Step~1 says $\mathcal R(\bar r)\le0$, so $\bar r\ge r^*$. Let $x^*=y-\Delta r^*$. Then $r^*=f'_{l_j}(x^*)$ and $y=x^*+\Delta f'_{l_j}(x^*)$. \proofstep{3}{Tweedie's formula.} Since $F(l_j,\cdot)=P_\Delta F(l,\cdot)$, the probability measure $\pi(\dd w)\propto F(l,w)\,\phi_\Delta(w-x^*)\dd w$ satisfies \begin{align*} \E_\pi W&=x^*+\Delta\,(\log F(l_j,\cdot))'(x^*)=x^*+\Delta f'_{l_j}(x^*)=y, \\ f'_{l_j}(x^*)&=\frac{(P_\Delta F(l,\cdot))'(x^*)}{P_\Delta F(l,\cdot)(x^*)}=\E_\pi f_l'(W). \end{align*} By Jensen's inequality ($f_l'$ is convex), $r^*=\E_\pi f_l'(W)\ge f'_l(\E_\pi W)=f_l'(y)$. The formula for $\E_\pi W$ is Tweedie's formula \cite{Robbins1956,Efron2011}. Combining the three steps with \eqref{II:flat:eq-mono1}, $(\log p_l)'(y)\le f'_l(y)-\bar r\le f'_l(y)-r^*\le0$. The flat density is continuous in $l$ at the breakpoints (at $l=l_j$, \eqref{II:flat:eq-doob} returns $p_{l_j}$), so the induction proceeds block by block. Since $p_l$ and $c$ are positive and nonincreasing, so is $p_lc$. \end{proof} \section{Convexity along ordered rays}\label{II:flat:ss-conv} Let $T_0,T_1$ be the inverse profiles of two normalized step profiles, on a common grid with $\sigma=1$ on $(0,l_0]$, such that $K:=T_1-T_0\ge0$ on $[0,1]$. For $\vartheta\in[0,1]$ put $T_\vartheta=T_0+\vartheta K$, so that $\sigma_\vartheta=T_\vartheta'=\sigma_0+\vartheta\dot\sigma$ with $\dot\sigma=K'$, and let $U^\vartheta$ be the corresponding solution and $I(\vartheta)=I(\zeta_\vartheta)$. The main example is the ray from replica symmetry: $T_0=\mathrm{id}$ and $T_1=T$ the inverse of an arbitrary normalized step profile, for which $K=T-\mathrm{id}\ge0$ because $T'\ge1$. For $\vartheta\in[0,1]$, $T_\vartheta$ is again convex with $T_\vartheta'\ge1$, $T_\vartheta'=1$ on $[0,l_0]$ and $T_\vartheta(1)\le\max\{T_0(1),T_1(1)\}$, so every $\zeta_\vartheta$ lies in the fixed class $\cK^1(l_0,\max\{T_0(1),T_1(1)\})$. Write $\dot U=\partial_\vartheta U^\vartheta$ and $\ddot U=\partial_\vartheta^2U^\vartheta$. Differentiating \eqref{II:flat:eq-lpde} in $\vartheta$, and using that $\sigma_\vartheta$ is affine in $\vartheta$, \begin{equation}\label{II:flat:eq-var} \dot U_l=\mathcal G\dot U-\tfrac{\dot\sigma}2\,c,\qquad \ddot U_l=\mathcal G\ddot U+\dot U_y^2+\dot\sigma\,\dot U_{yy},\qquad \dot U=\ddot U=0\ \text{ for } l\le l_0 . \end{equation} The function that makes the second variation transparent is \begin{equation}\label{II:flat:eq-Z} \mathcal Z=\dot U+\tfrac K2\,c,\qquad\text{which solves}\qquad \mathcal Z_l=\mathcal G\mathcal Z-\tfrac K2\,c^2,\quad \mathcal Z=0 \ (l\le l_0). \end{equation} Indeed, $c_l=\mathcal Gc-c^2$ (proof of \Cref{II:flat:lem-signs}), $K'=\dot\sigma$, and $K$ depends on $l$ only, so $\mathcal G(Kc)=K\mathcal Gc$. \begin{proposition}\label{II:flat:prop-zform} In the setting above, $\vartheta\mapsto I(\vartheta)$ is of class $C^2$ on $[0,1]$ and \begin{equation}\label{II:flat:eq-zform} I''(\vartheta)=\int_{l_0}^1\!\!\int_\R\Bigl[p_l\,\mathcal Z_y^2-2K\,(p_lc)_y\,\mathcal Z_y +\frac{K^2}2\,p_l\,c_y^2\Bigr]\dd y\,\dd l , \end{equation} where $p$, $c$, and $\mathcal Z$ are computed at $\sigma_\vartheta$. \end{proposition} The proof, in \Cref{II:appC:ss-zform}, starts from $I''=\iint p\,(\dot U_y^2+\dot\sigma\dot U_{yy})$ and substitutes $\dot U=\mathcal Z-Kc/2$. All terms that involve derivatives in $l$ combine into $(\partial_l-\mathcal G)(K\mathcal Z_{yy}-K^2c_{yy}/4)$, whose integral against $p$ vanishes; one integration by parts in $y$ then gives \eqref{II:flat:eq-zform}. \begin{corollary}\label{II:flat:cor-rayconv} In the setting of \Cref{II:flat:prop-zform}, \[ I''(\vartheta)\ge\frac12\int_{l_0}^1K(l)^2\int_\R p_l\,c_y^2\,\dd y\,\dd l\ \ge0 . \] Consequently, for every $t\ge0$, $\vartheta\mapsto F_t(\zeta_\vartheta)$ is convex on $[0,1]$. \end{corollary} \begin{proof} We show that each term in \eqref{II:flat:eq-zform} is nonnegative. By \eqref{II:flat:eq-Z} and the Feynman--Kac formula along the diffusion of \Cref{II:flat:def-density}, \[ \mathcal Z(l,y)=-\frac12\,\E\int_{l_0}^lK(\lambda)\,c(\lambda,Y_\lambda)^2\dd\lambda, \] where $Y$ starts at $(l,y)$ and runs downward in $l$. By \Cref{II:flat:lem-signs}, $c^2$ is nonincreasing in $y$, and the flow $y\mapsto Y_\lambda$ is increasing; since $K\ge0$, $\mathcal Z$ is nondecreasing in $y$, so $\mathcal Z_y\ge0$. By \Cref{II:flat:prop-mono}, $(p_lc)_y\le0$. Hence $-2K(p_lc)_y\mathcal Z_y\ge0$, and the other two terms are nonnegative. The functional $A$ is affine in $T$ by \Cref{II:flat:prop-inverse}, so $F_t=A+tI$ is convex along the segment for $t\ge0$. \end{proof} Only the ordering $K\ge0$ enters the proof. For directions $K$ of both signs the middle term of \eqref{II:flat:eq-zform} has no sign, and we do not know whether $I$ is convex along such segments; the ordered rays from the replica-symmetric profile suffice for \Cref{II:thm:flatuniq}. \Cref{II:flat:lem-signs,II:flat:prop-mono} hold for every positive step function $\sigma$, while the proof of \Cref{II:flat:prop-zform} uses the tail estimates of \Cref{II:flat:lem-tails}, which are proved for profiles. This convexity differs from the classical convexity of Parisi functionals: of the Crisanti--Sommers functional for spherical models \cite{Talagrand2006Spherical}, and of the Parisi functional for Ising spins \cite[Theorem~2]{AuffingerChen2015Uniqueness} (see \Cref{II:sec:intro}). Those results concern linear interpolation of distribution functions, and the proofs for Ising spins use the convexity of the terminal datum $\log\cosh$ in the spatial variable. The terminal datum $\log\Phi$ of \eqref{II:eq:intro-pde} is concave (see the opening of \Cref{II:sec:functional}), so these arguments do not apply to $\cP_\kappa$. \Cref{II:flat:cor-rayconv} concerns only the flat functional, interpolates linearly in the inverse profile $T$ rather than in the profile, and is proved only along ordered segments. \section{The first variation at the replica-symmetric profile}\label{II:flat:ss-fs} By \Cref{II:flat:cor-rayconv}, $F_{\td}(\zeta)-F_{\td}(\1_{[0,1]})$ is bounded below by the one-sided derivative of $F_{\td}$ at $T=\mathrm{id}$ in the direction $K=T-\mathrm{id}$. We now compute this derivative and show that it is positive, with cubic contact at the edge of the admissible cone. At $T=\mathrm{id}$ the solution is $U(l,y)=\log\Phi(y/\sqrt l)$, and the diffusion of \Cref{II:flat:def-density} is the Doob transform of Brownian motion by $\Phi(\cdot/\sqrt l)$. In the variables \[ Z=Y_l/\sqrt l,\qquad \varsigma=\log(1/l), \] It\^o's formula gives the time-homogeneous diffusion \begin{equation}\label{II:flat:eq-topdiff} \dd Z=\bigl(R(Z)+\tfrac Z2\bigr)\dd\varsigma+\dd B_\varsigma,\qquad \mathcal G_{\mathrm{top}}=\tfrac12\partial_{zz}+\bigl(R(z)+\tfrac z2\bigr)\partial_z . \end{equation} It describes the normalized position of the Parisi diffusion above the atom of a replica-symmetric order parameter, which is why it reappears in \Cref{II:sec:frsb}; its transition kernel is computed in \Cref{II:appD:lem-kernel}. We write $Z^z$ for the solution started at $Z_0=z$. For $0<\rho<1$ define \begin{equation}\label{II:flat:eq-calI} \mathcal Q(\rho)=\int_\R\Bigl[V(z)^2-\E\,V\bigl(Z^z_{\log(1/\rho)}\bigr)^2\Bigr]\dd z . \end{equation} This is the decrease of $\int V^2$ when Lebesgue measure is transported by the diffusion \eqref{II:flat:eq-topdiff} for the time $\log(1/\rho)$. The integrand is integrable by the Dynkin calculation in the proof below, although $V^2$ itself is not integrable. \begin{proposition}\label{II:flat:prop-fs} Let $T$ be the inverse profile of a normalized step profile, let $K=T-\mathrm{id}$, and let $\zeta_\vartheta$ be the normalized profile with inverse profile $\mathrm{id}+\vartheta K$. The function $\dot\sigma=K'=\sigma-1$ is a nonnegative nondecreasing step function that vanishes on $[0,l_0]$, and \begin{equation}\label{II:flat:eq-firstvar} \frac{\dd}{\dd\vartheta}F_{\td}(\zeta_\vartheta)\Big|_{\vartheta=0} =\int_{(0,1]}\mathcal E(r)\,\dd\dot\sigma(r),\qquad \mathcal E(r)=\frac1{8c_0}\int_r^1\mathcal Q(\rho)\,\frac{\rho-r}{\rho^2}\,\dd\rho , \end{equation} where $\dd\dot\sigma$ is the Stieltjes measure of $\dot\sigma$. Moreover $\mathcal Q>0$ on $(0,1)$, and with $c_{\mathrm{FS}}=1/96$, \begin{equation}\label{II:flat:eq-FS} \mathcal E(r)\ge c_{\mathrm{FS}}(1-r)^3\qquad(00\}$ therefore allows us to integrate first in $y$. Since $\int_\R(\1_{\{y\le0\}}-V)=0$ by \Cref{II:fun:gauss}(f), \[ I'(0)=\frac12\int_0^1\frac{\dot\sigma(l)}l\,M(l)\dd l,\qquad M(l):=\int_\R\Bigl[V(y)-\E V\bigl(Z^y_{\varsigma_l}\bigr)\Bigr]\dd y . \] By \eqref{II:flat:eq-Aaffine}, $A(\zeta_\vartheta)=\frac18-\frac\vartheta8\int_0^1\dot\sigma(l)(1-l)l^{-1}\dd l$. Since $\td/2=1/(8c_0)$, \begin{equation}\label{II:flat:eq-e} \frac{\dd}{\dd\vartheta}F_{\td}(\zeta_\vartheta)\Big|_{\vartheta=0}=\int_0^1\dot\sigma(l)\,\mathfrak e(l)\dd l,\qquad \mathfrak e(l)=\frac{M(l)-c_0(1-l)}{8c_0\,l}. \end{equation} \proofstep{2}{A Dynkin calculation for the transported Lebesgue measure.} The drift $R(z)+z/2$ of \eqref{II:flat:eq-topdiff} has derivative $\frac12-V\in(-\frac12,\frac12)$, so the flow $z\mapsto Z^z_\varsigma$ is an increasing $C^1$ bijection with Jacobian in $[\e^{-\varsigma/2},\e^{\varsigma/2}]$. Let $\nu_\varsigma$ be the density of the image of Lebesgue measure under this flow; then $\nu_\varsigma\le\e^{\varsigma/2}$. Using the identity \eqref{II:fun:Vpp} for $V''$, \begin{equation}\label{II:flat:eq-AV} \begin{split} \mathcal G_{\mathrm{top}}V&=-V(1-V),\qquad \mathcal G_{\mathrm{top}}(V^2)=k_1:=V'^2-2V^2(1-V),\\ \mathcal G_{\mathrm{top}}\bigl(V(1-V)\bigr)&=-V(1-V)-k_1 . \end{split} \end{equation} All three right sides are bounded and integrable by \Cref{II:appC:lem-gauss}(a). For $f\in\{V,V^2,V(1-V)\}$, Dynkin's formula for each starting point, and Fubini's theorem (justified by $\nu_\varsigma\le\e^{\varsigma/2}$ and $\mathcal G_{\mathrm{top}}f\in L^1$), give \[ \int_\R\bigl[\E f(Z^z_\varsigma)-f(z)\bigr]\dd z=\int_0^\varsigma\!\!\int_\R\nu_u\,\mathcal G_{\mathrm{top}}f\,\dd z\,\dd u . \] Put $\mathfrak j(u)=\int\nu_uV(1-V)$ and $\mathfrak k(u)=\int\nu_uk_1$, which are bounded on bounded intervals. With $f=V$, $f=V^2$, and $f=V(1-V)$ this gives \[ M(l)=\int_0^{\varsigma_l}\mathfrak j(u)\dd u,\qquad \mathcal Q(\rho)=-\int_0^{\varsigma_\rho}\mathfrak k(u)\dd u,\qquad \mathfrak j(u)=c_0-\int_0^u\bigl(\mathfrak j+\mathfrak k\bigr). \] The last identity is the integral form of $\mathfrak j'=-\mathfrak j-\mathfrak k$, $\mathfrak j(0)=c_0$, so $\mathfrak j(u)=c_0\e^{-u}-\int_0^u\e^{u'-u}\mathfrak k(u')\dd u'$. Integrating over $[0,\varsigma_l]$ and applying Fubini's theorem twice (with $\e^{-\varsigma_l}=l$), \[ M(l)-c_0(1-l)=-\int_0^{\varsigma_l}\mathfrak k(u')\bigl(1-l\,\e^{u'}\bigr)\dd u' =l\int_l^1\mathcal Q(\rho)\rho^{-2}\dd\rho . \] The second equality follows by inserting the formula for $\mathcal Q$, since $u'\le\varsigma_\rho$ if and only if $\rho\le\e^{-u'}$, and $l\int_l^{\e^{-u'}}\rho^{-2}\dd\rho=1-l\,\e^{u'}$. Hence $\mathfrak e(l)=(8c_0)^{-1}\int_l^1\mathcal Q(\rho)\rho^{-2}\dd\rho$. Since $\dot\sigma(l)=\int_{(0,l]}\dd\dot\sigma(r)$, Fubini's theorem turns \eqref{II:flat:eq-e} into \eqref{II:flat:eq-firstvar} with $\mathcal E(r)=\int_r^1\mathfrak e(l)\dd l=(8c_0)^{-1}\int_r^1\mathcal Q(\rho)(\rho-r)\rho^{-2}\dd\rho$. \proofstep{3}{One comparison gives positivity and the cubic bound.} Couple \eqref{II:flat:eq-topdiff}, started at $z$, with the Ornstein--Uhlenbeck process $\dd X=-\frac12X\dd\varsigma+\dd B_\varsigma$, $X_0=z$, driven by the same Brownian motion. The difference $\Xi=Z-X$ satisfies $\Xi'=R(Z)+\frac12Z+\frac12X=\Rbar(Z)-\frac12\Xi$ and $\Xi_0=0$, so \[ \Xi_\varsigma=\int_0^\varsigma\e^{-(\varsigma-u)/2}\,\Rbar(Z_u)\dd u>0, \] because $\Rbar>0$ (\Cref{II:fun:gauss}(a)). Thus $Z_\varsigma\ge X_\varsigma$, and $X_\varsigma\sim N(\sqrt\rho\,z,1-\rho)$ with $\rho=\e^{-\varsigma}$. This is a pathwise comparison of the type of \cite{IkedaWatanabe1977}, made explicit here by the linear equation for $\Xi$. Since $V^2$ is decreasing, with $P_v$ the convolution with the $N(0,v)$ density, \[ \E\,V(Z^z_\varsigma)^2\le P_{1-\rho}(V^2)(\sqrt\rho\,z). \] This comparison can be integrated despite the nonzero left limit of $V^2$. Put $f=V^2-\1_{\{z<0\}}$. By \Cref{II:fun:gauss}(f), $f\in L^1$ and $\int f=-c_0$. The function $\1_{\{z<0\}}-P_{1-\rho}\1_{\{\cdot<0\}}(\sqrt\rho\,z) =\1_{\{z<0\}}-\Phi(-az)$, $a=\sqrt{\rho/(1-\rho)}$, is integrable with integral $\int_{-\infty}^0\Phi(az)\dd z-\int_0^\infty\Phi(-az)\dd z=0$. The substitution $w=\sqrt\rho\,z+\sqrt{1-\rho}\,g$ shows $\int P_{1-\rho}f(\sqrt\rho\,z)\dd z=\rho^{-1/2}\int f$. Hence \[ \int_\R\Bigl[V(z)^2-P_{1-\rho}(V^2)(\sqrt\rho\,z)\Bigr]\dd z =\bigl(1-\rho^{-1/2}\bigr)\int_\R f =c_0\bigl(\rho^{-1/2}-1\bigr). \] The integrand of \eqref{II:flat:eq-calI} is integrable by Step~2 and pointwise at least the integrand on the left. Since $\rho^{-1/2}-1\ge1-\sqrt\rho=(1-\rho)/(1+\sqrt\rho)\ge(1-\rho)/2$, \begin{equation}\label{II:flat:eq-ougain} \mathcal Q(\rho)\ge c_0\bigl(\rho^{-1/2}-1\bigr)\ge\tfrac{c_0}{2}(1-\rho)>0\qquad(0<\rho<1). \end{equation} With the formula for $\mathfrak e$ from Step~2 and $\rho^{-2}\ge1$, \[ \mathfrak e(l)\ge\frac1{16}\int_l^1(1-\rho)\dd\rho=\frac{(1-l)^2}{32},\qquad \mathcal E(r)=\int_r^1\mathfrak e(l)\dd l\ge\frac{(1-r)^3}{96}. \] This proves \eqref{II:flat:eq-FS}. \end{proof} The number $\mathcal E(r)$ is the derivative of $F_{\td}$ in the one-step direction $K=(l-r)_+$. Its positivity for every $r$ is the flat counterpart of the statement that the obstacle potential of the replica-symmetric crossing order parameter is strictly increasing above its atom (\Cref{II:frsb:prop-rsobstacle}). The contact at $r=1$ is cubic because at $t=\td$ the term of order $1-l$ in $\mathfrak e(l)$ cancels, which is the flat analogue of the vanishing of the replicon eigenvalue at $\kd$. More precisely, $\mathfrak k$ is continuous at $0$ with $\mathfrak k(0)=\int k_1=A_V-2B_V$, where $A_V=\int V'^2$ and $B_V=\int V^2(1-V)$: for $f\in C_c^2$, Dynkin's formula gives $|\int\nu_uf-\int f|\le u\,\e^{u/2}\|\mathcal G_{\mathrm{top}}f\|_{L^1}$, and $k_1\in L^1$ is approximated by such $f$ while $\nu_u\le\e^{u/2}$. With $x_{\mathrm f}=A_V/(2B_V)$ the flat breakpoint of \Cref{II:frsb:def-curv}, this gives $\mathcal Q(\rho)=2B_V(1-x_{\mathrm f})(1-\rho)+o(1-\rho)$ and \[ \mathcal E(r)=\frac{B_V(1-x_{\mathrm f})}{24\,c_0}\,(1-r)^3+o\bigl((1-r)^3\bigr)\qquad(r\uparrow1). \] We do not use this expansion. The same computation at a general parameter $t$ shows that $\td$ is sharp. By Step~1, the derivative of $F_t(\zeta_\vartheta)$ at $\vartheta=0$ in the direction $K=(l-r)_+$ equals $\int_r^1[tM(l)-\frac14(1-l)]\,(2l)^{-1}\dd l$, and $M(l)=c_0(1-l)+O((1-l)^2)$ by Step~2 (because $|\mathfrak k|\le\e^{1/2}\|k_1\|_{L^1}$ on $[0,1]$). For $t<\td$ this derivative is negative for $r$ close to one, so $\1_{[0,1]}$ is not a minimizer of $F_t$. This is the flat counterpart of the instability of the replica-symmetric order parameter beyond $\kd$ (\Cref{II:prop:crossing}(ii)). \section{Uniqueness of the flat minimizer}\label{II:flat:ss-uniq} We now combine the ray convexity with the first variation. The cubic contact of \eqref{II:flat:eq-FS} becomes a cubic $L^1$ modulus. \begin{theorem}\label{II:thm:flatuniq} Let $\zeta$ be a normalized profile with bounded support and $\zeta=1$ on a neighborhood of $0$, that is, $\zeta\in\bigcup_{00$, $\vartheta\mapsto F_{\td}(\zeta_\vartheta)$ is convex on $[0,1]$, where $\zeta_\vartheta$ has inverse profile $\mathrm{id}+\vartheta K$; it is differentiable by \Cref{II:flat:prop-zform}. By \Cref{II:flat:prop-fs}, \[ F_{\td}(\zeta)-F_{\td}(\1_{[0,1]})\ge\frac{\dd}{\dd\vartheta}F_{\td}(\zeta_\vartheta)\Big|_{\vartheta=0} \ge\frac1{96}\int_{(0,1]}(1-r)^3\dd\dot\sigma(r). \] \proofstep{2}{Comparison with the $L^1$ distance.} Let $l_*=\ell(1)\in(0,1]$, so that $T(l_*)=1$. Since $\int\zeta=1=\int\1_{[0,1]}$, \[ \tfrac12\bigl\|\zeta-\1_{[0,1]}\bigr\|_{L^1}=\int_0^1(1-\zeta)=1-l_* . \] From $T(l)=l+\int_0^l\dot\sigma$ and $T(l_*)=1$, $1-l_*=\int_0^{l_*}\dot\sigma(u)\dd u=\int_{(0,1]}(l_*-r)_+\dd\dot\sigma(r)$. For $r\le l_*$ put $a=1-l_*$ and $b=l_*-r$. Then $(1-r)^3=(a+b)^3\ge\frac{27}4a^2b$, because $(a+b)^3-\frac{27}4a^2b=(b-\frac a2)^2(b+4a)\ge0$. Therefore \[ \int_{(0,1]}(1-r)^3\dd\dot\sigma(r)\ge\frac{27}4(1-l_*)^2\int_{(0,1]}(l_*-r)_+\dd\dot\sigma(r) =\frac{27}4(1-l_*)^3=\frac{27}{32}\bigl\|\zeta-\1_{[0,1]}\bigr\|_{L^1}^3 . \] With Step~1 and $\frac{27}{32}\cdot\frac1{96}=\frac9{1024}$, this proves \eqref{II:flat:eq-cubicmod} for step profiles. \proofstep{3}{General profiles and the parameters $t\ge\td$.} Let $\zeta\in\cK^1(d_1,S)$. By \eqref{II:flat:eq-ellbounds}, $d_1\le1\le S$, so $\1_{[0,1]}\in\cK^1(d_1,S)$. Let $\zeta_n\in\cK^1(d_1,S)$ be the step approximations \eqref{II:flat:eq-steps} of $\zeta$. By \Cref{II:flat:lem-tails}, $A$ and $\zeta\mapsto\int W_\zeta$ are continuous on $\cK^1(d_1,S)$, hence so is $F_{\td}$. Both sides of \eqref{II:flat:eq-cubicmod} are continuous in $L^1$, so the inequality passes to the limit along $\zeta_n\to\zeta$. Next, $\td\,(I(\zeta)-I(\1_{[0,1]}))=F_{\td}(\zeta)-F_{\td}(\1_{[0,1]})+A(\1_{[0,1]})-A(\zeta)$, and $A(\zeta)\le A(\1_{[0,1]})$ by \Cref{II:flat:cor-Amax}; hence $I(\zeta)-I(\1_{[0,1]})\ge\frac9{1024\,\td}\|\zeta-\1_{[0,1]}\|_{L^1}^3\ge0$. For $t\ge\td$, \[ F_t(\zeta)-F_t(\1_{[0,1]})=F_{\td}(\zeta)-F_{\td}(\1_{[0,1]})+(t-\td)\bigl(I(\zeta)-I(\1_{[0,1]})\bigr) \ge\frac9{1024}\,\frac t{\td}\,\bigl\|\zeta-\1_{[0,1]}\bigr\|_{L^1}^3 , \] which is \eqref{II:flat:eq-tmod} because $t/\td=4c_0t$. \end{proof} The constant in \eqref{II:flat:eq-cubicmod} is $\frac{27}{32}c_{\mathrm{FS}}$. The factor $\frac{27}{32}$ comes from Step~2 and cannot be improved there: for the one-step directions $\dot\sigma=2\cdot\1_{(r,1]}$ one has $l_*-r=\frac12(1-l_*)$, and the two sides of the inequality in Step~2 are equal. Without the elementary inequality $(a+b)^3\ge\frac{27}4a^2b$, the same argument gives the constant $\frac18c_{\mathrm{FS}}=\frac1{768}$. The theorem is a statement about one explicit variational problem, in which neither $e$ nor $\theta$ appears; its link with the functional $\cP_\kappa$ at finite $e$ is \Cref{II:thm:compactexp}. \section{Stability at the crossing}\label{II:flat:ss-conv-cross} From now on we assume, as announced in \Cref{II:rs:rem-Theta}, that $\thd$ lies in the interior of $\Theta$. Put $\theta_e=(\kRS-\kd)/e^2$, so that $\kappa_e(\theta_e)=\kd$. By \Cref{II:prop:crossing}(iii), $\theta_e\to\thd$; hence $\theta_e\in\Theta$ for small $e$, and \Cref{II:prop:crossing}(iv) applies at $\theta=\theta_e$, where $q^*(\kd)=\qd$. Together, \begin{equation}\label{II:flat:eq-crossing} \begin{split} \theta_e&=\frac{\kRS-\kd}{e^2}=\thd+O(eL_e),\\ d_{\mathrm d}&=\frac{1-\qd}{e^2}=2\phi_0\theta_e+O(e)\longrightarrow d_0=2\phi_0\thd=\frac1{64c_0^2\phi_0^2}. \end{split} \end{equation} The following estimate controls the mass and the shape of a profile at once. It is uniform in $\theta$ near $\thd$, which is what the uniform window of \Cref{II:sec:frsb} needs, and at $\theta=\theta_e$ it gives the convergence of minimizers at the crossing. \begin{proposition}\label{II:flat:cor-nearmin} Fix $00$, depending only on $(m,M)$ and $\Theta$, such that the following holds for $00$ with the following property for all small $e=\alpha-2>0$. For every margin in $[\kd-ce^2,\kd]$, the replica-symmetric order parameter is the unique global minimizer. For every margin in $(\kd,\kd+ce^2)$, every global minimizer has a Parisi measure supported on one nondegenerate interval, with atoms at both ends and a distribution function that is continuously differentiable on the interior, where its derivative is of order $e^{-2}$. The case $\kappa=\kd$ is recorded separately in \Cref{II:thm:uniqd}. The proof is Step~5 of \Cref{II:sec:intro-ideas}. \Cref{II:frsb:ss-kkt} writes the KKT condition through anchored obstacles, which involve only $S$ and $\lambda$, and defines the local curvature $\widetilde\Sigma$. \Cref{II:frsb:ss-rs} treats the obstacle potential of the crossing order parameter $\gamma_{\mathrm d}=\1_{[\qd,1]}$ at $\kd$, which has a strict cubic minimum at $\qd$ of depth proportional to $e$, and the replica-symmetric family near the crossing. \Cref{II:frsb:ss-nat} proves estimates at the natural scale in the crossing window. In \Cref{II:frsb:ss-loc}, the stability estimate \Cref{II:flat:cor-nearmin} gives closeness of every global minimizer to $\gamma_{\mathrm d}$ in $L^1$, on the scale $1-\qd\asymp e^2$ (\Cref{II:ext:cor-flat}), and a comparison of anchored obstacles upgrades this to localization of the supports (\Cref{II:prop:localize}). \Cref{II:frsb:ss-window} shows that a KKT point with support in a small window has no gap and satisfies a width estimate (\Cref{II:prop:window}), and \Cref{II:frsb:ss-uniform} combines the two steps. The relation to earlier work is discussed at the end of the chapter. Throughout this chapter, $\alpha=2+e$, $\kd=\kd(\alpha)$, and $\qd$, $\delta_{\mathrm d}=1-\qd$, and $\gamma_{\mathrm d}=\1_{[\qd,1]}$ are as in \Cref{II:prop:crossing}. We write \[ \hat s(q)=\frac{1-q}{\delta_{\mathrm d}},\qquad \cW_\eta=\bigl\{q\in[0,1]:|q-\qd|\le\eta e^2\bigr\}. \] Thus $\hat s(\qd)=1$, and $\cW_\eta$ is a window of fixed rescaled width around the atom of $\gamma_{\mathrm d}$. In this chapter we fix the window $\Theta=[\thd/2,2\thd]$ of \Cref{II:rs:rem-Theta}, which contains $\thd$ in its interior, so that the results of \Cref{II:sec:apriori,II:sec:depth,II:sec:compact,II:sec:flat} apply to it, and every constant that depends on $\Theta$ is numerical. The distance from $\thd$ to $\R\setminus\Theta$ is $\Delta_\Theta=\thd/2$. Recall from \Cref{II:sec:flat} that $L_e=\log(1/e)$, $\theta_e=(\kRS-\kd)/e^2$, $d_{\mathrm d}=\delta_{\mathrm d}/e^2$, and $d_0=2\phi_0\thd=1/(64c_0^2\phi_0^2)$. By \eqref{II:flat:eq-crossing}, which rests on \Cref{II:prop:crossing}(iii),(iv), there are numerical constants $C_{\mathrm{cr}}$ and $C_\Theta$ such that \begin{equation}\label{II:ext:eq-cross} |\theta_e-\thd|\le C_{\mathrm{cr}}\,eL_e,\qquad |d_{\mathrm d}-2\phi_0\theta_e|\le C_\Theta\,e \end{equation} for all small $e$. We fix $e_{\rm st}\in(0,1)$ below the thresholds of \Cref{II:prop:crossing,II:thm:ordering,II:thm:T1} and of \eqref{II:ext:eq-cross}, and such that for $00 . \] \end{lemma} The proof is in \Cref{II:appD:ss-const}. The breakpoint and the curvature enter through the derivatives of $S$. By \Cref{II:lem:generator}(c)(iv),(d), $S$ is Lipschitz on $[0,r]$ for every $r<1$, with one-sided derivatives \begin{equation}\label{II:frsb:eq-Sprime} S'(q+)=\frac{2\alpha\widetilde B(q)}{\lambda(q)}\bigl(\widetilde x(q)-\gamma(q)\bigr),\qquad S'(q-)=\frac{2\alpha\widetilde B(q)}{\lambda(q)}\bigl(\widetilde x(q)-\gamma(q-)\bigr), \end{equation} and on an interval where $\gamma\equiv m$ is constant, $S''=\alpha\Sigma(q;m/\lambda)$. The functions $\cN$ and $\widetilde B$ are Lipschitz on $[0,r]$, and $\widetilde\Sigma$ is continuous there, by \Cref{II:fun:diffusion}(c); since $\widetilde B>0$ is continuous, $\widetilde x$ is Lipschitz on $[0,r]$. The next lemma restates the gap principle \Cref{II:lem:gap} in terms of $\widetilde\Sigma$ and adds the density formula on support intervals. The right and left derivatives in \eqref{II:frsb:eq-Sprime} differ only through $\gamma(q)$ and $\gamma(q-)$, as in the analogue of \cite[Proposition~3]{AuffingerChenProperties} noted after \Cref{II:lem:generator}; for spherical models the density on support intervals is given explicitly in \cite[Theorem~6(ii)]{AuffingerChenProperties}. \begin{lemma}\label{II:frsb:lem-gapint} Let $\gamma$ be a KKT point of $\cP_\kappa$. \begin{enumerate}[label=\textup{(\alph*)}] \item If $p0$ such that the following holds for $00$ on $[0,\qd)$, $D_{\mathrm d}<0$ on $(\qd,1)$, and $\qd$ is the unique minimum point of $\cH_{\mathrm d}$ on $[0,1)$. \item For every $q\in[0,1)$, \[ G_{\mathrm d}(q)\ge c_{\mathrm o}\,e\,\min\bigl\{|\hat s(q)-1|,\tfrac34\bigr\}^3 . \] \item If $|\hat s(q)-1|\le\frac34$, then \begin{gather*} 0\le1-S_{\mathrm d}(q)\le C_{\mathrm o}e|\hat s(q)-1|,\qquad G_{\mathrm d}(q)\le C_{\mathrm o}e|\hat s(q)-1|^3,\\ |\delta_{\mathrm d}D_{\mathrm d}(q)|\le C_{\mathrm o}e|\hat s(q)-1|^2 . \end{gather*} \end{enumerate} \end{proposition} The proof, in \Cref{II:appD:ss-rs}, is explicit. Above $\qd$ the law of $(X_q-\kd)/\sqrt{1-q}$ evolves by the transition kernel of the diffusion \eqref{II:flat:eq-topdiff}, which is explicit and reversible for the weight $\Phi^2\e^{z^2/2}$ (\Cref{II:appD:lem-kernel}). Near $\alpha=2$ the initial density $\beta_{\mathrm d}\phi(k_{\mathrm d}+\beta_{\mathrm d}z)$, with $k_{\mathrm d}=\kd/\sqrt{\qd}$ and $\beta_{\mathrm d}=\sqrt{\delta_{\mathrm d}/\qd}$, is this weight times a strictly decreasing function, and a stochastic-order comparison shows that $S_{\mathrm d}=\alpha\E V^2$ decreases. The lower bounds in (b) come from exact positive formulas for $1-S_{\mathrm d}$ on both sides of $\qd$. They carry the factor $\beta_{\mathrm d}\asymp e$, and the potential is cubic because the replicon eigenvalue vanishes at the crossing. We also need the replica-symmetric family near the crossing. For $q\in(0,1)$ and a margin $\kappa$ let $\gamma_q=\1_{[q,1]}$, $\delta_q=1-q$, and $Z^q=(\sqrt q\,G-\kappa)/\sqrt{\delta_q}$ be as in \Cref{II:def:rs}, let $D_{\RS}(q)=D_{\RS}(q;\kappa)$ and $S_{\RS}(q)=S_{\RS}(q;\kappa)$ be the values of $D$ and $S$ of $\gamma_q$ at its atom, given explicitly in \eqref{II:rs:eq-DRS}, and put \[ A_w(q)=\E V'(Z^q)^2 . \] The order parameter $\gamma_q$ is a KKT point only if $D_{\RS}(q)=0$. For $\kappa<\kRS$ let $q^*(\kappa)$ be the stationary point of \Cref{II:prop:crossing}(i), and put $\tau(\kappa)=1-S_{\RS}(q^*(\kappa))$, the rescaled replicon eigenvalue. By \Cref{II:rs:lem-formulas}(b) and \Cref{II:def:rs}, $\tau(\kappa)=4(1-q^*(\kappa))^2\Lambda_{\mathrm R}(\kappa)$, so \Cref{II:prop:crossing}(ii) gives \begin{equation}\label{II:ext:eq-tausign} \tau(\kappa)>0\quad(\kRS-c_2\le\kappa<\kd),\qquad \tau(\kd)=0,\qquad \tau(\kappa)<0\quad(\kd<\kappa<\kRS). \end{equation} \begin{lemma}\label{II:frsb:lem-rsfamily} There are $\eta_1\in(0,\frac14d_0]$, $\eta_1'\in(0,\eta_1]$, $e_{\mathrm r}>0$, and $c_{\mathrm r},C_{\mathrm r}>0$ such that the following holds for $00$ there are $\varepsilon,e_{\mathrm c}>0$ such that for $00$ such that, in this setting with $\eta=\eta_2$ and $00$ (\Cref{II:frsb:lem-flatconst}). Part (a) is used only at the margin $\kd$; the dependence on the margin is controlled separately by \Cref{II:ext:lem-margin} below. The next lemma compares the bottom level of an order parameter supported in a window with the replica-symmetric order parameter of the same mean. For $\gamma\in\cU$ with bottom $q_0$ write \[ m=1-\lambda(q_0)=\int t\dd\mu_\gamma(t),\qquad \delta_m=1-m,\qquad \tilde\ell=\frac{Q-q_0}{e^2}, \] and let $\gamma_m=\1_{[m,1]}$ be the replica-symmetric order parameter with the same mean; it has the same value $\lambda(q_0)$. \begin{lemma}\label{II:frsb:lem-bottom} There are $\eta_3,e_{\mathrm{bt}},C>0$ such that for $00$ there are $\rho_1(\varepsilon)\in(0,\frac12\Delta_\Theta]$ and $e_1(\varepsilon)>0$ such that for $00$ and $e_l(\eta)>0$ such that for $00$ with \begin{equation}\label{II:ext:eq-omega} \frac{8\omega(u_+-u_-)^2}{u_-^2}\le\frac{c_{\mathrm o}\eta^3}8 . \end{equation} Let $\varepsilon_0,e_{\mathrm c}$ be given by \Cref{II:frsb:lem-natcont}(a) for the interval $[u_-,u_+]$ and the tolerance $\frac12\omega$. Decrease $\varepsilon_0$ so that $\varepsilon_0\le\frac12u_-$, $(1+\varepsilon_0/u_-)^{-2}\ge\frac12$, and $C_{\mathrm o}(\varepsilon_0^3+(u_+-u_-)\varepsilon_0^2)\le\frac14c_{\mathrm o}\eta^3$. Let $\rho_1=\rho_1(\varepsilon_0)$ and $e_1=e_1(\varepsilon_0)$ be given by \Cref{II:ext:cor-flat}, and put $\rho_l=\min\{\rho_1,1\}$. Finally let $e_l\le\min\{e_*,e_{\mathrm o},e_{\mathrm c},e_1,\frac14\omega\}$ be so small that $u_+\delta_{\mathrm d}\le\frac12$ and $S_*e^2\le\frac12$ for $e0 . \] This is a contradiction. Therefore every support point satisfies $|\hat s-1|<\eta$. This proves the proposition. \end{proof} The constants are chosen in the order $\eta$, $\omega$, $(\varepsilon_0,e_{\mathrm c})$, $(\rho_1,e_1)$, $(\rho_l,e_l)$. The margin enters only through the KKT identity $\cA_\gamma(p,t)=0$ at $\kappa$ and through the error term $2|\kappa-\kd|$ in \eqref{II:ext:eq-Scont}. In particular no estimate of the replica-symmetric obstacle at margins other than $\kd$ is needed. \section{Window rigidity}\label{II:frsb:ss-window} We now analyze KKT points whose support lies in a window $\cW_\eta$ of small rescaled width. Let $\eta_1,\eta_1'$, $\eta_2$, and $\eta_3$ be the window sizes of \Cref{II:frsb:lem-rsfamily,II:frsb:lem-natcont,II:frsb:lem-bottom}; recall that $\eta_1\le\frac14d_0$. \begin{proposition}\label{II:prop:window} There are $\eta_0\in(0,\min\{\eta_1',\eta_2,\eta_3,\frac14d_0,\frac12\}]$, $e_{\mathrm{win}}>0$, and $c',C'>0$ such that for $00$; and $\mu_\gamma(\{q_0\})=\widetilde x(q_0)>\frac12x_{\mathrm f}$, $\mu_\gamma(\{Q\})=1-\widetilde x(Q)>\frac14$. Moreover, with $\tilde\ell=(Q-q_0)/e^2$ and $m$ the mean of $\mu_\gamma$, \begin{equation}\label{II:frsb:eq-width} c'e\tilde\ell\le-\tau(\kappa),\qquad |m-q^*(\kappa)|\le C'e^3\tilde\ell . \end{equation} \item If $q_0=Q$, then $q_0=q^*(\kappa)$ and $\tau(\kappa)\ge0$. \item If $\kappa\le\kd$, then $\gamma=\1_{[q^*(\kappa),1]}$. If $\kappa>\kd$, then $q_00$. Hence the support has no gap. It is a closed set that contains every point between its minimum and its maximum, that is, $[q_0,Q]$. (b) The structure statements are \Cref{II:frsb:lem-gapint}(b) with $p=q_0$ and $p'=Q$, together with $\widetilde\Sigma>0$ and $\frac12x_{\mathrm f}<\widetilde x<\frac34$ from \eqref{II:frsb:eq-winconst}. Since $\gamma=0$ on $[0,q_0)$, $\mu_\gamma(\{q_0\})=\gamma(q_0)=\widetilde x(q_0)$; since $\widetilde x$ is continuous, $\mu_\gamma(\{Q\})=1-\gamma(Q-)=1-\widetilde x(Q)$. For the width estimate we record three bounds and then proceed in three steps. First, $m=1-\lambda(q_0)=q_0+\int_{q_0}^Q(1-\gamma)$ and $\gamma<\frac34$ on $[q_0,Q)$, so \[ \tfrac14e^2\tilde\ell\le m-q_0\le e^2\tilde\ell . \] Second, $m\in[q_0,Q]\subset\cW_{\eta_0}$ and $\eta_0\le\frac14d_0$, so $\delta_m\in[\frac14d_0e^2,3d_0e^2]$ and $m\in\cW_{\eta_1}$. Third, $\tilde\ell\le2\eta_0\le1$. \proofstep{1}{The $S$-sector.} Since $q_0$ is a support point, $S_\gamma(q_0)=1$. By \Cref{II:frsb:lem-bottom}, $|S_{\RS}(m)-1-\alpha(m-q_0)A_w(m)/\delta_m|\le C_{\mathrm{bt}}e\tilde\ell^2$. By \Cref{II:frsb:lem-rsfamily}(c) and the three bounds, with $2\le\alpha\le3$, \[ \frac{c_{\mathrm r}}{6d_0}e\tilde\ell\le\alpha(m-q_0)\frac{A_w(m)}{\delta_m}\le\frac{12C_{\mathrm r}}{d_0}e\tilde\ell . \] Hence, with $c_1=c_{\mathrm r}/(6d_0)$ and $C_1=12C_{\mathrm r}/d_0+C_{\mathrm{bt}}$, \begin{equation}\label{II:frsb:eq-Ssector} S_{\RS}(m)-1\ge e\bigl(c_1\tilde\ell-C_{\mathrm{bt}}\tilde\ell^2\bigr),\qquad |S_{\RS}(m)-1|\le C_1e\tilde\ell . \end{equation} \proofstep{2}{The mean.} Since $D_\gamma(q_0)=0$, \Cref{II:frsb:lem-bottom} gives $|D_{\gamma_m}(q_0)|\le C_{\mathrm{bt}}\tilde\ell^2/e$ and $\sup_{[q_0,m]}|S_{\gamma_m}-1|\le(C_{\mathrm{bt}}+C_1)e\tilde\ell$. By the formula for $D_{\gamma_m}(q_0)$ in that lemma, \[ |D_{\RS}(m)|\le\frac{C_{\mathrm{bt}}\tilde\ell^2}e+\frac{m-q_0}{\delta_m^2}(C_{\mathrm{bt}}+C_1)e\tilde\ell\le\frac{C_2\tilde\ell}e,\qquad C_2=C_{\mathrm{bt}}+\frac{16(C_{\mathrm{bt}}+C_1)}{d_0^2}. \] Both $m$ and $q^*(\kappa)$ lie in the interval $\cW_{\eta_1}$, on which $\partial_qD_{\RS}\le-c_{\mathrm r}/\delta_q^2\le-c_{\mathrm r}/(3d_0e^2)^2$, because $\delta_q\le\delta_{\mathrm d}+\eta_1e^2\le3d_0e^2$ there; and $D_{\RS}(q^*(\kappa))=0$. By the mean value theorem, \[ |m-q^*(\kappa)|\le\frac{9d_0^2e^4}{c_{\mathrm r}}|D_{\RS}(m)|\le C'e^3\tilde\ell,\qquad C'=\frac{9d_0^2C_2}{c_{\mathrm r}} . \] \proofstep{3}{Conclusion.} By \Cref{II:frsb:lem-rsfamily}(b), $|S_{\RS}(m)-S_{\RS}(q^*(\kappa))|\le(C_{\mathrm r}/e)|m-q^*(\kappa)|\le C_{\mathrm r}C'e^2\tilde\ell$. With \eqref{II:frsb:eq-Ssector}, \[ -\tau(\kappa)=S_{\RS}(q^*(\kappa))-1\ge e\tilde\ell\bigl(c_1-C_{\mathrm{bt}}\tilde\ell-C_{\mathrm r}C'e\bigr). \] The constants $c_1,C_{\mathrm{bt}},C_{\mathrm r},C'$ do not depend on $\eta_0$. Now decrease $\eta_0$ so that $2C_{\mathrm{bt}}\eta_0\le\frac14c_1$, and $e_{\mathrm{win}}$ so that $C_{\mathrm r}C'e_{\mathrm{win}}\le\frac14c_1$. Since $\tilde\ell\le2\eta_0$, this gives \eqref{II:frsb:eq-width} with $c'=\frac12c_1$. (c) If $\supp\mu_\gamma=\{q_0\}$, then $\gamma=\1_{[q_0,1]}$, and $D(q_0)=0$ says $D_{\RS}(q_0)=0$. Thus $q_0$ is RS-stationary at $\kappa\in[\kRS-c_2,\kRS)$, and \Cref{II:prop:crossing}(i) gives $q_0=q^*(\kappa)$. Finally $S(q_0)\le1$ says $\tau(\kappa)\ge0$. (d) Let $\kappa\le\kd$. Then $\tau(\kappa)\ge0$ by \eqref{II:ext:eq-tausign}. If $q_00$ and \eqref{II:frsb:eq-width} gives $0\kd$. Then $\tau(\kappa)<0$ by \eqref{II:ext:eq-tausign}, since $\kappa<\kRS$. If $q_0=Q$, then (c) gives $\tau(\kappa)\ge0$, a contradiction. Hence $q_00$ and $C<\infty$ such that the following holds for $0\tfrac12x_{\mathrm f},\quad 1-\widetilde x(Q)>\tfrac14 . \] Moreover, \begin{equation}\label{II:ext:eq-rates} Q-q_0\le C(\kappa-\kd),\qquad \sup_{t\in\supp\mu_\gamma}|t-\qd|\le C(\kappa-\kd),\qquad \|\gamma-\gamma_{\mathrm d}\|_{L^1}\le C(\kappa-\kd). \end{equation} Finally, for every $\varepsilon>0$ there is $\rho_{\mathrm a}(\varepsilon)>0$ such that if also $e<\rho_{\mathrm a}(\varepsilon)$ and $\kappa-\kd<\rho_{\mathrm a}(\varepsilon)e^2$, then $|\mu_\gamma(\{q_0\})-x_{\mathrm f}|\le\varepsilon$ and $|\mu_\gamma(\{Q\})-(1-x_{\mathrm f})|\le\varepsilon$. \end{enumerate} \end{theorem} \begin{proof} We divide the proof into four steps. \proofstep{1}{Choice of constants.} Let $\eta_0,e_{\mathrm{win}},c',C'$ be the constants of \Cref{II:prop:window}, and let $\eta_1',c_{\mathrm r},C_{\mathrm r}$ be those of \Cref{II:frsb:lem-rsfamily}. Put $\eta'=\eta_0/(2d_0)\in(0,\frac18]$, and let $\rho_l=\rho_l(\eta')$ and $e_l=e_l(\eta')$ be given by \Cref{II:prop:localize}. Put \[ c_W=\min\bigl\{\rho_l,\ \eta_0,\ \tfrac14\thd\bigr\}, \] and let $e_0\le\min\{e_{\mathrm{win}},e_l,e_{\rm st}\}$. All these constants are numerical. \proofstep{2}{Localization in the window.} Let $0c_We^2$, so $\kappa<\kc$ and $\cP_*(\kappa)>-\infty$ by \eqref{II:fun:kc}, and global minimizers of $\cP_\kappa$ exist by \Cref{II:fun:existence}. Let $\gamma$ be one. By \Cref{II:prop:localize} with $\eta'$, every support point $q$ satisfies $|\hat s(q)-1|<\eta'$, that is, $|q-\qd|<\eta'\delta_{\mathrm d}\le2d_0\eta'e^2=\eta_0e^2$. Hence \begin{equation}\label{II:ext:eq-window} \supp\mu_\gamma\subset\cW_{\eta_0},\qquad |\kappa-\kd|\le\eta_0e^2 . \end{equation} By \Cref{II:prop:kkt}(c), $\gamma$ is a KKT point of $\cP_\kappa$. Thus \Cref{II:prop:window} applies to $\gamma$. \proofstep{3}{Replica symmetry below the crossing.} Let $\kappa\in[\kd-c_We^2,\kd]$. By Step 2 and \Cref{II:prop:window}(d), every global minimizer equals $\1_{[q^*(\kappa),1]}$, and a global minimizer exists. Hence it is the unique global minimizer. Since $|\kappa-\kd|\le\eta_0e^2\le\eta_1'e^2$, \Cref{II:frsb:lem-rsfamily}(a) gives $|q^*(\kappa)-\qd|\le C_{\mathrm r}(\kd-\kappa)$. This proves (i). \proofstep{4}{Full replica symmetry breaking above the crossing.} Let $\kappa\in(\kd,\kd+c_We^2)$ and let $\gamma$ be a global minimizer. By \Cref{II:prop:window}(d), $q_0\tfrac12x_{\mathrm f},\qquad \mu_\gamma(\{Q\})=1-\widetilde x(Q)>\tfrac14 . \] Since $\gamma$ is continuously differentiable on $(q_0,Q)$, the restriction of $\mu_\gamma$ to $(q_0,Q)$ has the density $\gamma'$, and $\mu_\gamma$ has no mass outside $[q_0,Q]$. This proves the decomposition of $\mu_\gamma$. The bounds on $\gamma'$ in \eqref{II:ext:eq-density} are \Cref{II:frsb:cor-densityscale}. For the rates, \eqref{II:frsb:eq-width} and \Cref{II:frsb:lem-rsfamily}(b) give \[ c'e\tilde\ell\le-\tau(\kappa)\le\frac{C_{\mathrm r}(\kappa-\kd)}e,\qquad\text{hence}\qquad Q-q_0=e^2\tilde\ell\le\frac{C_{\mathrm r}}{c'}(\kappa-\kd). \] The mean $m$ of $\mu_\gamma$ lies in $[q_0,Q]$. For $t\in[q_0,Q]$, \eqref{II:frsb:eq-width} and \Cref{II:frsb:lem-rsfamily}(a) give, since $e^3\tilde\ell=e(Q-q_0)\le Q-q_0$, \begin{align*} |t-\qd|&\le(Q-q_0)+|m-q^*(\kappa)|+|q^*(\kappa)-\qd|\\ &\le\frac{(1+C')C_{\mathrm r}}{c'}(\kappa-\kd)+C_{\mathrm r}(\kappa-\kd)=:C_4(\kappa-\kd). \end{align*} Since $\|\gamma-\gamma_{\mathrm d}\|_{L^1}=\int|t-\qd|\dd\mu_\gamma(t)$ (Step 3 of the proof of \Cref{II:prop:localize}), this proves \eqref{II:ext:eq-rates} with $C\ge C_4$. It remains to treat the atoms. By \eqref{II:ext:eq-rates}, $\supp\mu_\gamma\subset\cW_\eta$ with $\eta=C_4(\kappa-\kd)/e^2$. Let $\varepsilon>0$. By \Cref{II:frsb:lem-natcont}(b) there are $\eta(\varepsilon)\in(0,\eta_0]$ and $e(\varepsilon)>0$ such that $|\widetilde x(q)-x_{\mathrm f}|\le\varepsilon$ whenever $ed_1$. All constants depend only on $\Theta$. \end{proof} \section{Proof of the flat-limit theorem} \begin{proof}[Proof of \Cref{II:thm:flat}] The uniqueness statement and the cubic modulus are \Cref{II:thm:flatuniq}. They hold for every normalized profile with bounded support that equals one on a neighborhood of zero. The expansion is \Cref{II:thm:compactexp}, combined with the scaling identity \eqref{II:flat:eq-scaling}, which expresses the first-order correction $\cP^{(1)}(\zeta)$ as the flat functional of the normalized profile at the parameter $t=2\phi_0\sqrt{\ell_0}$, plus an explicit function of $\ell_0$. By \Cref{II:rs:rem-Theta}, \Cref{II:thm:compactexp} holds for every compact interval $\Theta\subset(0,\infty)$. \end{proof} \section{Proof of the full replica symmetry breaking theorem} \begin{proof}[Proof of \Cref{II:thm:main}] Let $e_0,c,C$ be the numerical constants of \Cref{II:thm:uniform}. Part (i) of \Cref{II:thm:main} is \Cref{II:thm:uniform}(i); at $\kappa=\kd$ it is \Cref{II:thm:uniqd}. For part (ii), let $\kappa\in(\kd,\kd+ce^2)$. By \Cref{II:thm:uniform}(ii), global minimizers of $\cP_\kappa$ exist, and every global minimizer $\gamma$ has support $[q_0,Q]$ with $0\frac12x_{\mathrm f}$ and $\mu_\gamma(\{Q\})=1-\widetilde x(Q)>\frac14$, and the last assertion of \Cref{II:thm:uniform}(ii) gives their convergence to $x_{\mathrm f}$ and $1-x_{\mathrm f}$ as $e\to0$ and $(\kappa-\kd)/e^2\to0$, uniformly over global minimizers. The formula $x_{\mathrm f}=\frac12-c_0/(4B_V)\in(0,\frac14)$ is \Cref{II:frsb:lem-flatconst}. This proves (c). Finally (d) is \eqref{II:ext:eq-rates}. \end{proof} The constants in \Cref{II:thm:main,II:thm:thresholds}, and the function $\rho_{\mathrm a}$ of \Cref{II:thm:uniform}, are numerical; those in \Cref{II:thm:compact} depend only on $\Theta$. \appendix \addtocontents{toc}{\protect\addvspace{10pt plus 2pt}\protect\noindent\protect\textsc{Appendices to Part~II}\protect\par} \chapter{Regularity of the Parisi equation}\label{II:app:functional} This appendix proves \Cref{II:lem:regularity,II:fun:diffusion,II:fun:control}, together with the coefficient-continuity identity of \Cref{II:appA:diff}, which is used in \Cref{II:fun:colehopf}, \Cref{II:prop:kkt} and \Cref{II:fun:quad}. The argument has three stages. First, for step coefficients, the solution is an explicit composition of Gaussian integrals (\Cref{II:appA:sec:rec,II:appA:sec:step}); this gives smoothness, the signs, and polynomial bounds whose constants do not depend on the number of steps. Second, a Feynman--Kac representation of the difference of two solutions gives Lipschitz dependence on the coefficient in $L^1$ for every spatial derivative. This defines the solution for general coefficients and transfers all properties to it (\Cref{II:appA:sec:cont,II:appA:sec:ito}); \Cref{II:appA:sec:control} proves the control representation. Third, the expansion at $-\infty$ is derived from the martingale $\psi$ (\Cref{II:appA:sec:asym}), using \Cref{II:lem:generator,II:lem:sandwich}, whose proofs do not use \Cref{II:lem:regularity}(e). Throughout, $C$ and $p$ denote constants that may change from line to line; their dependence is stated in each section. For the Sherrington--Kirkpatrick datum $\log\cosh$, the construction through step coefficients and the Cole--Hopf transformation, and the passage to general coefficients by a limit in $L^1$, are standard \cite{Parisi1980Sequence,Guerra2003,Talagrand2006,TalagrandVolII,PanchenkoBook,JagannathTobasco2016,AuffingerChenProperties}. El Alaoui and Sellke carried them out for the present terminal datum $\log\Phi$, which is concave and decreases quadratically at $-\infty$; their results are recalled after \Cref{II:lem:regularity}. This appendix adds three things: the strict signs in \Cref{II:lem:regularity}(c), including $v_x<0$; the expansion (e) at $-\infty$; and the bound (f) for every derivative with a polynomial weight on the whole line, where \cite[Proposition~13]{ElAlaouiSellke2022} bounds the difference for $x$ in bounded sets. We therefore give the whole argument. We use three standard facts (see \cite{KaratzasShreve1991}). The first is It\^o's formula for a function that is $C^{1,2}$ on $[s,t]\times\R$, applied to a continuous semimartingale. The second is the Feynman--Kac formula. Let $Y$ solve $\dd Y_q=\beta(q,Y_q)\dd q+\dd B_q$ on $[s,t]$ with $\beta$ Lipschitz in $y$ and of linear growth, so that $\E\sup_{[s,t]}|Y|^p<\infty$ for all $p$. Let $\varphi$ be continuous on $[s,t]\times\R$ and $C^{1,2}$ on each $[t_{i-1},t_i]\times\R$ for a finite partition of $[s,t]$, and suppose that $\varphi_q+\frac12\varphi_{yy}+\beta\varphi_y+\mathcal V\varphi=\mathcal F$ there, with $\mathcal V$ bounded and with $\varphi,\varphi_y,\mathcal F$ of polynomial growth. Then \begin{equation}\label{II:appA:FK} \varphi(s,y)=\E\Big[\varphi(t,Y_t)\e^{\int_s^t\mathcal V(q,Y_q)\dd q}-\int_s^t\e^{\int_s^q\mathcal V(q',Y_{q'})\dd q'}\mathcal F(q,Y_q)\dd q\Big],\qquad Y_s=y . \end{equation} This follows from It\^o's formula for $\e^{\int_s^q\mathcal V}\varphi(q,Y_q)$ on each $[t_{i-1},t_i]$, since the stochastic integral has an integrand of polynomial growth and is therefore a true martingale. The third fact is a Gronwall comparison of two diffusions. Let $Y$ and $Y'$ solve $\dd Y=\beta(q,Y)\dd q+\dd B$ and $\dd Y'=\beta'(q,Y')\dd q+\dd B$ on $[s,t]$ with $Y_s=Y'_s$, where $\beta'$ is Lipschitz in $y$ with constant $L$ and $|\beta'(q,y)-\beta(q,y)|\le\Delta(q)(1+|y|)$. Then $|Y'_q-Y_q|\le\int_s^qL|Y'-Y|+\int_s^q\Delta\,(1+|Y|)$, and Gronwall's inequality gives \begin{equation}\label{II:appA:gronwall} \sup_{[s,t]}|Y'-Y|\le\e^{L(t-s)}\Big(1+\sup_{[s,t]}|Y|\Big)\int_s^t\Delta(q)\dd q . \end{equation} \section{One Gaussian step}\label{II:appA:sec:rec} \begin{lemma}\label{II:appA:rec} Let $g\in C^1(\R)$ be concave and nondecreasing, with $g'$ Lipschitz with constant $K$, $g\le0$, and $g(x)\ge-C_g(1+x^2)$. Let $m\ge0$, $s>0$, $w=\cT_{m,s}g$ and $Y=x+\sqrt s\,G$. \begin{enumerate}[label=(\roman*)] \item $w\in C^\infty(\R)$, $\E g(Y)\le w(x)\le0$, and $w$ is concave and nondecreasing with $w'$ Lipschitz with constant $K$. If $g'>0$, then $w'>0$. \item Let $\nu_x$ be the law with density proportional to $\e^{mg(y)}\phi_s(y-x)$, and write $\E_\nu$ and $\Var_\nu$. Then $w'(x)=\E_\nu[g'(Y)]$ and $-w''(x)=\E_\nu[-g''(Y)]-m\Var_\nu[g'(Y)]$, where $g''$ is the almost everywhere derivative of $g'$. Moreover $0\le w'(x)\le C_g(3+2s)+K|x|$. \item If $g\in C^k$ and $|g^{(j)}(y)|\le L(1+|y|)^{p}$ for $1\le j\le k$, then $|w^{(j)}(x)|\le L'(1+|x|)^{p'}$ for $1\le j\le k$, where $L',p'$ depend only on $L,p,k,m,s,K,C_g$ and remain bounded for $m$ and $s$ in bounded sets. As $s\downarrow0$, $\partial_x^j\cT_{m,s}g\to g^{(j)}$ locally uniformly for $0\le j\le k$; for $j\le1$ this holds without the extra hypothesis. \item The function $(s,x)\mapsto\cT_{m,s}g(x)$ is $C^\infty$ on $(0,\infty)\times\R$ and satisfies $\partial_s\cT_{m,s}g=\frac12\big(\partial_{xx}+m(\partial_x\cdot)^2\big)\cT_{m,s}g$. \end{enumerate} \end{lemma} \begin{proof} Take $m>0$; the case $m=0$ is the heat equation and is simpler. Since $0<\e^{mg}\le1$, the function $\e^{mw(x)}=\int \e^{mg(y)}\phi_s(y-x)\dd y$ is positive and $C^\infty$ in $(s,x)$ for $s>0$, and it solves the heat equation $\partial_s\e^{mw}=\frac12\partial_{xx}\e^{mw}$; this gives (iv) and smoothness. Jensen's inequality gives $w\ge\E g(Y)$, and $g\le0$ gives $w\le0$. Concavity of $w$ follows from Pr\'ekopa's theorem that the marginal of a log-concave function is log-concave \cite[Theorem~6]{Prekopa1973}, because $(x,y)\mapsto mg(y)+\log\phi_s(y-x)$ is jointly concave. Integrating by parts in $y$ (the function $\e^{mg}$ is $C^1$ with $\e^{mg}\le1$, and $mg'\e^{mg}$ has linear growth) gives $\partial_x\e^{mw}=\E[mg'(Y)\e^{mg(Y)}]$, which is the formula for $w'$. A second integration by parts, valid since $g'\e^{mg}$ is locally Lipschitz, gives the formula for $w''$. Then $w'\ge0$, $w'>0$ if $g'>0$, and $-w''\le\E_\nu[-g'']\le K$; also $-w''\ge0$ by concavity. By concavity and $w\le0$, \[ w'(0)\le w(0)-w(-1)\le-\E g(-1+\sqrt s\,G)\le C_g(2+s), \] and $w'(x)\le w'(0)+K|x|$ for $x<0$, while $w'(x)\le w'(0)$ for $x\ge0$. For (iii), the density of $\nu_x$ is $\e^{-U}$ with $U(y)=-mg(y)+(y-x)^2/(2s)+\mathrm{const}$, and $U'$ is Lipschitz with $s^{-1}\le U''\le s^{-1}+mK$ almost everywhere. Let $y^*$ be the minimizer of $U$. Then $y^*=x+smg'(y^*)$, so $x\le y^*\le x+smg'(x)$ by monotonicity of $g'$, and $|y^*|\le C(1+|x|)$, since $0\le g'(x)\le2C_g+K|x|$ by concavity of $g$, as in (ii). The bounds on $U''$ give $\e^{-U(y^*)}\sqrt{2\pi/(s^{-1}+mK)}\le1$ and $\e^{-U(y)}\le \e^{-U(y^*)}\e^{-(y-y^*)^2/(2s)}$, hence $\E_\nu|Y-y^*|^{p}\le C_{p}(1+mKs)^{1/2}s^{p/2}$ and $\E_\nu|Y|^{p}\le C(1+|x|)^{p}$. Differentiating $\E_\nu[F(Y)]=\E[F(Y)\e^{mg(Y)}]/\E[\e^{mg(Y)}]$ in $x$ gives \[ \frac{\dd}{\dd x}\E_\nu[F(Y)]=\E_\nu[F'(Y)]+m\Cov_\nu\big(F(Y),g'(Y)\big). \] Starting from $w'=\E_\nu[g']$ and iterating, $w^{(j)}$ is a finite sum of products of $\nu_x$-expectations of monomials in $g'(Y),\dots,g^{(j)}(Y)$, with coefficients that are polynomials in $m$. The moment bound gives the claimed polynomial bound. As $s\downarrow0$, $\nu_x$ converges to the point mass at $x$ locally uniformly in $x$, by the bounds on $y^*$ and on $\E_\nu|Y-y^*|^2$, and dominated convergence in the same representations gives the convergence of derivatives; when $\nu_x$ is replaced by the point mass, the representation of $w^{(j)}$ becomes $g^{(j)}(x)$. \end{proof} \section{Step order parameters}\label{II:appA:sec:step} Let $\gamma\in\cU_r$ be a step order parameter with grid $0=t_0<\dots0$ and $-f_r''=V(z)/(1-r)\in(0,(1-r)^{-1})$ by \Cref{II:fun:gauss}(a), and $f_r(x)\ge-C(1+x^2)$ by \Cref{II:fun:gauss}(c). Moreover all derivatives of $f_r$ have polynomial growth, because $(\log\Phi)'=R$, $R'=-R(z+R)$, and $00$, the identity for $1-v$ gives $1-v(q,x)=\E\big[(1-V(Z^{q,x}))\exp(-\int_q^r\frac\gamma\lambda v)\big]>0$. Consequently $0\le v<1$ on $[0,r]\times\R$, so that $\int_q^r\frac\gamma\lambda v\le\frac r{1-r}$ along every path, and the identity for $v$ has the nonnegative potential $\frac\gamma\lambda(1-v)$. Hence \begin{equation}\label{II:appA:vbounds} v(q,x)\ge\E\,V(Z^{q,x}),\qquad 1-v(q,x)\ge \e^{-r/(1-r)}\,\E\big[1-V(Z^{q,x})\big]. \end{equation} For $c_x$, the identity $\cL c_x-3\gamma c\,c_x=0$ with the potential $-3\gamma c\in[-3(1-r)^{-1},0]$ and the terminal value $c_x(r,\cdot)=V'(z)(1-r)^{-3/2}<0$ give \begin{equation}\label{II:appA:cxbound} \E\,c_x(r,X^{q,x}_r)\le c_x(q,x)\le \e^{-3r/(1-r)}\,\E\,c_x(r,X^{q,x}_r)<0 . \end{equation} \emph{Uniform polynomial bounds.} Write $u^{(k)}=\partial_x^ku$. Differentiating \eqref{II:fun:pde} $k\ge3$ times gives \[ \partial_qu^{(k)}+\tfrac12u^{(k+2)}+\gamma b\,u^{(k+1)}-k\,\gamma\,c\,u^{(k)}+\frac\gamma2\sum_{j=2}^{k-2}\binom kj u^{(j+1)}u^{(k-j+1)}=0 , \] because in $\partial_x^k(u_x^2)=\sum_{j=0}^k\binom kju^{(j+1)}u^{(k-j+1)}$ the terms $j=0$ and $j=k$ equal $2bu^{(k+1)}$ together, and the terms $j=1$ and $j=k-1$ equal $-2kc\,u^{(k)}$ together. The potential $-k\gamma c$ lies in $[-k(1-r)^{-1},0]$, and the sum involves only the orders $3,\dots,k-1$. Since $u^{(2)}=-c$ is bounded, and $|u^{(3)}|=|c_x|\le\E|c_x(r,X_r)|\le C(1+|x|^p)$ by \eqref{II:appA:cxbound}, induction on $k$ with \eqref{II:appA:FK} and the moment bounds gives $|u^{(k)}(q,x)|\le C_k(1+|x|^{p_k})$ with $C_k,p_k$ independent of the grid. The growth bound of \Cref{II:lem:regularity}(b) for $u$ follows from $u\ge\E f_r(x+\sqrt{r-q}\,G)$ and $f_r(y)\ge-C(1+y^2)$. \emph{Behavior at $+\infty$.} Let $t=x-\kappa\ge0$ and $Z=(t+\sqrt{r-q}\,G)/\sqrt{1-r}$. By \Cref{II:fun:gauss}(c), $\log\Phi(z)\ge-2\Phi(-z)-(z^2+2)\1_{\{z<0\}}$ for all $z$; for $z<0$ we also use $\log(1+|z|)\le\frac12(1+z^2)$ and $\Phi(-z)\ge\frac12$. Now $\E\Phi(-Z)=\Phi(-t/\sqrt{1-q})\le \e^{-t^2/(2(1-q))}$, and by the Cauchy--Schwarz inequality \[ \E\big[(Z^2+2)\1_{\{Z<0\}}\big]\le C(1+t^2)\Prob(Z<0)^{1/2}\le C(1+t^2)\e^{-t^2/(4(1-q))}, \] because $\Prob(Z<0)=\Phi(-t/\sqrt{r-q})\le \e^{-t^2/(2(1-q))}$. Since $(1+t^2)\e^{-t^2/8}$ is bounded and $1-q\le1$, we obtain $0\ge u(q,x)\ge\E f_r(x+\sqrt{r-q}\,G)\ge-C\e^{-t^2/(8(1-q))}$. \emph{Translation.} The recursion \eqref{II:fun:recursion} commutes with translations in $x$, and $f_r$ depends on $x-\kappa$ only. This proves \Cref{II:lem:regularity}(a)--(d) and (g) for step order parameters. \section{Continuity in the order parameter and the general case}\label{II:appA:sec:cont} The next identity is the basic tool: it expresses the difference of two solutions with the same terminal datum through the difference of the coefficients. We state it for the class $\mathfrak F$ of terminal data of \Cref{II:fun:control}. For models of Sherrington--Kirkpatrick type it appears in the proof of \cite[Lemma~14]{JagannathTobasco2016}, and for the present datum in the proof of \cite[Proposition~13]{ElAlaouiSellke2022}. \begin{lemma}\label{II:appA:diff} Let $f\in\mathfrak F$, let $\xi,\widetilde\xi\le M$ be nonnegative nondecreasing step functions on $[0,T)$, and put $U=\Upsilon_{\xi,f}$ and $\widetilde U=\Upsilon_{\widetilde\xi,f}$. Let $Y$ solve \[ \dd Y_t=\beta(t,Y_t)\dd t+\dd B_t\quad(q\le t\le T),\qquad Y_q=x,\qquad \beta=\tfrac12\widetilde\xi\,(\widetilde U_x+U_x). \] Then \begin{equation}\label{II:appA:diffid} \begin{split} \widetilde U(q,x)-U(q,x) &=\frac12\E\int_q^T(\widetilde\xi-\xi)(t)\,U_x(t,Y_t)^2\dd t, \\ |\widetilde U(q,x)-U(q,x)|&\le C(1+x^2)\|\widetilde\xi-\xi\|_{L^1(0,T)}, \end{split} \end{equation} where $C$ depends only on $T$, $M$, $C_f$, and $\Lip(f')$. The identity also holds for $T=r$, $f=f_r$, and general $\xi=\gamma$, $\widetilde\xi=\widetilde\gamma$ in $\cU_r$, with $U=u_\gamma$ and $\widetilde U=u_{\widetilde\gamma}$ (see the end of this section). For the quadratic datum $f_k(x)=-\frac12(k-x)_+^2$, which belongs to $\mathfrak F$ with $\Lip(f_k')=1$ and $C_{f_k}\le1+k^2$, the bound shows that $\xi\mapsto\Upsilon_{\xi,f_k}(q,x)$ is Lipschitz for the $L^1$ distance on nondecreasing step functions bounded by $M$, uniformly for $k$ in bounded sets, and therefore extends uniquely and continuously to all nondecreasing $\xi:[0,T)\to[0,M]$. \end{lemma} \begin{proof} The one-step bounds of \Cref{II:appA:rec} give $0\le-U_{xx},-\widetilde U_{xx}\le\Lip(f')$ for $q0$, $v>0$ and $1-v>0$ by \eqref{II:appA:vbounds}, and $c_x<0$ by \eqref{II:appA:cxbound}. The integrated equation in (a) passes to the limit since $\int_q^r|\gamma_n-\gamma|\,u_x^2\le C\|\gamma_n-\gamma\|_{L^1}(1+x^2)$. If $\gamma=m$ on $[s,t)$, choose the partitions to contain $s$ and $t$, so that $\gamma_n=m$ on $[s,t)$ and $u_{\gamma_n}(q,\cdot)=\cT_{m,t-q}u_{\gamma_n}(t,\cdot)$ for $q\in[s,t]$. Since $u_{\gamma_n}\to u$ locally uniformly, $0<\e^{mu_{\gamma_n}}\le1$, and $u_{\gamma_n}\ge-C(1+x^2)$ uniformly in $n$, dominated convergence gives $u(q,\cdot)=\cT_{m,t-q}u(t,\cdot)$ on $[s,t]$. This proves \Cref{II:fun:colehopf}(a) for general $\gamma$, and \Cref{II:appA:rec}(iii),(iv), applied to the function $u(t,\cdot)$, which is smooth with polynomially bounded derivatives, gives the smoothness statement of (a). The identity \eqref{II:appA:diffid} extends to $\gamma,\widetilde\gamma\in\cU_r$ in the same way. Let $\gamma_n\to\gamma$ and $\widetilde\gamma_n\to\widetilde\gamma$ be step approximations in $\cU_r$, and let $Y^n$ be the diffusion of \Cref{II:appA:diff} for the pair $(\gamma_n,\widetilde\gamma_n)$. The left sides converge by (f). The drifts of $Y^n$ are Lipschitz with constant $(1-r)^{-1}$ and, by (b) and (f) with $k=1$, differ from $\beta=\frac12\widetilde\gamma(b_{\widetilde\gamma}+b_\gamma)$ by at most $C(|\widetilde\gamma_n-\widetilde\gamma|(t)+\varepsilon_n)(1+|y|)$, where $\varepsilon_n=\|\gamma_n-\gamma\|_{L^1}+\|\widetilde\gamma_n-\widetilde\gamma\|_{L^1}$. By \eqref{II:appA:gronwall}, $Y^n\to Y$ in every moment, uniformly on $[q,r]$. Since $|b_{\gamma_n}^2-b_\gamma^2|\le C\varepsilon_n(1+y^2)$ and $|\partial_y(b_\gamma^2)|\le C(1+|y|)$, the expectations $\E b_{\gamma_n}(t,Y^n_t)^2$ converge to $\E b_\gamma(t,Y_t)^2$ uniformly in $t$, and they are bounded uniformly in $n$ and $t$. Together with $\int_q^r|(\widetilde\gamma_n-\gamma_n)-(\widetilde\gamma-\gamma)|\le\varepsilon_n$, this shows that the right sides converge. It remains to prove the Lipschitz bound on $\cP_\kappa$ and compactness. By \eqref{II:fun:entropy}, on $\cU_r$ the entropy term is $\frac12\int_0^r\lambda^{-1}+\frac12\log(1-r)$ with $\lambda\ge1-r$ and $|\lambda_{\widetilde\gamma}-\lambda_\gamma|\le\|\widetilde\gamma-\gamma\|_{L^1}$, so its variation is at most $\frac r{2(1-r)^2}\|\widetilde\gamma-\gamma\|_{L^1}$; the energy term is controlled by (f) with $k=0$ at $(0,0)$. Compactness of $\cU_r$ follows from Helly's selection theorem: a sequence in $\cU_r$ has a subsequence converging at every point to a nondecreasing function, which after right-continuous modification lies in $\cU_r$, and the convergence holds in $L^1$ by dominated convergence. \section{It\^o's formula along the diffusion}\label{II:appA:sec:ito} \emph{Proof of \Cref{II:fun:diffusion}.} Part (a) was shown in \Cref{II:appA:sec:step} for step order parameters, and the bounds pass to the limit; (b) was shown in \Cref{II:appA:sec:cont}. For (c), let first $\gamma$ be a step order parameter. Then every $g\in\mathscr D$ is $C^{1,2}$ on each closed block, where $\partial_qg$ is the formal expression with $\gamma$ equal to its value on the block, and $g$ is continuous across grid points. It\^o's formula on each block, the polynomial growth of $g_x$, and the moment bounds give the integrated identity. For general $\gamma\in\cU_r$ let $\gamma_n(q)=\gamma\big(\min\{(\lfloor nq\rfloor+1)/n,1\}\big)$ for $0\le q<1$, and $\gamma_n(1)=1$. Then $\gamma_n\ge\gamma$ is a step order parameter in $\cU_Q\subset\cU_r$, $\gamma_n(q)\to\gamma(q)$ for every $q$ by right-continuity, and $\|\gamma_n-\gamma\|_{L^1}\to0$. The identity holds for $\gamma_n$ and the corresponding $g_n\in\mathscr D$, given by the same formal expression with $u_{\gamma_n}$ and $\lambda_{\gamma_n}$ in place of $u_\gamma$ and $\lambda_\gamma$. By \Cref{II:lem:regularity}(b),(f), the bounds $|\lambda_{\gamma_n}-\lambda_\gamma|\le\|\gamma_n-\gamma\|_{L^1}$ and $\lambda_{\gamma_n}\ge1-r$, and \Cref{II:fun:diffusion}(b), we have $g_n(q,X^n_q)\to g(q,X_q)$ in probability with uniformly bounded moments of every order, and likewise for $\cL_0g_n$ and $\cL_1g_n$. Since $\gamma_n(q)\to\gamma(q)$ for every $q$ and the integrands are bounded uniformly in $n$ and $q$, dominated convergence gives the identity for $\gamma$. The functions $h\in\mathscr D$ are continuous on $[0,r]\times\R$ with polynomial growth, and $X$ has continuous paths, so $q\mapsto\E h(q,X_q)$ is continuous. Hence $e_g$ is Lipschitz, and the one-sided derivatives follow from the right-continuity of $\gamma$ and the existence of $\gamma(q-)$. If $\gamma=m$ on $[s',t')$, the integrand equals $e_{(\cL_0+m\cL_1)g}$ on $[s',t')$, which is continuous on $[s',t']$; iterating gives $e_g\in C^\infty([s',t'])$. \section{The control representation}\label{II:appA:sec:control} The control representation is credited after \Cref{II:fun:control}. The proof below is the verification argument of \cite[Lemma~18]{JagannathTobasco2016}. \emph{Proof of \Cref{II:fun:control}.} The regularity statements and the bounds on $\Upsilon$, $\Upsilon_x$, and $\Upsilon_{xx}$ follow from \Cref{II:appA:rec}, applied block by block from $T$ backwards. Each $\Upsilon(t_j,\cdot)$ satisfies the hypotheses of that lemma with the Lipschitz constant $K=\Lip(f')$ and with $\Upsilon(t_j,x)\ge\E f(x+\sqrt{T-t_j}\,G)\ge-C(1+x^2)$. The bound $\Upsilon_x\le C(1+|x|)$ is \Cref{II:appA:rec}(ii), and the continuity of $\Upsilon$ and $\Upsilon_x$ up to $q=T$ is \Cref{II:appA:rec}(iii) with $k\le1$. Let $\theta$ be a control and $X^\theta_q=x+B_q+\int_0^q\theta$. Since $\xi\le\|\xi\|_\infty$, $\E\int_0^T\theta^2\le\|\xi\|_\infty\E\int_0^T\theta^2/\xi<\infty$, so $\E\sup_{[0,T]}|X^\theta|^2<\infty$. On each closed block $[t_{j-1},t_j]$ with $t_j0$, and both sides vanish when $\xi=\theta=0$. The stochastic integral has a square-integrable integrand, since $\Upsilon_x$ has linear growth, and therefore has mean zero. Taking expectations, summing over blocks, and letting $\varepsilon\downarrow0$, using the continuity of $\Upsilon$ on $[0,T]\times\R$, its quadratic growth, and dominated convergence, gives \eqref{II:fun:controlineq}. For $\theta^*=\xi\Upsilon_x(q,X^*_q)$ the inequality is an equality at every step. The drift $\xi\Upsilon_x$ is Lipschitz in $x$ with constant $\|\xi\|_\infty\Lip(f')$ and has linear growth, so $X^*$ exists, is unique, and has moments of all orders, and $\theta^*$ is a control. Finally, differentiating the equation of $\Upsilon$ in $x$ on each block shows that $\Upsilon_x$ solves $\partial_q\Upsilon_x+\frac12\partial_{xx}\Upsilon_x+\xi\Upsilon_x\partial_x\Upsilon_x=0$, so by It\^o's formula $\dd\Upsilon_x(q,X^*_q)=\Upsilon_{xx}(q,X^*_q)\dd B_q$ on $[0,T)$. Since $|\Upsilon_{xx}|\le\Lip(f')$, this is a square-integrable martingale, and its limit as $q\uparrow T$ is $f'(X^*_T)$ by the continuity of $\Upsilon_x$ up to $T$. \section{The expansion at \texorpdfstring{$-\infty$}{minus infinity}}\label{II:appA:sec:asym} \emph{Proof of \Cref{II:lem:regularity}(e).} Constants depend only on $r$ and $K_0$. Write $y=\kappa-x>0$. \proofstep{1}{$q\in[Q,r]$.} Here $u=\log\Phi(-t)$ and $\psi=\sqrt{1-q}\,\Rbar(-t)$ with $t=y/\sqrt{1-q}$, by \eqref{II:fun:top} and \Cref{II:lem:generator}(b). Since $\phi(t)=\e^{-t^2/2}/\sqrt{2\pi}$, \[ \log\Phi(-t)=-\tfrac12t^2-\log t-\log\sqrt{2\pi}+\log\frac{t\Phi(-t)}{\phi(t)}, \] and the last term lies in $(-t^{-2},0)$, because $\frac{t^2}{1+t^2}<\frac{t\Phi(-t)}{\phi(t)}<1$ by \Cref{II:fun:gauss}(b). Since $\lambda(q)=1-q$ and $\omega(q)=\frac12\log(1-q)-\frac12\log(2\pi)$ on $[Q,1)$, this is the first claim. By \Cref{II:fun:gauss}(b), $\frac{1-q}y-\frac{3(1-q)^2}{y^3}<\psi<\frac{1-q}y$ for $t\ge2$, which is the second claim, because $\lambda b-(\kappa-x)=\psi$. \proofstep{2}{The estimate for $\psi$ when $q0$, $\Rbar>0$, $00$, and $0<1-V(-t)<3t^{-2}$ for $t\ge2$ (part (b)); $\Phi(-t)\le\e^{-t^2/2}$ for $t\ge0$ (part (c)); and $R(z)\le2\phi(z)$ and $V(z)\le2(1+z)\phi(z)$ for $z\ge0$ (part (e)). Put \[ r(z):=\frac{V'(z)^2}{V(z)^2\bigl(1-V(z)\bigr)}\qquad(z\in\R). \] \begin{lemma}\label{II:appB:lem-gauss} \begin{enumerate}[label=(\alph*)] \item For $z\ge0$: $|V'(z)|\le2(2+z)^2\phi(z)$ and $r(z)\le C(1+z)^2$. \item For $t\ge2$, \[ \bigl|1-V(-t)-t^{-2}\bigr|\le Ct^{-4},\qquad \bigl|V'(-t)+2t^{-3}\bigr|\le Ct^{-5}. \] Consequently, for $z\le0$: $c/(1+z^2)\le1-V(z)\le C/(1+z^2)$, $|V'(z)|\le C(1+|z|)^{-3}$, $|V''(z)|\le C(1+|z|)^{-2}$, and $r(z)\le C(1+z^4)^{-1}$. \end{enumerate} \end{lemma} \begin{proof} (a) Let $z\ge0$. Since $R\le R(0)<1$, $\Rbar=z+R\le1+z$. As $00$ because $V$ is decreasing. Hence $r(z)\le C(1+z)^2$. (b) Let $t\ge2$ and $I_k:=\int_0^\infty w^k\exp(-tw-w^2/2)\dd w$. Then $\E T_{-t}=I_1/I_0$ and $\E T_{-t}^2=I_2/I_0$, so that \[ 1-V(-t)=\frac{I_2I_0-I_1^2}{I_0^2},\qquad V'(-t)=R(-t)\bigl(\Var T_{-t}-(\E T_{-t})^2\bigr)=R(-t)\,\frac{I_2I_0-2I_1^2}{I_0^2}, \] by \eqref{II:fun:Vpp}. Since $1-w^2/2\le\exp(-w^2/2)\le1-w^2/2+w^4/8$ and $\int_0^\infty w^j\e^{-tw}\dd w=j!\,t^{-j-1}$, \begin{align*} I_0&=\tfrac1t\bigl(1-t^{-2}+O(t^{-4})\bigr),\\ I_1&=\tfrac1{t^2}\bigl(1-3t^{-2}+O(t^{-4})\bigr),\\ I_2&=\tfrac2{t^3}\bigl(1-6t^{-2}+O(t^{-4})\bigr), \end{align*} where here and below $O(t^{-j})$ denotes a quantity bounded by a numerical constant times $t^{-j}$ for $t\ge2$. Hence \begin{align*} I_2I_0&=\tfrac2{t^4}\bigl(1-7t^{-2}+O(t^{-4})\bigr),\\ I_1^2&=\tfrac1{t^4}\bigl(1-6t^{-2}+O(t^{-4})\bigr),\\ I_0^2&=\tfrac1{t^2}\bigl(1-2t^{-2}+O(t^{-4})\bigr), \end{align*} and $I_0^2\ge\frac9{16}t^{-2}$ because $I_0\ge t^{-1}(1-t^{-2})$. Therefore \[ \begin{aligned} 1-V(-t)&=\frac{1-8t^{-2}+O(t^{-4})}{t^2\bigl(1-2t^{-2}+O(t^{-4})\bigr)}=t^{-2}+O(t^{-4}),\\ \frac{I_2I_0-2I_1^2}{I_0^2}&=\frac{-2t^{-2}+O(t^{-4})}{t^2\bigl(1+O(t^{-2})\bigr)}=-2t^{-4}+O(t^{-6}). \end{aligned} \] With $R(-t)=t+O(t^{-1})$ this gives $V'(-t)=-2t^{-3}+O(t^{-5})$. For the consequences, let $t_0\ge2$ be a large numerical constant. For $z\le-t_0$ the bounds on $1-V$ and $V'$ follow from the expansions. On $[-t_0,0]$ the functions $1-V$ and $|V'|$ are continuous, and $1-V\ge1-V(-t_0)>0$. The bound on $V''$ follows from \eqref{II:fun:Vpp}, since $00$ on $(-\infty,0]$. \end{proof} The next lemma is the monotonicity that drives the Chebyshev argument of \Cref{II:appB:ssec-topcheb}. \begin{lemma}\label{II:appB:lem-mlr} The function $r=(V')^2/(V^2(1-V))$ is strictly increasing on $\R$. \end{lemma} \begin{proof} Write $\mathsf a=1-V>0$ and $\mathsf b=V'/R=1-V-\Rbar^2$, so that $\Rbar^2=\mathsf a-\mathsf b$ and $\Rbar(R+\Rbar)=V+\Rbar^2=1-\mathsf b$. \proofstep{1}{Reduction.} Since $(\log r)'=2V''/V'-2V'/V+V'/(1-V)$, we have $(\log r)'\cdot V'V(1-V)=-\mathcal C$ with \[ \mathcal C:=(V')^2(2-3V)-2V''V(1-V) . \] As $V'V(1-V)<0$, the derivative $(\log r)'$ has the sign of $\mathcal C$. Inserting $V''=-(R+\Rbar)V'-2V(1-V)$ from \eqref{II:fun:Vpp}, $V'=R\mathsf b$, and $V=R\Rbar$, \[ \mathcal C=R^2\bigl[\mathsf b^2(3\mathsf a-1)+2(1-\mathsf b)\mathsf a\mathsf b+4\mathsf a^2(\mathsf a-\mathsf b)\bigr]=R^2(2\mathsf a-\mathsf b)(2\mathsf a^2-\mathsf a\mathsf b+\mathsf b), \] where the factorization is verified by expanding both sides to $4\mathsf a^3-4\mathsf a^2\mathsf b+\mathsf a\mathsf b^2+2\mathsf a\mathsf b-\mathsf b^2$. Since $2\mathsf a-\mathsf b=\mathsf a+\Rbar^2>0$, it suffices to show that \[ \mathcal Q:=2\mathsf a^2-\mathsf a\mathsf b+\mathsf b=(1-V)(2-V)-V\Rbar^2 \] is positive on $\R$. \proofstep{2}{An equation for $\mathcal Q$.} We have $\mathcal Q'=-V'(3-2V+\Rbar^2)-2V\Rbar(1-V)$ and $V'=R\mathsf b=(V/\Rbar)(1-V-\Rbar^2)$. Eliminating $\Rbar^2=((1-V)(2-V)-\mathcal Q)/V$, so that $1-V-\Rbar^2=(\mathcal Q-2(1-V)^2)/V$ and $3-2V+\Rbar^2=(2-V^2-\mathcal Q)/V$, and expanding, \begin{equation}\label{II:appB:eq-Qode} V\Rbar\,\mathcal Q'=4(1-V)^3-\mathcal Q\bigl(1+3(1-V)^2-\mathcal Q\bigr). \end{equation} \proofstep{3}{Barrier.} The function $\mathfrak k:=(1+3(1-V)^2-\mathcal Q)/(V\Rbar)$ is continuous, since $V\Rbar>0$. Fix $z_*\in\R$ and put $\mathcal F(z):=\mathcal Q(z)\exp\int_{z_*}^z\mathfrak k$. By \eqref{II:appB:eq-Qode}, $\mathcal F'=4(1-V)^3(V\Rbar)^{-1}\exp\int_{z_*}^z\mathfrak k>0$, so $\mathcal F$ is strictly increasing. As $z=-t\to-\infty$ we have $\Rbar^2\le t^{-2}$ and $1-V\le3t^{-2}$; hence $|\mathcal Q|\le2(1-V)+\Rbar^2\to0$, and the numerator of $\mathfrak k$ tends to one. Choosing $z_*$ sufficiently negative, $\mathfrak k>0$ on $(-\infty,z_*]$, and then $|\mathcal F(z)|\le|\mathcal Q(z)|\to0$ as $z\to-\infty$. An increasing function with limit zero at $-\infty$ is positive, so $\mathcal Q>0$. \end{proof} \begin{lemma}\label{II:appB:lem-xPhi} The constant $x_\Phi$ of \eqref{II:cpt:eq-xPhi} satisfies \[ x_\Phi=1-\frac{\int_\R\Phi V^2\dd z}{4\int_\R\Phi V^2(1-V)\dd z}<1 . \] \end{lemma} \begin{proof} Let $\mathcal G_{\mathrm{top}}f:=\frac12f''+(R+\frac z2)f'$. This is the generator of the diffusion \eqref{II:flat:eq-topdiff} of \Cref{II:sec:flat}, but only the formula is used here. By \eqref{II:fun:Vpp}, \[ \mathcal G_{\mathrm{top}}(V^2)=(V')^2+VV''+(2R+z)VV'=(V')^2-2V^2(1-V). \] The formal adjoint is $\mathcal G_{\mathrm{top}}^*w=\frac12w''-((R+\frac z2)w)'$. With $\Phi'=R\Phi$, $\Phi''=-zR\Phi$, $(R+\frac z2)'=\frac12-V$, and $V=zR+R^2$, \[ \mathcal G_{\mathrm{top}}^*\Phi=-\tfrac12zR\Phi-\Phi\bigl(\tfrac12-V+R^2+\tfrac12zR\bigr)=-\tfrac12\Phi . \] For $N>0$, \[ \int_{-N}^N\bigl(\Phi\,\mathcal G_{\mathrm{top}}(V^2)-V^2\mathcal G_{\mathrm{top}}^*\Phi\bigr)\dd z=\Bigl[\tfrac12\bigl(\Phi(V^2)'-\Phi'V^2\bigr)+\bigl(R+\tfrac z2\bigr)\Phi V^2\Bigr]_{-N}^N . \] As $N\to\infty$ the boundary terms vanish: at $+N$ because $V$ and $V'$ are $O(N^2\phi(N))$ by \Cref{II:appB:lem-gauss}(a) while $\Phi$ and $R$ are bounded, and at $-N$ because $\Phi(-N)\le\phi(N)/N$, $|V|\le1$, $|V'|\le C$, and $R(-N)\le N+1$. All integrands are integrable. Hence $\int\Phi(V')^2-2\int\Phi V^2(1-V)=-\frac12\int\Phi V^2$, which is the formula, and $\int\Phi V^2>0$. \end{proof} \section{The last constant piece in scale-free form}\label{II:appB:ssec-scale} For $m\in(0,1]$ let $M_m$ be a random variable with distribution function $\Phi^m$, independent of $G$. Its density is $f_M:=m\phi\Phi^{m-1}$. For $\beta\in[0,1]$ put $\sigma_\beta:=\sqrt{(1-\beta)/m}$ and \[ \mathsf Y:=\sqrt\beta\,M_m+\sigma_\beta G,\qquad F_m(z;\beta):=\Prob(\mathsf Y\le z),\qquad f:=\partial_zF_m , \] \[ R_m(z;\beta):=\partial_z\log F_m(z;\beta),\qquad v_m(z;\beta):=-\frac1m\,\partial_z^2\log F_m(z;\beta) . \] When $m$ and $\beta$ are fixed we write $F$, $v$, and so on. \begin{lemma}\label{II:appB:lem-scalefree} Let $\gamma\in\cU$ and $\kappa\in\R$, and suppose that $\gamma\equiv m\in(0,1]$ on $[p,Q)$ for some $p0$, and therefore \[ v_m=\frac1m\Bigl(\frac{f^2}{F^2}-\frac{f'}F\Bigr),\qquad \partial_zv_m=-\frac1m\Bigl(\frac{f''}F-\frac{3ff'}{F^2}+\frac{2f^3}{F^3}\Bigr) \] are continuous. For $m=1$ or $\beta\in\{0,1\}$, $v_m\in(0,1)$ because $0|\mu|/2)\le\exp(-\mu^2/16)$. For $|\mu|\le1$ uniform Gaussian moments suffice. Thus all constants in \eqref{II:appB:eq-smoothedpower} are independent of $\bar m$, $\mu$, and $\sigma$. Since $f_M=m\phi^mR^{\bar m}$ and $\phi^m=(2\pi)^{\bar m/2}m^{-1/2}\phi_{1/m}$, completing the square in the Gaussian convolution defining $f$ gives \begin{equation}\label{II:appB:eq-gaussfactor} f(z)=C_m\phi_{1/m}(z)H(z),\qquad C_m:=\sqrt m(2\pi)^{\bar m/2},\qquad H(z):=\E\,\mathfrak g\bigl(\sqrt\beta z+\sigma_\beta G\bigr). \end{equation} For $0<\beta<1$ this follows by writing the density of $\sqrt\beta M_m$ as $C_m\phi_{\beta/m}(x)\mathfrak g(x/\sqrt\beta)$ and using $\phi_{\beta/m}(x)\phi_{(1-\beta)/m}(z-x)=\phi_{1/m}(z)\phi_{\beta(1-\beta)/m}(x-\beta z)$. The same identity holds at $\beta\in\{0,1\}$ by continuity. By \eqref{II:appB:eq-smoothedpower} with $\mu=\sqrt\beta z$ and $\sigma=\sigma_\beta\le\sqrt2$, and since $\sqrt\beta/(1+\sqrt\beta t)\le1/(1+t)$ for $t\ge0$, we have for $z\le0$ and $t,u\ge0$ \begin{equation}\label{II:appB:eq-Hbounds} |H^{(j)}(z)|\le \frac{C_jH(z)}{(1+|z|)^j}\quad(1\le j\le3), \qquad \frac{H(-t-u)}{H(-t)}\le C\Bigl(1+\frac{u}{1+t}\Bigr)^{\bar m} . \end{equation} Also $H$ is nonincreasing. For $t\ge1$, integrate \eqref{II:appB:eq-gaussfactor} over $z\le-t$ and substitute $z=-t-\rho/(mt)$. Since $\frac m2(t+\frac\rho{mt})^2=\frac m2t^2+\rho+\frac{\rho^2}{2mt^2}$, \begin{equation}\label{II:appB:eq-tailfactor} \begin{split} F(-t)&=\frac{C_m\sqrt{m/(2\pi)}}{mt}\exp(-mt^2/2)\,\Xi(t),\\ \Xi(t)&:=\int_0^\infty\exp\Bigl(-\rho-\frac{\rho^2}{2mt^2}\Bigr) H\Bigl(-t-\frac\rho{mt}\Bigr)\dd\rho. \end{split} \end{equation} We claim that \begin{equation}\label{II:appB:eq-Aderiv} cH(-t)\le\Xi(t)\le CH(-t),\qquad |\Xi^{(j)}(t)|\le C_jt^{-j}H(-t)\quad(1\le j\le3). \end{equation} For the lower bound, restrict the integral to $0\le\rho\le1$, where the exponential is at least $\e^{-2}$, and use $H(-t-\rho/(mt))\ge H(-t)$. For the upper bounds let $\xi(t)=-t-\rho/(mt)$. Then $|\xi'|\le1+2\rho$, $|\xi''|\le4\rho t^{-3}$, and $|\xi'''|\le12\rho t^{-4}$. Together with $|\xi|\ge t$ and \eqref{II:appB:eq-Hbounds}, the chain rule bounds the $j$th derivative of $H(\xi(t))$ by $C_jt^{-j}P_j(\rho)H(\xi(t))$, where $P_j$ is a fixed polynomial. The $j$th derivative of $\exp(-\rho^2/(2mt^2))$ has the same type of bound, with that exponential in place of $H(\xi(t))$. Finally \eqref{II:appB:eq-Hbounds} bounds $H(\xi(t))/H(-t)$ by $C(1+2\rho)^{\bar m}$. The integrable majorants $C\e^{-\rho}P(\rho)$ justify each differentiation and prove \eqref{II:appB:eq-Aderiv}. Consequently $|(\log\Xi)^{(j)}(t)|\le C_jt^{-j}$ for $j=1,2,3$. Taking logarithms in \eqref{II:appB:eq-tailfactor}, \[ \log F(-t)=\text{constant}-\frac m2t^2-\log t+\log\Xi(t) . \] Its second derivative in $t$ differs from $-m$ by $O(t^{-2})$, and its third derivative is $O(t^{-3})$, uniformly in $m$ and $\beta$. Since $v_m=-m^{-1}(\log F)''$, this proves (V3) and (V5) for $z\le-1$. The compact interval $[-1,0]$ is covered by (V0). \proofstep{4}{Proof of (V2).} The bounds on $\mathfrak g,\mathfrak g',\mathfrak g''$ above give $|\mathfrak g^{(j)}(x)|\le C(1+|x|)^2$ for all $x$ and $j\le2$, hence $|H^{(j)}(z)|\le C(1+z)^2$ for $z\ge0$ and $j\le2$. Thus \eqref{II:appB:eq-gaussfactor} gives $|f^{(j)}(z)|\le C\exp(-cz^2)$ for $z\ge0$ and $j\le2$. Since $F\ge F(0)\ge1/4$ there by (V1), the formulas for $v_m$ and $\partial_zv_m$ in the proof of (V0) prove (V2). \proofstep{5}{Proof of (V4).} By (V0) and compactness, $1-v_m\ge c_1>0$ on $[-1,0]$, uniformly in $m$ and $\beta$. Let $z<-1$, and first let $m\in[m_0,1)$ and $\beta\in(0,1)$. We use the realization \eqref{II:appB:eq-realize} at level $q=q_1$, where $\beta_q=\beta$, and the point $x=\sqrt{\lambda(q)}\,z$. Choose $q'\in[q,Q]$ with \[ \bar\beta:=\lambda(q')/\lambda(q)=\max\{\beta,z^{-2}\} , \] which is possible because $\lambda(\cdot)/\lambda(q)$ decreases continuously from $1$ to $\beta$ on $[q,Q]$. Let $X=X^{q,x}$ be the Parisi diffusion started at $(q,x)$. The generator identity $\cL(1-v)=(m/\lambda)v(1-v)$ (\Cref{II:lem:generator}(a)) and It\^o's formula on the constant block $[q,q']$ show that $(1-v(t,X_t))\exp(-\int_q^t(m/\lambda)v)$, $t\in[q,q']$, is a local martingale bounded by one, hence a martingale. Since $v<1$ and $\int_q^{q'}m/\lambda=\log(1/\bar\beta)$, \[ 1-v(q,x)=\E\Bigl[\bigl(1-v(q',X_{q'})\bigr)\exp\Bigl(-\int_q^{q'}\frac m\lambda v\dd t\Bigr)\Bigr]\ge\bar\beta\,\E\bigl[1-v(q',X_{q'})\bigr]. \] Moreover $\psi=\lambda b+x>0$ (\Cref{II:lem:generator}(b), with $\kappa=0$), and It\^o's formula with $\lambda'=-m$ gives \[ \dd\Bigl(\frac{X_t}{\lambda(t)}\Bigr)=\frac{m\psi(t,X_t)}{\lambda(t)^2}\dd t+\frac{\dd B_t}{\lambda(t)} , \qquad\text{hence}\qquad Z':=\frac{X_{q'}}{\sqrt{\lambda(q')}}\ge\sqrt{\bar\beta}\,z+N , \] where $N=\sqrt{\lambda(q')}\int_q^{q'}\dd B_t/\lambda(t)$ is centered Gaussian with variance $(1-\bar\beta)/m$. By \Cref{II:appB:lem-scalefree}, $v(q',\cdot)=v_m(\cdot/\sqrt{\lambda(q')};\beta/\bar\beta)$, and $v_m$ is decreasing in $z$ (\Cref{II:lem:regularity}(c)). On $\{N\ge0\}$, whose probability is at least $\frac12$, \[ 1-v_m(Z';\beta/\bar\beta)\ge1-v_m(\sqrt{\bar\beta}\,z;\beta/\bar\beta) . \] If $\bar\beta=z^{-2}$, then $\sqrt{\bar\beta}\,z=-1$ and the right side is at least $c_1$, which gives $1-v_m(z;\beta)\ge c_1/(2z^2)$. Otherwise $\bar\beta=\beta>z^{-2}$ and $q'=Q$, so $v_m(\cdot;1)=V$ and \Cref{II:appB:lem-gauss}(b) gives \[ 1-v_m(z;\beta)\ge\frac\beta2\bigl(1-V(\sqrt\beta z)\bigr)\ge\frac{c\beta}{2(1+\beta z^2)}\ge\frac c{4z^2}, \] because $1+\beta z^2\le2\beta z^2$. In both cases $1-v_m(z;\beta)\ge c/(1+z^2)$. The cases $m=1$ or $\beta\in\{0,1\}$, where $v_m$ is $V$ or $V(\sqrt m\,\cdot)$, follow from \Cref{II:appB:lem-gauss}(b). This proves (V4). \proofstep{6}{The bound on $r_m$.} By (V3), (V0), and compactness, $v_m\ge c>0$ on $(-\infty,0]$. Together with (V4) and (V5), this gives $r_m\le C(1+|z|)^{-6}(1+z^2)\le C(1+z^4)^{-1}$ for $z\le0$, and completes the proof of \Cref{II:appB:prop-vunif}. \end{proof} \section{Proof of the breakpoint bound}\label{II:appB:ssec-topcheb} We first record Chebyshev's association inequality in the form needed here: an average against a nonincreasing weight is bounded by the largest average over half-lines. For a nondecreasing integrand it reduces to the classical inequality (see \cite{BoucheronLugosiMassart2013}). \begin{lemma}\label{II:appB:lem-mix} Let $\nu$ be a finite measure on $\R$ with a positive density, let $\tilde g\ge0$ be nonincreasing with $0<\int\tilde g\dd\nu<\infty$, and let $h\ge0$ be measurable with $\int h\dd\nu<\infty$. For $\tau\in(-\infty,\infty]$ put $\mathcal R_h(\tau):=\nu((-\infty,\tau))^{-1}\int_{(-\infty,\tau)}h\dd\nu$. Then \[ \frac{\int h\,\tilde g\dd\nu}{\int\tilde g\dd\nu}\le\sup_{\tau\in(-\infty,\infty]}\mathcal R_h(\tau). \] If $h$ is nondecreasing, then $\mathcal R_h(\tau)\le\mathcal R_h(\infty)$ for all $\tau$, so the left side is at most $\int h\dd\nu/\nu(\R)$. \end{lemma} \begin{proof} Changing $\tilde g$ on a countable set does not affect the integrals, so we may assume that $\tilde g$ is right-continuous. Then $\tilde g(z)=\tilde g(\infty)+\varsigma((z,\infty))$ for the measure $\varsigma\ge0$ with $\varsigma((z_1,z_2])=\tilde g(z_1)-\tilde g(z_2)$. By Tonelli's theorem, for every measurable $k\ge0$, \[ \int k\,\tilde g\dd\nu=\tilde g(\infty)\int k\dd\nu+\int\varsigma(\dd\tau)\int_{(-\infty,\tau)}k\dd\nu . \] Taking $k=h$ and $k=1$ shows that the ratio is a weighted average of the values $\mathcal R_h(\tau)$, $\tau\in\R$, with weights $\nu((-\infty,\tau))\,\varsigma(\dd\tau)$, and of $\mathcal R_h(\infty)$, with weight $\tilde g(\infty)\nu(\R)$. If $h$ is nondecreasing, then $h(z)\le h(\tau)$ for $z<\tau$ and $h(z)\ge h(\tau)$ for $z\ge\tau$, so \[ \nu\bigl([\tau,\infty)\bigr)\int_{(-\infty,\tau)}h\dd\nu\le\nu\bigl((-\infty,\tau)\bigr)\int_{[\tau,\infty)}h\dd\nu , \] which is equivalent to $\mathcal R_h(\tau)\le\mathcal R_h(\infty)$. \end{proof} \begin{proof}[Proof of \Cref{II:cpt:prop-topcheb}] We divide the proof into four steps. \proofstep{1}{The law of the scale-free variable.} Let $\eps>0$ be given; $\bar\eta\in(0,1/4]$ will be chosen at the end. Let $|\kappa|\le\bar\eta$ and $\gamma\in\cU$. \emph{Case (b).} Suppose that $\gamma\equiv m$ on $[p,Q)$ with $1-\bar\eta\le m<1$ and $\lambda(p)\le\bar\eta$, and let $q\in[p,Q)$. Then $m\ge3/4\ge m_0$. Since $\gamma\ge m$ on $[p,1]$, we have $\lambda(p)\ge m(1-p)$, so $1-q\le1-p\le2\bar\eta\le\frac12$ and $q\ge\frac12$; also $\lambda:=\lambda(q)\le\bar\eta$. Put $\beta:=\delta/\lambda$. By \Cref{II:lem:tilted}, $X_q$ has density $\e^{mu(q,x)}\phi_q(x)g_q(x)$ with $g_q$ nonincreasing, and by \Cref{II:appB:lem-scalefree}, $\e^{mu(q,x)}=F_m(z;\beta)$ with $z=(x-\kappa)/\sqrt\lambda$. Hence $Z_q:=(X_q-\kappa)/\sqrt\lambda$ has density proportional to $F_m(z;\beta)\,\mathcal G(z)\,\tilde g(z)$, where \[ \mathcal G(z):=\frac{\phi_q(\kappa+\sqrt\lambda z)}{\phi_q(\kappa)}=\exp\Bigl(-\frac{2\kappa\sqrt\lambda z+\lambda z^2}{2q}\Bigr),\qquad \tilde g(z):=g_q(\kappa+\sqrt\lambda z), \] and $\tilde g$ is nonincreasing. Since $2|\kappa|\sqrt\lambda\le2\bar\eta^{3/2}\le1\le2q$ and $\lambda\le2q$, we have $\exp(-|z|-z^2)\le\mathcal G(z)\le\exp(|z|)$, and $\mathcal G\to1$ locally uniformly as $\bar\eta\to0$. By \Cref{II:appB:lem-scalefree} and \eqref{II:cpt:eq-xt}, \begin{align*} \widetilde x_\gamma(q)&=\frac{\E(\partial_zv_m)^2(Z_q;\beta)}{2\E\bigl[v_m^2(1-v_m)\bigr](Z_q;\beta)}=\frac12\,\frac{\int r_m\,\tilde g\dd\nu}{\int\tilde g\dd\nu},\\ \nu(\dd z)&:=F_m(z;\beta)\,\mathcal G(z)\,\bigl(v_m^2(1-v_m)\bigr)(z;\beta)\dd z . \end{align*} The measure $\nu$ is finite and has a positive density: $F_m\le2\exp(-z^2/16)$ for $z\le0$ and $v_m\le C\exp(-cz^2)$ for $z\ge0$ by \Cref{II:appB:prop-vunif}, while $\mathcal G\le\exp(|z|)$. Also $0<\int\tilde g\dd\nu<\infty$, since $\int\tilde gF_m\mathcal G\dd z$ is a normalizing constant of the density of $Z_q$. \emph{Case (a).} Suppose that $\delta\le\bar\eta$ and take $q=Q$. Then $Q\ge\frac12$, $\gamma(Q)=1$, $u(Q,x)=\log\Phi(z)$ with $z=(x-\kappa)/\sqrt\delta$, and $v(Q,x)=V(z)$. By \Cref{II:lem:tilted}, the same computation gives $\widetilde x_\gamma(Q)=\frac12\int r\,\tilde g\dd\nu_Q/\int\tilde g\dd\nu_Q$ with $\nu_Q(\dd z):=\Phi(z)\mathcal G(z)V^2(1-V)(z)\dd z$, where $\lambda=\delta$ in $\mathcal G$. \proofstep{2}{The main term.} Recall $r=r_1=(V')^2/(V^2(1-V))$. In case (b), $\widetilde x_\gamma(q)\le\frac12\mathcal M+\frac12\mathcal E$, where \[ \mathcal M:=\frac{\int r\,\tilde g\dd\nu}{\int\tilde g\dd\nu},\qquad \mathcal E:=\frac{\int|r_m-r|\,\tilde g\dd\nu}{\int\tilde g\dd\nu}. \] In case (a), $\widetilde x_\gamma(Q)=\frac12\mathcal M$ with $\nu_Q$ in place of $\nu$, and there is no error term. Since $r$ is nondecreasing (\Cref{II:appB:lem-mlr}) and $\tilde g$ is nonincreasing, \Cref{II:appB:lem-mix} gives $\mathcal M\le\int r\dd\nu/\nu(\R)$. We claim that \[ \frac{\int r\dd\nu}{\nu(\R)}\longrightarrow\frac{\int\Phi(V')^2}{\int\Phi V^2(1-V)}=2x_\Phi\qquad\text{as }\bar\eta\to0 , \] uniformly over all admissible $(\kappa,\gamma,m,q)$, and likewise for $\nu_Q$. Otherwise there are $\eps_1>0$ and admissible sequences with $\bar\eta_n\to0$ that violate the claim by $\eps_1$. Along a subsequence $\beta_n\to\beta_\infty\in[0,1]$, while $m_n\to1$, $\kappa_n\to0$, $\lambda_n\to0$, and $q_n\in[\frac12,1]$. By (V0), $F_{m_n}(\cdot;\beta_n)\to F_1(\cdot;\beta_\infty)=\Phi$ and $v_{m_n}(\cdot;\beta_n)\to V$ pointwise, and $\mathcal G_n\to1$. The densities of $r\,\nu_n$ and $\nu_n$ are dominated by an integrable function independent of $n$: for $z\le0$ by $C\exp(|z|-z^2/16)$, using $r\le C$ there (\Cref{II:appB:lem-gauss}(b)); for $z\ge0$ by $C(1+z)^2\exp(|z|-cz^2)$, using $r\le C(1+z)^2$ (\Cref{II:appB:lem-gauss}(a)) and (V2). Dominated convergence gives the claim, since $\int\Phi V^2(1-V)>0$ and $\int\Phi rV^2(1-V)=\int\Phi(V')^2$. The case of $\nu_Q$ is the same with $m=1$. \proofstep{3}{The error term.} By \Cref{II:appB:lem-mix} with $h=|r_m-r|$, $\mathcal E\le\sup_\tau\mathcal R_{|r_m-r|}(\tau)$. Fix $T\ge1$ and $T'\ge1$. \emph{Far left.} By \Cref{II:appB:prop-vunif} and \Cref{II:appB:lem-gauss}(b), \[ \mathfrak t_1(T):=\sup\bigl\{r_{m'}(z;\beta)+r(z):\ z\le-T,\ m'\in[m_0,1],\ \beta\in[0,1]\bigr\}\le CT^{-4} . \] For $\tau\le-T$, bounding the numerator of $\mathcal R_{|r_m-r|}(\tau)$ by $\mathfrak t_1(T)\,\nu((-\infty,\tau))$ gives $\mathcal R_{|r_m-r|}(\tau)\le\mathfrak t_1(T)$. For $\tau>-T$, the part of the numerator over $(-\infty,-T)$ is at most $\mathfrak t_1(T)\,\nu((-\infty,-T))\le\mathfrak t_1(T)\,\nu((-\infty,\tau))$, so after division its contribution is again at most $\mathfrak t_1(T)$. No reciprocal of a small Gaussian mass appears. \emph{Bulk and right tail.} For $\tau>-T$ the denominator satisfies $\nu((-\infty,\tau))\ge\nu((-T-1,-T))\ge c_T>0$, uniformly: on $[-T-1,-T]$ the factors $F_m$ and $v_m^2(1-v_m)$ are bounded below by (V0) and compactness, and $\mathcal G\ge\exp(-|z|-z^2)$. The remaining part of the numerator is at most \[ \omega_{T,T'}(1-m)\,\nu(\R)+\int_{T'}^\infty(r_m+r)\dd\nu , \] where \[ \omega_{T,T'}(\eps'):=\sup\bigl\{|r_{m'}(z;\beta)-r(z)|:\ 1-\eps'\le m'\le1,\ \beta\in[0,1],\ z\in[-T,T']\bigr\}. \] Here $\nu(\R)\le C$. By (V0), $r_{m'}$ is jointly continuous on the compact set $[m_0,1]\times[0,1]\times[-T,T']$, its denominator being positive there, and $r_1(\cdot;\beta)=r$ for every $\beta$; hence $\omega_{T,T'}(\eps')\to0$ as $\eps'\to0$. The exact cancellation $r_m\,\dd\nu=F_m\mathcal G(\partial_zv_m)^2\dd z$ and (V2) give a uniform Gaussian majorant for the $r_m$ term, and \Cref{II:appB:lem-gauss}(a) and (V2) give one for the $r$ term. Hence \[ \int_{T'}^\infty(r_m+r)\dd\nu\le C\int_{T'}^\infty(1+z)^2\exp(z-cz^2)\dd z=:\mathfrak t_2(T')\longrightarrow0\qquad(T'\to\infty). \] Consequently \[ \mathcal E\le\mathfrak t_1(T)+c_T^{-1}\bigl(C\,\omega_{T,T'}(\bar\eta)+\mathfrak t_2(T')\bigr). \] \proofstep{4}{Choice of $\bar\eta$.} Given $\eps>0$, first choose $T$ with $\mathfrak t_1(T)\le\eps/2$, then $T'$ with $\mathfrak t_2(T')\le c_T\eps/4$, and then $\bar\eta\in(0,1/4]$ so small that $C\omega_{T,T'}(\bar\eta)\le c_T\eps/4$ and, by the claim above, $\int r\dd\nu/\nu(\R)\le2x_\Phi+\eps$ and $\int r\dd\nu_Q/\nu_Q(\R)\le2x_\Phi+\eps$ for all admissible data. Then $\mathcal E\le\eps$ and $\mathcal M\le2x_\Phi+\eps$. Hence $\widetilde x_\gamma(q)\le x_\Phi+\eps$ in case (b), and $\widetilde x_\gamma(Q)\le x_\Phi+\eps/2$ in case (a). This proves \Cref{II:cpt:prop-topcheb}. \end{proof} \chapter{Proofs for the flat limit}\label{II:app:flat} This appendix proves the technical statements of \Cref{II:sec:flat}. \Cref{II:appC:ss-gauss} records the decay of the derivatives of $V$, which \Cref{II:sec:flat,II:sec:frsb} use in addition to \Cref{II:fun:gauss}. \Cref{II:appC:ss-tails} proves the uniform tail estimates of \Cref{II:flat:lem-tails}: a dilation turns every profile of a compact class into one fixed instance of the finite-cutoff problem of \Cref{II:sec:functional}, whose asymptotics give parts (a) and (d), and a transport estimate for the rescaled diffusion gives parts (b) and (c). \Cref{II:appC:ss-exp} proves the compact-class expansion \Cref{II:thm:compactexp}, and \Cref{II:appC:ss-zform} proves the differentiability of $I$ along ordered rays and the second-variation identity of \Cref{II:flat:prop-zform}. \section{Tails of \texorpdfstring{$V$}{V} and its derivatives}\label{II:appC:ss-gauss} We use the facts of \Cref{II:fun:gauss}, in particular $R'=-V$, $\Rbar'=1-V$, $z_-0$. For $z\ge0$ we use $R\le2\phi$ and $R\le R(0)<1$ (\Cref{II:fun:gauss}(a),(e)). By induction, using $R'=-zR-R^2$, $V^{(k)}=P_k(z,R)$ for a polynomial $P_k$ of total degree at most $k+2$ each of whose monomials contains $R$: this holds for $V=zR+R^2$, and $\frac{\dd}{\dd z}P(z,R)=\partial_zP-(zR+R^2)\partial_RP$. A monomial $z^iR^j$ of $P_k$ has $j\ge1$ and $i\le k+1$, so $|z^iR^j|\le z^iR\le2(1+z)^{k+1}\phi(z)$. (b) The three integrands are nonnegative and continuous, and integrable by (a): on $(-\infty,0]$ they are bounded by $C(1+|z|)^{-6}$, $1-V$, and $C(1+|z|)^{-8}$, and on $[0,\infty)$ they have Gaussian decay. They are not identically zero, because $00$, $\kappa\in\R$, $e^2s_{\rm hi}<1$, and $\zeta_\gamma=\zeta$ on $[0,s_{\rm hi}]$, then $U(s,y)=u_\gamma(1-e^2s,\kappa+ey)$ for $s\in[s_{\rm lo},s_{\rm hi}]$. \end{lemma} \begin{proof} With $\widetilde q=1-s/s_{\rm hi}$ and $\widetilde x=y/\sqrt{s_{\rm hi}}$ we have $\partial_{\widetilde q}\widetilde u=-s_{\rm hi}U_s$, $\widetilde u_{\widetilde x}=\sqrt{s_{\rm hi}}\,U_y$ and $\widetilde u_{\widetilde x\widetilde x}=s_{\rm hi}U_{yy}$, so the Parisi equation $\widetilde u_{\widetilde q}=-\frac12(\widetilde u_{\widetilde x\widetilde x}+\widetilde\gamma\,\widetilde u_{\widetilde x}^2)$ is \eqref{II:flat:eq-pde}. The function $\widetilde\gamma$ is nondecreasing because $\zeta$ is nonincreasing, and $\widetilde\gamma=1$ on $[r,1]$ because $\zeta=1$ on $[0,s_{\rm lo}]$. At $\widetilde q=r$ the terminal datum of \Cref{II:lem:regularity} is $\log\Phi(\widetilde x/\sqrt{1-r})=\log\Phi(y/\sqrt{s_{\rm lo}})$. The identity for the $L^1$ distance is the substitution $\widetilde q=1-s/s_{\rm hi}$, and the diffusion statement is Brownian scaling. For the last statement, the affine map $(\widetilde q,\widetilde x)\mapsto(1-e^2s_{\rm hi}(1-\widetilde q),\,\kappa+e\sqrt{s_{\rm hi}}\,\widetilde x)$ carries each Gaussian step of the Cole--Hopf recursion for $\widetilde\gamma$ at margin $0$ into the corresponding step for $\gamma$ at margin $\kappa$ on $[1-e^2s_{\rm hi},1)$, and at level $1-e^2s_{\rm lo}\ge Q_\gamma$ both solutions are $\log\Phi$ of the normalized position, by \eqref{II:fun:top}. This proves the claim for step profiles, and the general case follows by the $L^1$ limit that defines the solution (\Cref{II:def:orderparam}(iii)). \end{proof} Consequently \Cref{II:lem:regularity,II:lem:generator,II:lem:sandwich} apply to $U_\zeta$, with constants that depend only on $(s_{\rm lo},s_{\rm hi})$. We use them in the following rescaled form, for $\zeta\in\cK(d_1,S)$ (so $s_{\rm lo}=d_1$ and $s_{\rm hi}=S$), levels $s\in[d_1,S]$, and $Y$ the rescaled diffusion \eqref{II:appC:eq-Y} started at $(s,y)$ and run down to level $d_1$. Expectations are over $Y$. Under the dilation, $\widetilde\lambda=\ell/S$, $\widetilde b=\sqrt S\,U_y$, $\widetilde v=v$, and $\widetilde\psi=\Psi/\sqrt S$, where $\widetilde\psi=\widetilde\lambda\widetilde b+\widetilde x$ is the function of \eqref{II:fun:psi}. \begin{enumerate}[label=\textup{(R\arabic*)}] \item $U_y(s',Y_{s'})$ and $\Psi(s',Y_{s'})$ are martingales. At the wall, $U_y(d_1,y)=R(y/\sqrt{d_1})/\sqrt{d_1}$ and $\Psi(d_1,y)=\sqrt{d_1}\,\Rbar(y/\sqrt{d_1})$. \item $\Psi>0$, $\Psi_y=1-v\in(0,1)$, $\Psi_{yy}=-v_y\ge0$, and $v\le U_y\Psi\le(s/\ell)v$. \item $\widehat Y=Y/\ell$ satisfies $\dd\widehat Y=(\zeta\Psi/\ell^2)\dd\varsigma+\ell^{-1}\dd B_\varsigma$. \item If $\zeta$ is a step profile, then \begin{align*} 1-v(s,y)&=\E\Bigl[\bigl(1-V(Y_{d_1}/\sqrt{d_1})\bigr)\,\e^{-\int_{d_1}^s\frac\zeta\ell v}\Bigr],\qquad v(s,y)\le\frac{\ell(s)}{d_1}\,\E V(Y_{d_1}/\sqrt{d_1}),\\ v_y(s,y)&=d_1^{-1/2}\,\E\Bigl[V'(Y_{d_1}/\sqrt{d_1})\,\e^{-\int_{d_1}^s\frac\zeta\ell(3v-1)}\Bigr],\\ v_{yy}(s,y)&=d_1^{-1}\,\E\Bigl[V''(Y_{d_1}/\sqrt{d_1})\,\e^{-\int_{d_1}^s\frac\zeta\ell(4v-1)}\Bigr]\\ &\quad-3\,\E\int_{d_1}^s\frac{\zeta(s')}{\ell(s')}\,v_y(s',Y_{s'})^2\,\e^{-\int_{s'}^s\frac\zeta\ell(4v-1)}\dd s' . \end{align*} Here each integral in an exponent is taken along the path, over the levels indicated. Since $\int_{d_1}^s\zeta/\ell=\log(\ell(s)/d_1)\le\log(S/d_1)$ and $00$, and $v_x<0$ from \Cref{II:lem:generator}(a),(b) and \Cref{II:lem:regularity}(c), together with the sandwich \eqref{II:fun:sandwich}, read through the dilation; the wall values come from $U(d_1,y)=\log\Phi(y/\sqrt{d_1})$ and $\ell(d_1)=d_1$. (R3) is It\^o's formula, since $\dd\ell/\dd\varsigma=-\zeta$. For (R4), write $\mathcal G_s=\frac12\partial_{yy}+\zeta\,U_y\partial_y$. Differentiating \eqref{II:flat:eq-pde} gives, on each block of a step profile, \[ (\partial_s-\mathcal G_s)v=\frac\zeta\ell\,v(1-v),\qquad (\partial_s-\mathcal G_s)(-U_{yy})=-\zeta\,U_{yy}^2\le0, \] and, using $\partial_y(\mathcal G_sf)=\mathcal G_sf_y-(\zeta v/\ell)f_y$, \[ (\partial_s-\mathcal G_s)v_y=-\frac\zeta\ell(3v-1)v_y,\qquad (\partial_s-\mathcal G_s)v_{yy}=-\frac\zeta\ell(4v-1)v_{yy}-3\,\frac\zeta\ell\,v_y^2 . \] These are the rescaled forms of \Cref{II:lem:generator}(a). On each block $U$ is smooth with derivatives of polynomial growth, the drift $\zeta U_y$ is Lipschitz, and $v$, $v_y$, $v_{yy}$ are continuous in $s$ at the breakpoints. It\^o's formula on each block and the Markov property give the Feynman--Kac representations, with the values $v(d_1,y)=V(y/\sqrt{d_1})$, $v_y(d_1,y)=V'(y/\sqrt{d_1})/\sqrt{d_1}$ and $v_{yy}(d_1,y)=V''(y/\sqrt{d_1})/d_1$ at the wall, and with $-U_{yy}(d_1,\cdot)=V(\cdot/\sqrt{d_1})/d_1$ for the upper bound on $v=-\ell U_{yy}$. The analogous bounds for the Sherrington--Kirkpatrick model are cited before \Cref{II:flat:lem-signs}. The next lemma controls where the rescaled diffusion travels. Paths started far to the left stay far to the left, and paths started to the right stay to the right, with Gaussian probability. \begin{lemma}\label{II:appC:lem-transport} There is $C=C(d_1,S)$ such that for $\zeta\in\cK(d_1,S)$, $s\in[d_1,S]$ and $Y$ started at $(s,y)$ the following hold. \begin{enumerate}[label=\textup{(\alph*)}] \item If $y\le-C$, then with probability at least $1-2\e^{-y^2/C}$, $Y_{s'}\le y/C$ for all $s'\in[d_1,s]$. \item If $y\ge0$, then with probability at least $1-\e^{-y^2/C}$, $Y_{s'}\ge y/2$ for all $s'\in[d_1,s]$. \item $U_y(s',0)\le C$ for all $s'\in[d_1,S]$, and $0<\Psi(s',x)/\ell(s')\le C$ for $x\le0$. \end{enumerate} \end{lemma} \begin{proof} (c) Start $Y$ at $(s',0)$. The drift $\zeta U_y$ is nonnegative, so $Y_{d_1}$ is at least a centered Gaussian variable $N$ of variance $s'-d_1\le S$. By (R1), since $R$ is decreasing and $R(z)\le R(0)+z_-$ (\Cref{II:fun:gauss}(a)), \[ U_y(s',0)=\E\,\frac{R(Y_{d_1}/\sqrt{d_1})}{\sqrt{d_1}} \le\frac{R(0)+\E N_-/\sqrt{d_1}}{\sqrt{d_1}}\le\frac{R(0)+\sqrt{S/d_1}}{\sqrt{d_1}}=:C_\Psi . \] Next, $\partial_x(\Psi/\ell)=(1-v)/\ell\ge0$ by (R2), and $\Psi(s',0)=\ell U_y(s',0)$; so $0<\Psi(s',x)/\ell(s')\le U_y(s',0)\le C_\Psi$ for $x\le0$. (a) By (R3) and (c), as long as $Y\le0$ the drift of $\widehat Y$ is $(\zeta/\ell)(\Psi/\ell)\le C_\Psi\zeta/\ell$. Hence $\widehat Y_{s'}\le y/\ell(s)+C_\Psi\int_{s'}^s\zeta/\ell+N_{s'}$ up to the first level at which $Y$ reaches $0$, where $N_{s'}=\int\ell^{-1}\dd B$ is a continuous Gaussian martingale with quadratic variation at most $\int_{d_1}^S\ell^{-2}\le S/d_1^2$, and $\int\zeta/\ell\le\log(S/d_1)$. By the reflection principle, the event $\{\sup|N|\le|y|/(4\ell(s))\}$ has probability at least $1-2\exp(-y^2d_1^2/(32S^3))$, because $\ell(s)\le S$. On this event, and if $|y|\ge4SC_\Psi\log(S/d_1)$, we get $\widehat Y\le y/(2\ell(s))<0$ up to that level; hence $Y$ never reaches $0$, and $Y_{s'}=\ell(s')\widehat Y_{s'}\le d_1y/(2S)$ for all $s'\in[d_1,s]$. (b) The drift is nonnegative, so $Y_{s'}$ is at least $y$ plus a Brownian increment over a time at most $S$. By the reflection principle, the running minimum of that increment is below $-y/2$ with probability at most $2\Phi(-y/(2\sqrt S))\le\e^{-y^2/(8S)}$. \end{proof} \begin{proof}[Proof of \Cref{II:flat:lem-tails}] All constants depend only on $(d_1,S)$. Let $\zeta\in\cK(d_1,S)$ and let $\widetilde\gamma\in\cU_r$, $r=1-d_1/S$, be its dilation (\Cref{II:appC:lem-dilation} with $s_{\rm lo}=d_1$, $s_{\rm hi}=S$). Level $s\in[d_1,S]$ corresponds to $\widetilde q=1-s/S\in[0,r]$, and $\widetilde x=y/\sqrt S$. Under the dilation $\widetilde\lambda(\widetilde q)=\ell(s)/S$, the top of $\widetilde\gamma$ is $1-d/S$ with $d$ as in \Cref{II:flat:ss-def}, and the substitution $t=1-u/S$ gives $\int_{\widetilde q}^{1-d/S}\widetilde\lambda^{-1}=\int_d^s\ell^{-1}$. Hence the function $\omega$ of \Cref{II:lem:regularity}(e), computed for $\widetilde\gamma$, is \[ \log\frac{\ell(s)}S-\frac12\log\frac dS-\frac12\int_d^s\frac{\dd u}{\ell(u)}-\log\sqrt{2\pi} =\omega(s)-\frac12\log S , \] with $\omega(s)$ from \eqref{II:flat:eq-Gamma}. (a) By \Cref{II:lem:regularity}(e) for $\widetilde\gamma$ at margin $0$, there are $C,x_\infty$ such that $|\widetilde\lambda\widetilde b-(-\widetilde x)-\widetilde\lambda/(-\widetilde x)|\le C|\widetilde x|^{-3}$ for $\widetilde x\le-x_\infty$. Here $\widetilde\lambda\widetilde b=\ell U_y/\sqrt S$ and $-\widetilde x=|y|/\sqrt S$, and $\Psi=\ell U_y-|y|$ for $y<0$. The inequality becomes $|\Psi(s,y)-\ell(s)/|y||\le CS^2|y|^{-3}$ for $y\le-\sqrt S\,x_\infty$. (d) \emph{Left tail.} The first estimate of \Cref{II:lem:regularity}(e) at $\widetilde q=0$, together with $\widetilde x^2/(2\widetilde\lambda(0))=y^2/(2\ell_0)$ and $\log(-\widetilde x)=\log|y|-\frac12\log S$, gives \[ \Bigl|U(S,y)+\frac{y^2}{2\ell_0}+\log|y|-\omega(S)\Bigr|\le\frac{CS}{y^2}\qquad(y\le-\sqrt S\,x_\infty). \] For $y\le-1$ the left side is $|W_\zeta(y)|$. \emph{Bounded range.} By \Cref{II:lem:regularity}(b), $-C(1+y^2/S)\le U(S,y)\le0$, and $|\omega(S)|$ is bounded by \eqref{II:flat:eq-ellbounds}; hence $W_\zeta$ is bounded on $[-\max\{1,\sqrt S\,x_\infty\},0]$. \emph{Right tail.} For $y>0$, $W_\zeta(y)=U(S,y)$, and \Cref{II:lem:regularity}(d) at $\widetilde q=0$ gives $0\ge U(S,y)\ge-C\e^{-y^2/(8S)}$. Together, $|W_\zeta(y)|\le C(1+y^2)^{-1}$. (b) Let first $\zeta$ be a step profile, and let $C_{\rm T}$ be the constant of \Cref{II:appC:lem-transport}. For $y\le-C_{\rm T}$, let $E$ be the event of \Cref{II:appC:lem-transport}(a). On $E$, $Y_{d_1}/\sqrt{d_1}\le y/(C_{\rm T}\sqrt{d_1})$, and since $1-V$ is increasing, \Cref{II:appC:lem-gauss}(a) gives $1-V(Y_{d_1}/\sqrt{d_1})\le C(1+y^2)^{-1}$. The weight in (R4) is at most one, and $\Prob(E^c)\le2\e^{-y^2/C_{\rm T}}$. Hence $1-v(s,y)\le C(1+y^2)^{-1}$ for $y\le-C_{\rm T}$, and the same bound holds for $-C_{\rm T}\le y\le0$ because $1-v\le1$. For $y\ge0$, the second formula in (R4), \Cref{II:appC:lem-transport}(b) and $V(z)\le2(1+z)\phi(z)\le C\e^{-z^2/4}$ for $z\ge0$ give $v(s,y)\le C\e^{-y^2/C}$. (c) The same argument applies to the representations of $v_y$ and $v_{yy}$ in (R4), with the weights bounded by $S/d_1$, with $|V'(z)|+|V''(z)|\le C(1+|z|)^{-3}$ for $z\le0$ and Gaussian decay for $z\ge0$ (\Cref{II:appC:lem-gauss}(a)), and, in the source term of $v_{yy}$, with the bound on $v_y$ just obtained at every level $s'$: on the event $E$ every $Y_{s'}$ lies below $y/C_{\rm T}$. \emph{General profiles in (b) and (c).} Let $\zeta_n\in\cK(d_1,S)$ be the step approximations \eqref{II:flat:eq-steps} of $\zeta$, so that $\zeta_n\to\zeta$ in $L^1$. By \Cref{II:appC:lem-dilation} their dilations converge in $L^1(0,1)$, so by \Cref{II:lem:regularity}(f) the derivatives $\partial_y^kU_{\zeta_n}(s,y)$, $k\le4$, converge pointwise; also $\ell_n\to\ell$ uniformly. Hence $v$, $v_y$, $v_{yy}$ of $\zeta_n$ converge pointwise to those of $\zeta$, and the bounds, whose constants do not depend on $n$, pass to the limit. \emph{Continuity.} Let $\zeta_n\to\zeta$ in $\cK(d_1,S)$. As in the previous paragraph, $U_{\zeta_n}(S,y)\to U_\zeta(S,y)$ for every $y$, and $\omega_n(S)\to\omega(S)$ because $\ell_n\to\ell$ uniformly and $\ell_n,\ell\ge d_1$ on $[d_1,S]$. Thus $W_{\zeta_n}\to W_\zeta$ pointwise, and the majorant $C(1+y^2)^{-1}$ of (d) gives convergence in $L^1(\R)$ by dominated convergence. Finally, $|\ell_n-\ell|\le\|\zeta_n-\zeta\|_{L^1}$, so $|\widetilde A(\zeta_n)-\widetilde A(\zeta)|\le\frac18(2S+d_1)d_1^{-2}\|\zeta_n-\zeta\|_{L^1}$. \end{proof} \section{Proof of the compact-class expansion}\label{II:appC:ss-exp} \begin{proof}[Proof of \Cref{II:thm:compactexp}] Constants below depend only on $(d_1,S)$ and on $\theta_\pm$. Let $e$ be so small that $e^2S\le\frac12$ and that \Cref{II:lem:level0} applies; in particular $|\kappa|\le e/2$ by \Cref{II:lem:level0}(a). Let $\gamma\in\cU$ have rescaled profile $\zeta\in\cK(d_1,S)$, and use the declared level $q_1=1-e^2S\ge\frac12$, with $Y_1=(B_{q_1}-\kappa)/e$ and $k_1=\kappa/\sqrt{q_1}$. The level $q_1$ may lie below $q_0=\min\supp\mu_\gamma$; since $\zeta=0$ on $(S,\infty)$, $\gamma=0$ on $[0,q_1)$. We divide the proof into five steps. \proofstep{1}{An exact decomposition.} By \Cref{II:appC:lem-dilation}, $U_\zeta(S,y)=u_\gamma(q_1,\kappa+ey)$, so $U_\zeta(S,\cdot)$ is the function $U(S,\cdot)$ of \Cref{II:def:rescaled}. Also $\Gamma(S)$ is the function $\Gamma$ of \Cref{II:def:rescaled} at level $S$ (see after \eqref{II:flat:eq-Gamma}), so $\omega(S)$ is the constant $\log\ell_0-\Gamma(S)-\log\sqrt{2\pi}$ of \Cref{II:prop:decomp} at that level. Hence \Cref{II:prop:decomp}(a), applied at the declared level $q_1$ (as permitted by the paragraph after \Cref{II:lem:valueformula}), holds with $W=W_\zeta$ and with $Y_1$ in place of $Y_0$. Put $\rho=\alpha\Phi(k_1)$; this is the number $\rho$ of \Cref{II:prop:decomp} at the declared level. Since $\widetilde\Gamma=\Gamma(S)-\frac12\log\ell_0$ at the level $S$, $\frac12(1+\rho)\log\ell_0+(1-\rho)\widetilde\Gamma=\rho\log\ell_0+(1-\rho)\Gamma(S)$, and \Cref{II:prop:decomp}(a) reads \begin{equation}\label{II:appC:eq-decomp} \cP_\kappa(\gamma)=K_e(q_1)+\frac{c_e(q_1)}{\ell_0}+\rho\log\ell_0+(1-\rho)\Gamma(S)+\alpha\E W_\zeta(Y_1), \end{equation} where $c_e(q_1)=q_1(1-\alpha m(k_1))/(2e^2)$ and $K_e(q_1)=\log e-\alpha\E[\log|Y_1|;Y_1\le-1]-\rho\log\sqrt{2\pi}$ are the coefficients of \Cref{II:prop:decomp} at the level $q_1$. \proofstep{2}{The coefficients.} By \Cref{II:lem:level0}(b) at the level $q_1$, that is, with $s_0$ replaced by $S$: \begin{enumerate}[label=\textup{(\roman*)}] \item $c_e(q_1)=2\phi_0\theta+e[(\phi_0-\frac1{8\phi_0})\theta+\frac S8]+O(e^2)$, and $\ell_0\ge d_1$; \item $\rho=1+\frac e4+O(e^2)$; since $\ell_0\in[d_1,S]$ and $\Gamma(S)\in[\frac12\log d_1,\frac12\log d_1+S/(2d_1)]$, $\rho\log\ell_0=\log\ell_0+\frac e4\log\ell_0+O(e^2)$ and $(1-\rho)\Gamma(S)=-\frac e4\Gamma(S)+O(e^2)$; \item $|K_e(q_1)-K_e|\le Ce^2L_e$. \end{enumerate} \proofstep{3}{The remainder.} The density of $Y_1$ is $p_e(y)=(e/\sqrt{q_1})\,\phi((ey+\kappa)/\sqrt{q_1})$. Since $q_1^{-1/2}\le1+e^2S$, $\phi_0-\phi(w)\le\phi_0\min\{1,w^2/2\}$ and $(ey+\kappa)^2/q_1\le2e^2(1+|y|)^2$, \begin{align*} |\alpha p_e(y)-2e\phi_0|&\le2e\phi_0\bigl|(1+\tfrac e2)q_1^{-1/2}-1\bigr| +\frac{(2+e)e\phi_0}{\sqrt{q_1}}\min\bigl\{1,e^2(1+|y|)^2\bigr\} \\ &\le Ce\bigl[e+\min\{1,e(1+|y|)\}\bigr]. \end{align*} By \Cref{II:flat:lem-tails}(d), splitting the integral at $|y|=1/e$, \[ \Bigl|\alpha\E W_\zeta(Y_1)-2e\phi_0\int_\R W_\zeta\Bigr| \le Ce\int_\R\frac{e+\min\{1,e(1+|y|)\}}{1+y^2}\dd y\le Ce^2L_e . \] \proofstep{4}{Summation.} Summing Steps 2 and 3 in \eqref{II:appC:eq-decomp}, the terms of order one are $K_e+2\phi_0\theta/\ell_0+\log\ell_0=K_e+\Psi_\theta(\ell_0)$, the other terms are $O(e^2L_e)$, and the terms of order $e$ are \begin{gather*} g_\theta(\ell_0)+\frac S{8\ell_0}-\frac{\Gamma(S)}4+2\phi_0\int W_\zeta,\\ \frac S{8\ell_0}-\frac{\Gamma(S)}4=\frac18\Bigl[\int_{d_1}^S\Bigl(\frac1{\ell_0}-\frac1{\ell(s)}\Bigr)\dd s+\frac{d_1}{\ell_0}-\log d_1\Bigr]=\widetilde A(\zeta). \end{gather*} This proves \eqref{II:flat:eq-compactexp}. The continuity of $\cP^{(1)}$ is part of \Cref{II:flat:lem-tails}. \proofstep{5}{The scaling identity.} Let $c=\ell_0$ and $\widetilde\zeta=\zeta(c\,\cdot)\in\cK^1(d_1/c,S/c)$. The equation \eqref{II:flat:eq-pde}, including its initial condition, is invariant under $s\mapsto cs$, $y\mapsto\sqrt c\,y$, $\zeta\mapsto\zeta(c\,\cdot)$, so $U_\zeta(s,y)=U_{\widetilde\zeta}(s/c,y/\sqrt c)$. Also $\ell_\zeta(s)=c\,\ell_{\widetilde\zeta}(s/c)$, and the substitution $s=cs'$ in \eqref{II:flat:eq-Gamma} gives $\Gamma_\zeta(S)=\Gamma_{\widetilde\zeta}(S/c)+\frac12\log c$ and $\omega_\zeta(S)=\omega_{\widetilde\zeta}(S/c)+\frac12\log c$, where the functions of $\widetilde\zeta$ are computed with the class parameters $(d_1/c,S/c)$. Hence, with $y=\sqrt c\,y'$, \[ W_\zeta(y)=W_{\widetilde\zeta}(y')+\tfrac12\log c\,\bigl[\1_{\{y'\le-c^{-1/2}\}}-\1_{\{y'\le0\}}\bigr] +\log|y'|\,\bigl[\1_{\{y'\le-c^{-1/2}\}}-\1_{\{y'\le-1\}}\bigr]. \] Put $b=c^{-1/2}$, so that $\frac12\log c=-\log b$. The first bracket equals $-\1_{(-b,0]}$, and its term integrates to $b\log b$. If $b\ge1$, the second bracket equals $-\1_{(-b,-1]}$, and its term integrates to $-\int_1^b\log u\,\dd u=-(b\log b-b+1)$; if $b<1$, it equals $\1_{(-1,-b]}$, and its term integrates to $\int_b^1\log u\,\dd u=-1-b\log b+b$. In both cases the two terms integrate to $b-1$, so $\int W_\zeta=\sqrt c\int W_{\widetilde\zeta}+1-\sqrt c$. The substitution $s=cs'$ also gives $\widetilde A(\zeta)=A(\widetilde\zeta)-\frac18\log c$. Therefore $\cP^{(1)}(\zeta)=A(\widetilde\zeta)+2\phi_0\sqrt c\,(\int W_{\widetilde\zeta}-1)-\frac18\log c+2\phi_0$, which is \eqref{II:flat:eq-scaling}. \end{proof} \section{Differentiability along ordered rays and the second variation}\label{II:appC:ss-zform} We work in the setting of \Cref{II:flat:ss-conv}: $T_\vartheta=T_0+\vartheta K$, with $T_0,T_1$ the inverse profiles of normalized step profiles on a common grid and $K\ge0$. For $\vartheta\in[0,1]$ all $\zeta_\vartheta$ lie in the fixed class $\cK^1(l_0,\max\{T_0(1),T_1(1)\})$, so the constants of \Cref{II:flat:lem-tails,II:appC:lem-transport} are uniform in $\vartheta$. In the variable $l$ we have $v=lc$, $v_y=lc_y$, and $v_{yy}=lc_{yy}$, and the diffusion of \Cref{II:flat:def-density} is the rescaled diffusion \eqref{II:appC:eq-Y} after the change of variable $s=T_\vartheta(l)$, because $\sigma_\vartheta\,\zeta_\vartheta(T_\vartheta(l))=1$. \begin{lemma}\label{II:appC:lem-derivs} For each $(l,y)$, $\vartheta\mapsto U^\vartheta(l,y)$ is of class $C^2$, and $\dot U,\ddot U$ solve \eqref{II:flat:eq-var}, with \[ \dot U(l,y)=-\frac12\,\E\int_{l_0}^l\dot\sigma(\lambda)\,c(\lambda,Y_\lambda)\dd\lambda,\qquad \ddot U(l,y)=\E\int_{l_0}^l\bigl(\dot U_y^2+\dot\sigma\,\dot U_{yy}\bigr)(\lambda,Y_\lambda)\dd\lambda, \] where $Y$ starts at $(l,y)$ and runs downward in $l$. Put $\dot\Gamma(l)=\frac12\int_0^l\dot\sigma(\lambda)\lambda^{-1}\dd\lambda$. There is $C$, uniform in $\vartheta\in[0,1]$ and $l\in[l_0,1]$, such that: \begin{enumerate}[label=\textup{(\alph*)}] \item $|\dot U(l,y)+\dot\Gamma(l)\1_{\{y\le0\}}|+|\ddot U(l,y)|\le C(1+y^2)^{-1}$; \item $|c_y|+|c_{yy}|+|\dot U_y|+|\dot U_{yy}|+|\mathcal Z_y|+|\mathcal Z_{yy}|\le C(1+|y|)^{-3}$ for $y\le0$, and $\le C\e^{-y^2/C}$ for $y\ge0$, where $\mathcal Z=\dot U+\frac K2c$. \end{enumerate} Consequently $I\in C^2([0,1])$, and \begin{align*} I'(\vartheta)&=\int_\R\bigl[\dot U(1,y)+\dot\Gamma(1)\1_{\{y\le0\}}\bigr]\dd y,\\ I''(\vartheta)&=\int_\R\ddot U(1,y)\dd y=\int_{l_0}^1\!\!\int_\R p_l\bigl(\dot U_y^2+\dot\sigma\,\dot U_{yy}\bigr)\dd y\,\dd l . \end{align*} \end{lemma} \begin{proof} \emph{Smoothness.} On each block, \eqref{II:flat:eq-colehopf} expresses $U^\vartheta$ as $\sigma_j\log$ of a Gaussian integral of $\exp(U^\vartheta(l_{j-1},\cdot)/\sigma_j)$, where $\sigma_j=\sigma_j(\vartheta)$ is affine in $\vartheta$ and at least one on $[0,1]$, hence positive on a neighborhood of $[0,1]$. The integrand is at most one, and derivatives in $\vartheta$ and $y$ produce polynomial factors. The Gaussian integral is at least $\exp(-C(1+y^2))$, because $U^\vartheta(l_{j-1},x)\ge U^\vartheta(l_{j-1},0)+b(l_{j-1},0)x-x^2/(2l_{j-1})$ by concavity and $c\le1/l$ (\Cref{II:flat:lem-signs}, which holds for every positive $\sigma$). To control derivatives of the logarithm, use the tilted Gaussian moment argument of \Cref{II:appA:rec}(iii): the negative log density has curvature between $\tau^{-1}$ and $\tau^{-1}+(\sigma_j l_{j-1})^{-1}$, where $\tau=\sigma_j(l-l_{j-1})$, and its mode has at most linear growth in $y$. Its polynomial moments therefore have polynomial growth in $y$, uniformly for block parameters in compact sets. Parameter derivatives of the Gaussian variance can be taken as $\partial_\tau P_\tau f=\frac12P_\tau f_{yy}$, so this argument remains valid as $\tau\downarrow0$. Induction over the finite set of blocks now shows that $U^\vartheta$ is smooth in $(\vartheta,y)$ on each block, with derivatives of polynomial growth in $y$, locally uniformly in $\vartheta$, and continuous in $l$. Differentiating \eqref{II:flat:eq-lpde} in $\vartheta$ gives \eqref{II:flat:eq-var}. The drift $b$ is Lipschitz in $y$, and the equations for $\dot U,\ddot U$ are linear, without potential, with sources of polynomial growth and zero data at $l_0$; It\^o's formula on each block and the Markov property give the representations. \emph{Tails.} The bounds (b) for $c_y,c_{yy}$ are \Cref{II:flat:lem-tails}(c), because $l\ge l_0$. Differentiate the representation of $\dot U$ in $y$ along the flow. Its Jacobian $J_\lambda=\partial_yY_\lambda=\exp(-\int_\lambda^lc(\lambda',Y_{\lambda'})\dd\lambda')$ lies in $(0,1]$ and satisfies $\partial_yJ_\lambda=-J_\lambda\int_\lambda^lc_y(\lambda',Y_{\lambda'})J_{\lambda'}\dd\lambda'$. Since $c_y$ and $c_{yy}$ are bounded, differentiation under the expectation is legitimate, and \[ |\dot U_y(l,y)|\le C\,\E\int_{l_0}^l|c_y(\lambda,Y_\lambda)|\dd\lambda,\qquad |\dot U_{yy}(l,y)|\le C\,\E\int_{l_0}^l\Bigl(|c_{yy}|+|c_y|\int_\lambda^l|c_y|\Bigr)(\lambda,Y_\lambda)\dd\lambda . \] By \Cref{II:appC:lem-transport}(a),(b), read in the variable $l$, for $y\le-C$ the path stays below $y/C$ with probability at least $1-2\e^{-y^2/C}$, and for $y\ge0$ it stays above $y/2$ with probability at least $1-\e^{-y^2/C}$. Since $c_y$ and $c_{yy}$ are bounded, this transfers their tails to $\dot U_y$ and $\dot U_{yy}$, and hence to $\mathcal Z_y=\dot U_y+\frac K2c_y$ and $\mathcal Z_{yy}=\dot U_{yy}+\frac K2c_{yy}$. For (a), when $y\le0$, \[ \dot U(l,y)+\dot\Gamma(l)=\frac12\,\E\int_{l_0}^l\frac{\dot\sigma(\lambda)}\lambda\bigl(1-v(\lambda,Y_\lambda)\bigr)\dd\lambda , \] and $1-v$ is bounded by one and by $C(1+Y_\lambda^2)^{-1}$ when $Y_\lambda\le0$ (\Cref{II:flat:lem-tails}(b)); bounding $|\dot\sigma|$ and using the transport estimate gives $|\dot U+\dot\Gamma|\le C(1+y^2)^{-1}$. No sign of $\dot\sigma=K'$ is required here: an ordered segment $K\ge0$ can have $K'$ of both signs. When $y>0$, $|\dot U|\le C\,\E\int c(\lambda,Y_\lambda)\dd\lambda$, and $c=v/\lambda$ has Gaussian decay on the right. The bound on $\ddot U$ follows from its representation, (b), and the transport estimate. \emph{Derivatives of $I$.} In \eqref{II:flat:eq-Il}, the counterterms are affine in $\sigma$, so the first and second $\vartheta$-derivatives of the integrand are $\dot U(1,y)+\dot\Gamma(1)\1_{\{y\le0\}}$ and $\ddot U(1,y)$. They are continuous in $\vartheta$ and dominated by $C(1+y^2)^{-1}$ uniformly in $\vartheta$, so we may differentiate under the integral. Finally, the representation of $\ddot U$ and Fubini's theorem give $\int\ddot U(1,y)\dd y=\int_{l_0}^1\int p_l\,(\dot U_y^2+\dot\sigma\dot U_{yy})$, because $p_l$ is the density of the image of Lebesgue measure under the flow from level $1$ to level $l$. The integrand is absolutely integrable by (b) and $p_l\le1/l_0$. \end{proof} The same argument gives a transport identity for the flat density. Let $l_0\le l'0$. Finally $x_{\mathrm f}=A_V/(2B_V)=\frac12-c_0/(4B_V)$ is positive because $A_V>0$, and $x_{\mathrm f}<\frac14$ because $c_0=\int V(1-V)>\int V^2(1-V)=B_V$, as $00$, and $g_c(z_0)\ge\varkappa_2$ whenever $\rho\in[\frac14,1)$, $c\in[-1,0]$ and $z_0\in(c,c+\sqrt{1-\rho})$. \item Let $p_0=\nu f_0$ be a probability density with $f_0$ nonincreasing, let $Z_0\sim p_0$, and let $Z_\rho$ have the law $\int p_0(z_0)\Pi_\rho(z_0,\cdot)\dd z_0$. Then $Z_\rho$ has the density $\nu f_\rho$ with $f_\rho(z)=\int\Pi_\rho(z,z_0)f_0(z_0)\dd z_0$ nonincreasing, and \[ \Prob(Z_\rho\le c)-\Prob(Z_0\le c)=\int\nu\bigl(f_0-f_0(c)\bigr)g_c\le0\qquad(c\in\R). \] If $f_0$ is strictly decreasing, then $f_\rho$ is strictly decreasing, the inequality is strict for every $c$, and $\E h(Z_\rho)<\E h(Z_0)$ for every bounded, strictly decreasing $h\in C^1(\R)$ with $h'$ integrable. \end{enumerate} \end{lemma} \begin{proof} (a) Consider the order parameter $\1_{[0,1]}$ at the margin $0$. By \eqref{II:fun:top}, $u(q,x)=\log H(q,x)$ with $H(q,x)=\Phi(x/\sqrt{1-q})$ on $[0,1)$, and $H_q+\frac12H_{xx}=0$. Hence the diffusion \eqref{II:fun:X} with drift $b=H_x/H$, started at $(0,z_0)$, is the Doob transform of Brownian motion by $H$, with transition densities $\phi_{q_2-q_1}(x_2-x_1)H(q_2,x_2)/H(q_1,x_1)$ by \eqref{II:fun:doob}. Put $Z_\varsigma=X_q/\sqrt{1-q}$ with $\varsigma=\log(1/(1-q))$. It\^o's formula gives $\dd Z=(R(Z)+\frac12Z)\dd\varsigma+(1-q)^{-1/2}\dd B_q$, and the martingale part is a Brownian motion in the time $\varsigma$; so $Z$ solves \eqref{II:flat:eq-topdiff}, whose drift has the bounded derivative $\frac12-V$, so that its solution is unique. In the variables $x_i=\sqrt{1-q_i}\,z_i$, with $\rho=(1-q_2)/(1-q_1)$, the transition density of $Z$ becomes $\sqrt{1-q_2}\,\phi_{q_2-q_1}(\sqrt{1-q_2}\,z_2-\sqrt{1-q_1}\,z_1)\Phi(z_2)/\Phi(z_1)=\Pi_\rho(z_1,z_2)$. This proves the first assertion; the semigroup property follows from the Markov property and time homogeneity. For detailed balance, $\nu(z_0)\Pi_\rho(z_0,z)=\Phi(z_0)\Phi(z)\sqrt\rho\,(2\pi(1-\rho))^{-1/2}\e^{E}$ with $E=(2\sqrt\rho\,zz_0-\rho z^2-\rho z_0^2)/(2(1-\rho))$, which is symmetric in $(z_0,z)$. Next $(\log\nu)'=2R+z>0$, because $R(z)>z_-$ (\Cref{II:fun:gauss}(a)). For $z<-1$, $\Phi(z)\le\phi(z)/|z|$ (\Cref{II:fun:gauss}(b)) gives $\nu(z)\le\e^{-z^2/2}/(2\pi z^2)$, so $\int_{-\infty}^c\nu<\infty$. The mixed derivative $\partial_z\partial_{z_0}\log\Pi_\rho(z_0,z)=\sqrt\rho/(1-\rho)>0$ gives the strict likelihood ratio order. For the representation, let $Y'\sim N(z_0/\sqrt\rho,(1-\rho)/\rho)$ and $G'\sim N(0,1)$ be independent. The density of $Y'$ is $\sqrt\rho\,\phi_{1-\rho}(\sqrt\rho z-z_0)$, and $\Prob(G'\le Y')=\Phi(z_0)$, so $\Pi_\rho(z_0,\cdot)$ is the law of $Y'$ conditioned on $\{G'\le Y'\}$. Put $W=G'-Y'$, so that $\Var W=1/\rho$ and $\Cov(Y',W)=-(1-\rho)/\rho$. Gaussian regression gives $Y'=\sqrt\rho\,z_0-(1-\rho)W+\sqrt{1-\rho}\,N$ with $N\sim N(0,1)$ independent of $W$. Writing $-W=\mathsf T/\sqrt\rho$ with $\mathsf T\sim N(z_0,1)$, the event $\{G'\le Y'\}$ is $\{\mathsf T\ge0\}$, which proves the representation. (b) Detailed balance and Tonelli's theorem give \[ \int\nu(z_0)\Pi_\rho\bigl(z_0,(-\infty,c]\bigr)\dd z_0=\int_{-\infty}^c\nu(z)\Bigl(\int\Pi_\rho(z,z_0)\dd z_0\Bigr)\dd z=\int_{-\infty}^c\nu<\infty . \] Subtracting $\int_{-\infty}^c\nu$ gives $\int\nu g_c=0$, and the sum of the two integrals bounds $\int\nu|g_c|$. (c) On $[0,1]$, $2R+z\ge2R\ge2\phi(1)$. On $[-1,0]$, $R\ge R(0)=2\phi_0$, so $2R+z\ge4\phi_0-1>0$. Let $\rho\in[\frac14,1)$, $c\in[-1,0]$ and $z_0\in(c,c+\sqrt{1-\rho})\subset(-1,1)$. Then $g_c(z_0)=\Pi_\rho(z_0,(-\infty,c])$. On the event $\{\mathsf T\le1,\ N\le-4\}$ the variable of (a) is at most $\sqrt\rho z_0+2(1-\rho)-4\sqrt{1-\rho}$, because $(1-\rho)/\sqrt\rho\le2(1-\rho)$. Moreover $c-\sqrt\rho z_0\ge c(1-\sqrt\rho)-\sqrt\rho\sqrt{1-\rho}\ge-(1-\rho)-\sqrt{1-\rho}$, and $3(1-\rho)\le3\sqrt{1-\rho}$. Hence the variable is at most $c$ on this event. Since $\Prob(\mathsf T\le1)\ge\Phi(1-z_0)-\Phi(-z_0)\ge\phi(2)$ and $\mathsf T$, $N$ are independent, $g_c(z_0)\ge\varkappa_2$. (d) By detailed balance, $\int\nu(z_0)f_0(z_0)\Pi_\rho(z_0,z)\dd z_0=\nu(z)f_\rho(z)$. By the likelihood ratio order (see \cite{ShakedShanthikumar2007}), $\Pi_\rho(z,\cdot)$ is stochastically increasing in $z$, strictly because it has a positive density; so $f_\rho$ is nonincreasing, and strictly decreasing if $f_0$ is. Next, $\Prob(Z_\rho\le c)-\Prob(Z_0\le c)=\int p_0g_c=\int\nu(f_0-f_0(c))g_c$, where we used $\int\nu g_c=0$ and the integrability of $\nu g_c$ from (b). The integrand is nonpositive: for $z_0\le c$, $g_c\le0$ and $f_0(z_0)\ge f_0(c)$; for $z_0>c$, $g_c\ge0$ and $f_0(z_0)\le f_0(c)$. If $f_0$ is strictly decreasing, the integrand is negative on $(-\infty,c)$, where $g_c=-\Pi_\rho(\cdot,(c,\infty))<0$. Finally, for $h$ as stated, $h(z)=h(+\infty)+\int_z^\infty|h'(c)|\dd c$, and Tonelli's theorem gives $\E h(Z_\rho)-\E h(Z_0)=\int|h'(c)|[\Prob(Z_\rho\le c)-\Prob(Z_0\le c)]\dd c<0$. \end{proof} We now describe the Parisi diffusion of $\gamma_{\mathrm d}$ at $\kd$. Write $\beta=\beta_{\mathrm d}=\sqrt{\delta_{\mathrm d}/\qd}$ and $k=\kd/\sqrt{\qd}$, and let $Z_{\mathrm d}=(X_{\qd}-\kd)/\sqrt{\delta_{\mathrm d}}$. Since $X_{\qd}=B_{\qd}$, $Z_{\mathrm d}$ has the density $w_{\mathrm d}(z)=\beta\phi(k+\beta z)$. \emph{Above the atom.} On $[\qd,1)$, $\gamma_{\mathrm d}=1$ and $u(q,x)=\log\Phi((x-\kd)/\sqrt{1-q})$ by \eqref{II:fun:top}. The computation in the proof of \Cref{II:appD:lem-kernel}(a), with the margin $\kd$ and the initial level $\qd$, shows that $Z_\rho:=(X_q-\kd)/\sqrt{1-q}$, with $\rho=\hat s(q)\in(0,1]$, starts from $Z_1=Z_{\mathrm d}$ and has the transition kernels $\Pi$. By \eqref{II:fun:topderiv}, $v(q,X_q)=V(Z_\rho)$, so $S_{\mathrm d}(q)=\alpha\E V(Z_\rho)^2$. \emph{Below the atom.} On $[0,\qd)$, $\gamma_{\mathrm d}=0$, $X=B$, and $\lambda_{\mathrm d}=\delta_{\mathrm d}$. By \Cref{II:lem:generator}(a), $\cL c=\gamma c^2=0$ there, so $c(q,\cdot)$ is the heat flow of $c(\qd,\cdot)$. With $Z'_q=(X_q-\kd)/\sqrt{\delta_{\mathrm d}}$ and $t=\hat s(q)-1=(\qd-q)/\delta_{\mathrm d}$, this gives $v(q,X_q)=\E[V(Z'_q+\sqrt tG)\mid Z'_q]$, where $G\sim N(0,1)$ is independent of $Z'_q$, and $Z'_q+\sqrt tG$ has the law of $Z_{\mathrm d}$. We use the positive constants $\varkappa_1,\varkappa_2$ of \Cref{II:appD:lem-kernel}(c) and \begin{align*} \varkappa_3&=(2\Phi(1)-1)\min_{[-2,2]}|V'|,\\ \varkappa_4&=\frac{(V(-1)^2-V(0)^2)\,\nu(-1)\,\phi(1)\,\varkappa_1\varkappa_2}{4\nu(1)},\\ \varkappa_5&=2\phi(1)\varkappa_3^2 ; \end{align*} they are positive because $V'<0$ (\Cref{II:fun:gauss}(a)). \begin{proof}[Proof of \Cref{II:frsb:prop-rsobstacle}] By \Cref{II:prop:crossing}(iii) and $|\kd|\le e$, we may choose $e_{\mathrm o}$ so small that for $00$. (a) \emph{Below $\qd$.} By \eqref{II:frsb:eq-Sprime} with $\gamma_{\mathrm d}=0$, $S_{\mathrm d}'=\alpha\cN>0$ on $[0,\qd)$, since $v_x<0$; so $S_{\mathrm d}$ is strictly increasing on $[0,\qd]$. \emph{Above $\qd$.} Let $f_0=w_{\mathrm d}/\nu$. Then $(\log f_0)'(z)=-\beta(k+\beta z)-(2R(z)+z)$. For $z\in[0,1]$, $2R(z)+z\ge2\phi(1)>\phi_0$; for $z>1$, $2R+z>1>\phi_0$; in both cases $-\beta(k+\beta z)\le\beta|k|$. For $z=-t<0$, $-\beta(k+\beta z)\le\beta|k|+\beta^2t\le\beta|k|+\frac12t$, and \[ 2R(-t)-\tfrac32t>\max\bigl\{\tfrac12t,\ 4\phi_0-\tfrac32t\bigr\}\ge\phi_0, \] because $R(-t)>t$ (\Cref{II:fun:gauss}(b)) and $R(-t)>R(0)=2\phi_0$, as $R$ is decreasing. In all cases $(\log f_0)'<\beta|k|-\phi_0\le0$, so $f_0$ is strictly decreasing. Let $0<\rho'<\rho\le1$. By \Cref{II:appD:lem-kernel}(d) (with nothing to prove if $\rho=1$), $Z_\rho$ has a density $\nu f_\rho$ with $f_\rho$ strictly decreasing. By the semigroup property, $Z_{\rho'}$ is obtained from $Z_\rho$ by the kernel $\Pi_{\rho'/\rho}$. Applying \Cref{II:appD:lem-kernel}(d) again, with the initial density $\nu f_\rho$ and the strictly decreasing function $h=V^2$ (for which $|h'|=2V|V'|$ is integrable), gives $\E V(Z_{\rho'})^2<\E V(Z_\rho)^2$. Hence $S_{\mathrm d}$ is strictly decreasing on $[\qd,1)$. Since $S_{\mathrm d}(\qd)=1$ and $D_{\mathrm d}(\qd)=0$, the function $D_{\mathrm d}(q)=\int_{\qd}^q(S_{\mathrm d}-1)\lambda_{\mathrm d}^{-2}$ is positive on $[0,\qd)$ and negative on $(\qd,1)$. As $\cH_{\mathrm d}'=-D_{\mathrm d}/2$, the point $\qd$ is the unique minimum point of $\cH_{\mathrm d}$ on $[0,1)$. (b) \emph{Above $\qd$.} Let $\rho\in[\frac14,1)$ and let $q$ be the level with $\hat s(q)=\rho$. Put $h=V^2$. Since $h$ is decreasing with $h(+\infty)=0$, Tonelli's theorem gives $\E h(Z)=\int|h'(c)|\Prob(Z\le c)\dd c$ for every random variable $Z$. Since $1-S_{\mathrm d}(q)=\alpha(\E h(Z_1)-\E h(Z_\rho))$, and since the integrand in \Cref{II:appD:lem-kernel}(d) is nonpositive everywhere, \begin{align*} 1-S_{\mathrm d}(q)&=\alpha\int|h'(c)|\bigl[\Prob(Z_1\le c)-\Prob(Z_\rho\le c)\bigr]\dd c,\\ \Prob(Z_1\le c)-\Prob(Z_\rho\le c)&=\int\nu\,|f_0-f_0(c)|\,|g_c| . \end{align*} Restrict to $c\in[-1,0]$ and $z_0\in(c,c+\sqrt{1-\rho})\subset(-1,1)$. There $g_c(z_0)\ge\varkappa_2$ by \Cref{II:appD:lem-kernel}(c), and $\nu(z_0)\ge\nu(-1)$. On $[-1,1]$ we have $w_{\mathrm d}\ge\beta\phi(1)$, because $|k+\beta z|\le1$, and $\nu\le\nu(1)$, so $f_0\ge\beta\phi(1)/\nu(1)$; also $|(\log f_0)'|\ge\varkappa_1-\beta|k|-\beta^2\ge\frac12\varkappa_1$. Hence $|f_0(z_0)-f_0(c)|\ge\beta\phi(1)\varkappa_1(z_0-c)/(2\nu(1))$, and integrating over $z_0$ and then over $c\in[-1,0]$, where $\int_{-1}^0|h'|=V(-1)^2-V(0)^2$, \[ 1-S_{\mathrm d}(q)\ge\alpha \beta\,\varkappa_4\,(1-\rho)\qquad(\tfrac14\le\hat s(q)=\rho<1). \] Since $D_{\mathrm d}(\qd)=0$, \eqref{II:frsb:eq-anchor} for $\gamma_{\mathrm d}$ gives $G_{\mathrm d}(q)=\cA_{\mathrm d}(\qd,q)$. For $q\ge\qd$ we have $\lambda_{\mathrm d}(r')=1-r'=\delta_{\mathrm d}\hat s(r')$, and in the variable $\hat s'=\hat s(r')$, \begin{align*} G_{\mathrm d}(q)&=\frac12\int_{\hat s(q)}^1\bigl(\hat s'-\hat s(q)\bigr)\bigl(1-S_{\mathrm d}\bigr)\hat s'^{-2}\dd\hat s'\\ &\ge\frac{\alpha \beta\varkappa_4}2\int_{\hat s(q)}^1(\hat s'-\hat s(q))(1-\hat s')\dd\hat s'=\frac{\alpha \beta\varkappa_4}{12}\bigl(1-\hat s(q)\bigr)^3 \end{align*} for $\frac14\le\hat s(q)\le1$. \emph{Below $\qd$.} Let $q<\qd$ with $t=\hat s(q)-1\in(0,1]$. Since $\alpha\E V(Z_{\mathrm d})^2=S_{\mathrm d}(\qd)=1$, the description below the atom gives \[ 1-S_{\mathrm d}(q)=\alpha\E\Var\bigl(V(Z'_q+\sqrt tG)\bigm|Z'_q\bigr). \] For a $C^1$ function $F$ with bounded derivative, the Cauchy--Schwarz inequality and Gaussian integration by parts give $\Var F(\sqrt tG)\ge(\E[F(\sqrt tG)G])^2=t(\E F'(\sqrt tG))^2$, which is the lower bound of \cite{Cacoullos1982}. For $|z|\le1$ and $t\le1$, $|\E V'(z+\sqrt tG)|\ge\E[|V'(z+\sqrt tG)|;|G|\le1]\ge\varkappa_3$, because $V'<0$. The density of $Z'_q$ at $z\in[-1,1]$ is $\sqrt{\delta_{\mathrm d}/q}\,\phi((\kd+\sqrt{\delta_{\mathrm d}}z)/\sqrt q)\ge\beta\phi(1)$, because $q\le\qd$ and $|\kd|+\sqrt{\delta_{\mathrm d}}\le\sqrt{\qd-\delta_{\mathrm d}}\le\sqrt q$. Hence $1-S_{\mathrm d}(q)\ge\alpha \beta\varkappa_5t$ for $t\in(0,1]$. Since $\lambda_{\mathrm d}=\delta_{\mathrm d}$ on $[0,\qd]$, in the variable $t'=\hat s(r')-1$, \[ G_{\mathrm d}(q)=\frac12\int_0^t(t-t')\bigl(1-S_{\mathrm d}\bigr)\dd t'\ge\frac{\alpha \beta\varkappa_5}{12}t^3 . \] \emph{Saturation.} By (a), $G_{\mathrm d}$ decreases in $\hat s$ on $(0,1]$ and increases on $[1,\infty)$. For $\hat s<\frac14$ and $\hat s>2$ we may therefore use the values at $\frac14$ and at $2$. With $\varkappa_{\mathrm o}=\min\{\varkappa_4,\varkappa_5\}$ this gives $G_{\mathrm d}\ge\frac1{12}\alpha \beta\varkappa_{\mathrm o}\min\{|\hat s-1|,\frac34\}^3$, and $\alpha \beta\ge2c_\beta e$. (c) \emph{Above $\qd$.} Let $\rho=\hat s(q)\in[\frac14,1)$ and $\varsigma=\log(1/\rho)$. By \eqref{II:flat:eq-AV}, the generator of \eqref{II:flat:eq-topdiff} maps $V^2$ to $k_1=V'^2-2V^2(1-V)$, which is bounded and integrable. Dynkin's formula gives $S_{\mathrm d}(\qd)-S_{\mathrm d}(q)=-\alpha\int_0^{\varsigma}\E k_1(Z_{\e^{-r}})\dd r$. The drift of \eqref{II:flat:eq-topdiff} has derivative $\frac12-V\ge-\frac12$. Hence two solutions driven by the same Brownian motion and started at $z_0t$ and $1-V(z)\ge\frac13\Rbar(z)^2\ge\frac13(t^{-1}-3t^{-3})^2\ge\frac1{48}t^{-2}$, so $R^2(1-V)\ge\frac1{48}$ on $(-\infty,-2]$; and $\Prob(Z^q\le-2)=\Phi((\kappa-2\sqrt{\delta_q})/\sqrt q)\ge\frac14$ for small $e$. Hence \begin{equation}\label{II:appD:eq-rsbounds} \begin{gathered} |\partial_qS_{\RS}|\le\frac Ce,\qquad |\partial_\kappa S_{\RS}|\le C,\qquad |\partial_\kappa D_{\RS}|\le\frac C{e^4},\\ ce\le A_w(q)\le Ce,\qquad \E\bigl[R^2(1-V)\bigr](Z^q)\ge\frac1{192} . \end{gathered} \end{equation} This proves (c) and the first sentence of (b). (a) At $(\qd,\kd)$ we have $S_{\RS}=1$ and $D_{\RS}=0$. By \eqref{II:appD:eq-rsbounds}, $|S_{\RS}-1|\le C\eta e$ on $\cW_\eta\times[\kd-\eta e^2,\kd+\eta e^2]$, and $|D_{\RS}(\qd;\kappa)|\le C\eta/e^2$. The identity \eqref{II:appD:eq-DRS} is a linear differential equation for $D_{\RS}$ in $q$, with coefficient $2/\delta_q\le8/(d_0e^2)$ and with the other terms bounded by $C/e^4$. Gronwall's inequality over a $q$-interval of length at most $\eta e^2$ gives $|D_{\RS}|\le C\eta/e^2$ on $\cW_\eta$. Inserting these bounds and $\alpha\ge2$ into \eqref{II:appD:eq-DRS}, and using $e^{-2}\le3d_0/\delta_q$, \[ \partial_qD_{\RS}\le\frac{C\eta e-\frac1{48}+C\eta}{\delta_q^2}\le-\frac1{96\,\delta_q^2} \] for $\eta=\eta_1$ small. Now let $|\kappa-\kd|\le\eta_1'e^2$. By \eqref{II:appD:eq-rsbounds}, $|D_{\RS}(\qd;\kappa)|\le C|\kappa-\kd|/e^4$, while $\partial_qD_{\RS}\le-1/(864d_0^2e^4)$ on $\cW_{\eta_1}$, because $\delta_q\le3d_0e^2$. Hence $D_{\RS}(\cdot;\kappa)$ changes sign within distance $C_1|\kappa-\kd|$ of $\qd$, with $C_1=864d_0^2C$, and has a zero there if $C_1\eta_1'<\eta_1$. This zero is RS-stationary at $\kappa$. Since $\kRS-\kd=\thd e^2+o(e^2)$ by \Cref{II:prop:crossing}(iii), after decreasing $\eta_1'$ (so that $\eta_1'<\frac12\thd$) and $e_{\mathrm r}$ we have $\kappa\in[\kRS-c_2,\kRS)$, and by the uniqueness in \Cref{II:prop:crossing}(i) the zero is $q^*(\kappa)$. This proves (a) with $c_{\mathrm r}\le\frac1{96}$ and $C_{\mathrm r}\ge C_1$. (b) The segments from $(\qd,\kd)$ to $(\qd,\kappa)$ and from $(\qd,\kappa)$ to $(q^*(\kappa),\kappa)$ lie in $\cW_{\eta_1}\times[\kd-\eta_1e^2,\kd+\eta_1e^2]$. By \eqref{II:appD:eq-rsbounds} and (a), \[ |\tau(\kappa)|=\bigl|S_{\RS}(q^*(\kappa);\kappa)-S_{\RS}(\qd;\kd)\bigr|\le\frac Ce\,C_1|\kappa-\kd|+C|\kappa-\kd|\le\frac{C_{\mathrm r}|\kappa-\kd|}e \] after enlarging $C_{\mathrm r}$. \end{proof} \section{Natural-scale continuity}\label{II:appD:ss-nsc} \begin{proof}[Proof of \Cref{II:frsb:lem-natcont}] We divide the proof into four steps. The first two hold for a fixed interval $[u_-,u_+]$ with $00$ it stays above $y/2>0$ with probability at least $1-\e^{-y^2/C}$; and bounded $|y|$ is absorbed in the constant. By \Cref{II:flat:lem-tails}(b),(c), the functions $v_\zeta$, $\partial_yv_\zeta$, and $\partial_y^2v_\zeta$ are bounded, and \begin{align*} |1-v_\zeta|+|\partial_yv_\zeta|+|\partial_y^2v_\zeta|&\le C(1+z^2)^{-1}\quad(z<0),\\ v_\zeta+|\partial_yv_\zeta|+|\partial_y^2v_\zeta|&\le C\e^{-z^2/C}\quad(z\ge0), \end{align*} at $(s,z)$. Since $P$ is Lipschitz on bounded sets and $P(0,0,0)=0$, $|r_{P,\zeta}(s,z)|$ obeys the same two bounds. With \Cref{II:appC:lem-transport}(a),(b) once more, $\E|r_{P,\zeta}(s,Y^\zeta_s(y))|\le C(1+y^2)^{-1}$. \proofstep{3}{Convergence.} Let $e_n\to0$ and $\kappa_n\to0$, let $\zeta_n\in\cK(s_{\rm lo},s_{\rm hi})$ converge in $L^1$ to $\zeta^0=\1_{[0,d_0]}$, let $s_n\to s_\infty\in(s_{\rm lo},s_{\rm hi})$, and let $\pi_n=\pi_{e_n,\kappa_n}$. We claim that \begin{equation}\label{II:appD:eq-lim} \begin{aligned} \int\pi_n\,\E\,r_{P,\zeta_n}\bigl(s_n,Y^{\zeta_n}_{s_n}(y)\bigr)\dd y&\longrightarrow\phi_0\int\E\,r_{P,\zeta^0}\bigl(s_\infty,Y^{\zeta^0}_{s_\infty}(y)\bigr)\dd y,\\ \int\pi_n\bigl[\Prob\bigl(Y^{\zeta_n}_{s_n}(y)<0\bigr)-\1_{\{y<0\}}\bigr]\dd y&\longrightarrow\phi_0\int\bigl[\Prob\bigl(Y^{\zeta^0}_{s_\infty}(y)<0\bigr)-\1_{\{y<0\}}\bigr]\dd y . \end{aligned} \end{equation} By \Cref{II:appC:lem-dilation}, the dilated order parameters of $\zeta_n$ converge in $L^1$ to that of $\zeta^0$. Hence, by \Cref{II:lem:regularity}(f), $\partial_y^kU_{\zeta_n}\to\partial_y^kU_{\zeta^0}$ for $k\le4$, uniformly on $[s_{\rm lo},s_{\rm hi}]\times[-M,M]$ for every $M$; by \Cref{II:lem:regularity}(a) the limits are continuous; and $\ell_{\zeta_n}\to\ell_{\zeta^0}$ uniformly because $|\ell_{\zeta_n}-\ell_{\zeta^0}|\le\|\zeta_n-\zeta^0\|_{L^1}$. Since $\partial_y^jv_\zeta=-\ell_\zeta\partial_y^{j+2}U_\zeta$, it follows that $P_{\zeta_n}(s_n,z_n)\to P_{\zeta^0}(s_\infty,z)$ whenever $z_n\to z$. Fix a Brownian path. The difference $Y^{\zeta_n}-Y^{\zeta^0}$ of two solutions of \eqref{II:appC:eq-Y} started from $y$ solves an ordinary differential equation in the time $\varsigma=s_{\rm hi}-s$, whose right side is $\zeta_n U_{\zeta_n,y}(Y^{\zeta_n})-\zeta^0U_{\zeta^0,y}(Y^{\zeta^0})$. The function $U_{\zeta_n,y}$ is Lipschitz in $y$ with constant $1/s_{\rm lo}$, because $|U_{\zeta,yy}|=v_\zeta/\ell_\zeta\le1/s_{\rm lo}$ on $[s_{\rm lo},s_{\rm hi}]$. The remaining terms $\zeta_n|U_{\zeta_n,y}-U_{\zeta^0,y}|(Y^{\zeta^0})$ and $|\zeta_n-\zeta^0|\,U_{\zeta^0,y}(Y^{\zeta^0})$ have integrals over $\varsigma\in[0,s_{\rm hi}-s_{\rm lo}]$ that tend to zero, because the path of $Y^{\zeta^0}$ is bounded and $\|\zeta_n-\zeta^0\|_{L^1}\to0$. Gronwall's inequality gives $\sup_s|Y^{\zeta_n}_s(y)-Y^{\zeta^0}_s(y)|\to0$, and continuity of the path gives $Y^{\zeta_n}_{s_n}(y)\to Y^{\zeta^0}_{s_\infty}(y)$. The law of $Y^{\zeta^0}_{s_\infty}(y)$ has a density, because $s_\infty0$. Then there are $e_n\to0$, order parameters $\gamma_n$ with $\hat s(\supp\mu_{\gamma_n})\subset[u_-,u_+]$ and $\|\gamma_n-\gamma_{\mathrm d}\|_{L^1}/\delta_{\mathrm d}\to0$, and levels $q_n$ with $\hat s(q_n)\in[u_-,u_+]$ and $|S_{\gamma_n,\kd}(q_n)-S_{\mathrm d}(q_n)|\ge\omega_0e_n$, where $\kd=\kd(2+e_n)$. Then $s_n=(1-q_n)/e_n^2=d_{\mathrm d}\hat s(q_n)\in[\frac12u_-d_0,2u_+d_0]$; pass to a subsequence with $s_n\to s_\infty\in(s_{\rm lo},s_{\rm hi})$. The change of variables $q=1-e_n^2s$ gives $\|\zeta_{\gamma_n}-\zeta_{\mathrm d}\|_{L^1}=e_n^{-2}\|\gamma_n-\gamma_{\mathrm d}\|_{L^1}\to0$, and $\zeta_{\mathrm d}=\1_{[0,d_{\mathrm d}]}\to\zeta^0$ because $d_{\mathrm d}\to d_0$. So both $\zeta_{\gamma_n}$ and $\zeta_{\mathrm d}$ converge to $\zeta^0$. Both profiles are evaluated at the margin $\kd\to0$, hence with the same $\pi_n$. Apply \eqref{II:appD:eq-resc} with $P(a_1,a_2,a_3)=a_1^2$, for which $\alpha e\cdot(\text{left side})=S$, to $\gamma_n$ and to $\gamma_{\mathrm d}$, writing $\Prob(Y<0)=\1_{\{y<0\}}+[\Prob(Y<0)-\1_{\{y<0\}}]$, and subtract. The terms $\int\pi_n\1_{\{y<0\}}$ cancel, and by \eqref{II:appD:eq-lim} for both sequences the remaining terms have the same limit. Hence $(S_{\gamma_n,\kd}(q_n)-S_{\mathrm d}(q_n))/(\alpha e_n)\to0$, a contradiction. (b) Let $\eta\le\frac14d_0$ and $e0$, $e_n^{-1}\E(e_nv_x)^2\to\phi_0\sqrt{d_0}\,A_V/d_0$, so $\widetilde x(q_n)\to x_{\mathrm f}$, and then $e_n^{-1}\widetilde\Sigma(q_n)\to\phi_0\sqrt{d_0}\,\Sigma_{\rm flat}$. Thus $\widetilde\Sigma(q_n)/\widetilde B(q_n)\to\sigma_{\mathrm f}$, a contradiction. The consequence \eqref{II:frsb:eq-winconst} follows because $x_{\mathrm f}\in(0,\frac14)$ and $\sigma_{\mathrm f}>0$ (\Cref{II:frsb:lem-flatconst}). \end{proof} \section{The bottom comparison}\label{II:appD:ss-bottom} We first record a comparison principle for the linear parabolic equations that arise, with weights adapted to the Gaussian tails. Its first step is the maximum principle for parabolic equations on unbounded domains, for functions of polynomial growth (see \cite{Friedman1964}). For $k\ge0$ and $c>0$ put \[ \mathfrak d_{k,c}(y)=(1+|y|)^{-k}\quad(y\le0),\qquad \mathfrak d_{k,c}(y)=\e^{-cy^2}\quad(y>0). \] Then $\mathfrak d_{k,c}\le\mathfrak d_{k',c'}$ if $k\ge k'$ and $c\ge c'$, and $\mathfrak d_{k,c}\mathfrak d_{k',c'}\le\mathfrak d_{k+k',\min\{c,c'\}}$. \begin{lemma}\label{II:appD:lem-parabolic} Let $00$, and $C_0,C_H\ge0$. Let $0=t_0p_0/2$ and put $\Psi(t,y)=\e^{Kt}(1+y^2)^p$. Then $|\partial_{yy}\Psi|\le C_p\Psi$, and $|\beta_1\partial_y\Psi|\le C_0(1+|y|)\,2p|y|(1+y^2)^{p-1}\e^{Kt}\le4pC_0\Psi$, because $(1+|y|)|y|\le2(1+y^2)$; also $-\beta_0\Psi\ge0$. Hence $(\partial_t-\mathsf L)\Psi\ge(K-\frac12C_p-4pC_0)\Psi\ge\Psi$ for $K=K(p,C_0)$ large. For $\varepsilon>0$ the function $w_\varepsilon=w-\varepsilon\Psi$ satisfies $(\partial_t-\mathsf L)w_\varepsilon\le-\varepsilon\Psi<0$ on each block, $w_\varepsilon(0,\cdot)<0$, and $w_\varepsilon\to-\infty$ as $|y|\to\infty$, uniformly in $t\in[0,T]$. Suppose that $\sup w_\varepsilon>0$. The supremum is then attained at some $(t^*,y^*)$ with $t^*>0$. Choose $j$ with $t^*\in(t_{j-1},t_j]$. At $(t^*,y^*)$ we have $\partial_yw_\varepsilon=0$ and $\partial_{yy}w_\varepsilon\le0$, and the left $t$-derivative on the block $[t_{j-1},t_j]$ is nonnegative. Since $w_\varepsilon(t^*,y^*)>0$ and $\beta_0\le0$, \[ (\partial_t-\mathsf L)w_\varepsilon(t^*,y^*)=\partial_tw_\varepsilon-\tfrac12\partial_{yy}w_\varepsilon-\beta_1\partial_yw_\varepsilon-\beta_0w_\varepsilon\ge0, \] a contradiction. Hence $w_\varepsilon\le0$, and letting $\varepsilon\to0$ gives $w\le0$. \proofstep{2}{Bounded sources.} Since $(\partial_t-\mathsf L)(C_Ht)=C_H-\beta_0C_Ht\ge C_H$, the functions $w=\pm f-C_Ht$ satisfy $(\partial_t-\mathsf L)w\le\pm H-C_H\le0$ and $w(0,\cdot)=0$. Step 1 gives (i). \proofstep{3}{Weighted sources.} Let $\chi\in C^\infty(\R)$ satisfy $0\le\chi\le1$, $\chi=1$ on $(-\infty,-1]$ and $\chi=0$ on $[1,\infty)$. Put $c(t)=c_H/(1+2c_Ht)$, so that $\dot c(t)=-2c(t)^2$ and $c(t)\ge\hat c$ for $t\le T$, and \[ \mathfrak w_0(t,y)=\chi(y)(1+y^2)^{-k/2}+\bigl(1-\chi(y)\bigr)\e^{-c(t)y^2}. \] For $y\ge1$, $\mathfrak w_0=\e^{-c(t)y^2}$ and \[ \bigl(\partial_t-\tfrac12\partial_{yy}\bigr)\mathfrak w_0=\bigl(-\dot c(t)y^2-2c(t)^2y^2+c(t)\bigr)\mathfrak w_0=c(t)\mathfrak w_0\ge0 ; \] moreover $-\beta_1\partial_y\mathfrak w_0=2c(t)y\beta_1\mathfrak w_0\ge0$ and $-\beta_0\mathfrak w_0\ge0$, so $(\partial_t-\mathsf L)\mathfrak w_0\ge0$. For $y\le-1$, $\mathfrak w_0=(1+y^2)^{-k/2}$ does not depend on $t$, $|\partial_y\mathfrak w_0|=k|y|(1+y^2)^{-1}\mathfrak w_0\le k|y|^{-1}\mathfrak w_0$, and $|\partial_{yy}\mathfrak w_0|\le C_k\mathfrak w_0$; since $\beta_1\le2C_0|y|$ there, $(\partial_t-\mathsf L)\mathfrak w_0\ge-(\frac12C_k+2kC_0)\mathfrak w_0$. On $[0,T]\times[-1,1]$, $\mathfrak w_0$ is smooth, bounded below by a positive constant, and its derivatives are bounded, with constants depending only on $k$ and $c_H$ (because $c(t)\in[\hat c,c_H]$); also $\beta_1\le2C_0$ there. Hence $(\partial_t-\mathsf L)\mathfrak w_0\ge-C_2\mathfrak w_0$ on $[0,T]\times\R$, with $C_2$ depending only on $k,c_H,C_0$. The barrier $\mathfrak w=\e^{(C_2+1)t}\mathfrak w_0$ therefore satisfies \[ (\partial_t-\mathsf L)\mathfrak w\ge\mathfrak w . \] Comparing the two pieces of $\mathfrak w_0$ with $\mathfrak d_{k,\cdot}$ on $y\le-1$, $y\ge1$, and $|y|\le1$, and using $(1+|y|)^2\le2(1+y^2)$, we find $C_3,C_4$, depending only on $k$, $c_H$, and $C_2$, such that \[ \mathfrak d_{k,c_H}\le C_3\mathfrak w,\qquad \mathfrak w\le C_4\mathfrak d_{k,\hat c}\qquad\text{on }[0,T]\times\R . \] Put $A=\frac12C_HC_3$. Then $(\partial_t-\mathsf L)(t^2\mathfrak w)=2t\mathfrak w+t^2(\partial_t-\mathsf L)\mathfrak w\ge2t\mathfrak w\ge2t\mathfrak d_{k,c_H}/C_3$, so $w=\pm f-At^2\mathfrak w$ satisfies $(\partial_t-\mathsf L)w\le C_Ht\mathfrak d_{k,c_H}-2At\mathfrak d_{k,c_H}/C_3=0$ and $w(0,\cdot)=0$. Step 1 gives $|f|\le At^2\mathfrak w\le AC_4t^2\mathfrak d_{k,\hat c}$. This proves (ii). \end{proof} We now expand the solution near the top of the support around the Gaussian reference with the current value of $\lambda$. Fix $Q\in[\frac12,1)$, put $\delta=1-Q$, and fix a margin $\kappa$. For $\gamma'\in\cU$ with $\gamma'=1$ on $[Q,1]$, $t\in[0,\frac18]$ and $y\in\R$ put \begin{align*} \bar\gamma'(t)&=\gamma'(Q-\delta t),\quad l(t)=\frac{\lambda_{\gamma'}(Q-\delta t)}\delta,\\ \hat b(t,y)&=\sqrt\delta\,b_{\gamma'}\bigl(Q-\delta t,\kappa+\sqrt\delta y\bigr),\quad \varepsilon(t)=t-l(t)+1 . \end{align*} Then $l(0)=1$ and $l'=\bar\gamma'\in[0,1]$, so $1\le l\le1+t\le\frac98$, $0\le\varepsilon\le t$, and $\varepsilon'=1-\bar\gamma'$. The Gaussian reference is \[ f(l,y)=l^{-1/2}R\bigl(y/\sqrt l\bigr)=\partial_y\log\Phi\bigl(y/\sqrt l\bigr),\qquad g=\tfrac12\partial_{yy}f . \] \begin{lemma}\label{II:appD:lem-vw} There are numerical constants $C,c>0$ such that, in this setting, for $t\in[0,\frac18]$ and $y\in\R$, \begin{equation}\label{II:appD:eq-cancellation} \begin{split} \hat b(t,y)&=f\bigl(l(t),y\bigr)+\varepsilon(t)\,g\bigl(l(t),y\bigr)+\mathcal R(t,y),\\ |\mathcal R(t,y)|&\le Ct^2\mathfrak d_{3,c}(y),\quad |\partial_y\mathcal R(t,y)|\le Ct^2\mathfrak d_{4,c}(y). \end{split} \end{equation} \end{lemma} \begin{proof} We first let $\gamma'$ be a step profile, and write $\bar\gamma=\bar\gamma'$. We divide the proof into four steps. \proofstep{1}{Equations and a priori bounds.} By \Cref{II:lem:generator}(a), $\cL b=0$, that is, $b_q=-\frac12b_{xx}-\gamma'bb_x$. Since $\partial_t=-\delta\partial_q$ and $\partial_y=\sqrt\delta\partial_x$, on each block of constancy of $\bar\gamma$ \[ \hat b_t=\tfrac12\hat b_{yy}+\bar\gamma\,\hat b\,\hat b_y,\qquad \hat b(0,\cdot)=f(1,\cdot), \] the initial condition by \eqref{II:fun:topderiv}. By \Cref{II:lem:regularity}(a),(b), $\hat b$ is of class $C^{1,\infty}$ on each closed block, and it and its $y$-derivatives have polynomial growth. By \Cref{II:lem:regularity}(b),(c), $\hat b>0$ and $\hat b_y=-\delta c=-v/l\in[-1,0]$. By \Cref{II:lem:sandwich}(b), $\lambda b(q,\kappa)\le\sqrt{1-q}$, so $\hat b(t,0)\le\sqrt{1+t}/l\le2$. Hence \[ 0<\hat b(t,y)\le2+y_- . \] Since $\Phi(y/\sqrt l)$ solves $\partial_l=\frac12\partial_{yy}$, the function $\log\Phi(y/\sqrt l)$ solves $\partial_l=\frac12(\partial_{yy}+(\partial_y\,\cdot)^2)$, and differentiation in $y$ gives \[ f_l=\tfrac12f_{yy}+ff_y,\qquad g_l=\tfrac12g_{yy}+fg_y+3f_yg, \] the second by differentiating the first twice and using $(ff_y)_{yy}=ff_{yyy}+3f_yf_{yy}$. Moreover $f_y=-l^{-1}V(y/\sqrt l)\in[-1,0]$ and $\partial_y^jf=-l^{-(j+1)/2}V^{(j-1)}(y/\sqrt l)$ for $j\ge1$. By \Cref{II:appC:lem-gauss}(a), $|V^{(i)}(z)|\le C_i(1+|z|)^{-2-i}$ for $z\le0$ and $|V^{(i)}(z)|\le C_i\e^{-z^2/4}$ for $z\ge0$, for $i\ge1$. Since $1\le l\le\frac98$, with $c_*=\frac29$, \[ |\partial_y^jf(l,y)|\le C\mathfrak d_{j+1,c_*}(y)\qquad(2\le j\le5). \] In particular $|g|\le C\mathfrak d_{3,c_*}$, $|g_y|\le C\mathfrak d_{4,c_*}$, $|g_{yy}|\le C\mathfrak d_{5,c_*}$ and $|g_{yyy}|\le C\mathfrak d_{6,c_*}$. \proofstep{2}{First order.} Put $\Delta=\hat b-f(l(t),\cdot)$. Using $\frac{\dd}{\dd t}f(l(t),y)=\bar\gamma f_l$ and $\hat b\hat b_y-ff_y=\hat b\Delta_y+f_y\Delta$, on each block \[ \Delta_t=\mathsf L\Delta+(1-\bar\gamma)g,\qquad \mathsf L=\tfrac12\partial_{yy}+\bar\gamma\hat b\,\partial_y+\bar\gamma f_y,\qquad \Delta(0,\cdot)=0 . \] The drift $\beta_1=\bar\gamma\hat b$ satisfies $0\le\beta_1\le2+y_-$, and the reaction $\beta_0=\bar\gamma f_y$ is nonpositive. The source is bounded, so \Cref{II:appD:lem-parabolic}(i) gives $|\Delta|\le Ct$. Differentiating in $y$, \[ \partial_t\Delta_y=\mathsf L'\Delta_y+\bar\gamma f_{yy}\Delta+(1-\bar\gamma)g_y,\qquad \mathsf L'=\tfrac12\partial_{yy}+\bar\gamma\hat b\,\partial_y+\bar\gamma(\hat b_y+f_y), \] with the same drift, the nonpositive reaction $\bar\gamma(\hat b_y+f_y)$, and a bounded source. Hence also $|\Delta_y|\le Ct$. \proofstep{3}{Second order.} Put $\mathcal R=\Delta-\varepsilon g(l(t),\cdot)$. Since $\varepsilon'=1-\bar\gamma$ and $\frac{\dd}{\dd t}g(l(t),y)=\bar\gamma g_l$, the equations of Steps 1 and 2 give exactly \[ \mathcal R_t=\mathsf L\mathcal R+h,\qquad h=\varepsilon\Bigl[\tfrac12(1-\bar\gamma)g_{yy}+\bar\gamma\Delta g_y-2\bar\gamma f_yg\Bigr],\qquad \mathcal R(0,\cdot)=0 . \] By Steps 1 and 2 and $0\le\varepsilon\le t$, \[ |h|\le Ct\mathfrak d_{3,c_*},\qquad |h_y|=\Bigl|\varepsilon\Bigl[\tfrac12(1-\bar\gamma)g_{yyy}+\bar\gamma\Delta_yg_y+\bar\gamma\Delta g_{yy}-2\bar\gamma f_{yy}g-2\bar\gamma f_yg_y\Bigr]\Bigr|\le Ct\mathfrak d_{4,c_*}. \] \Cref{II:appD:lem-parabolic}(ii) with $k=3$ gives $|\mathcal R|\le Ct^2\mathfrak d_{3,c_1}$ with $c_1=c_*/(1+2c_*)$. \proofstep{4}{The derivative, and general profiles.} Differentiating the equation for $\mathcal R$ in $y$, \[ \partial_t\mathcal R_y=\mathsf L'\mathcal R_y+\bar\gamma f_{yy}\mathcal R+h_y . \] Here $|\bar\gamma f_{yy}\mathcal R|\le C\mathfrak d_{3,c_*}\,t^2\mathfrak d_{3,c_1}\le Ct\mathfrak d_{4,c_1}$ and $|h_y|\le Ct\mathfrak d_{4,c_1}$, so \Cref{II:appD:lem-parabolic}(ii) with $k=4$ gives $|\mathcal R_y|\le Ct^2\mathfrak d_{4,c}$ with $c=c_1/(1+2c_1)$. All constants are numerical. This proves \eqref{II:appD:eq-cancellation} for step profiles. For a general $\gamma'$, choose step profiles $\gamma'_n\in\cU_Q$ with $\gamma'_n\to\gamma'$ in $L^1$. For instance, for partitions $0=r_0<\dots