From cca6ae689762dd447a464be37a2b5229248235e1 Mon Sep 17 00:00:00 2001 From: Jaron Kent-Dobias Date: Tue, 28 Nov 2023 17:00:33 +0100 Subject: Big text reshuffle as more is put in the appendix. --- 2-point.tex | 1284 ++++++++++++++++++++++++++++++----------------------------- 1 file changed, 643 insertions(+), 641 deletions(-) (limited to '2-point.tex') diff --git a/2-point.tex b/2-point.tex index 9be13e9..ce2d1bc 100644 --- a/2-point.tex +++ b/2-point.tex @@ -612,718 +612,720 @@ $\pmb\sigma_1$ is special among the set of $\pmb\sigma$ replicas, since it alone is constrained to lie a given overlap from the $\mathbf s$ replicas. This replica asymmetry will be important later. -\subsection{The Hessian factors} - -The double partial derivatives of the energy are Gaussian with the variance -\begin{equation} - \overline{(\partial_i\partial_jH(\mathbf s))^2}=\frac1Nf''(1) -\end{equation} -which means that the matrix of partial derivatives belongs to the GOE class. Its spectrum is given by the Wigner semicircle +The resulting expression for the complexity, which must +still be extremized over the parameters $\hat\beta_1$, $r^{01}$, +$r^{11}_\mathrm d$, $r^{11}_0$, and $q^{11}_0$, is \begin{equation} - \rho(\lambda)=\begin{cases} - \frac2{\pi}\sqrt{1-\big(\frac{\lambda}{\mu_\text m}\big)^2} & \lambda^2\leq\mu_\text m^2 \\ - 0 & \text{otherwise} - \end{cases} + \begin{aligned} + &\Sigma_{12}(E_0,\mu_0,E_1,\mu_1,q) + =\mathop{\mathrm{extremum}}_{\hat\beta_1,r^{11}_\mathrm d,r^{11}_0,r^{01},q^{11}_0}\Bigg\{ + \mathcal D(\mu_1)-\frac12+\hat\beta_1E_1-r^{11}_\mathrm d\mu_1 + +\hat\beta_1\big(r^{11}_\mathrm df'(1)-r^{11}_0f'(q^{11}_0)\big)\\ + &\qquad+\hat\beta_0\hat\beta_1f(q)+(\hat\beta_0r^{01}+\hat\beta_1r^{10}+r^{00}_\mathrm d r^{01})f'(q) + +\frac{r^{11}_\mathrm d-r^{11}_0}{1-q^{11}_0}(r^{10}-qr^{00}_\mathrm d)f'(q)\\ + &+\frac12\Bigg[ + \hat\beta_1^2\big(f(1)-f(q^{11}_{0})\big) + +(r^{11}_\mathrm d)^2f''(1)+2r^{01}r^{10}f''(q)-(r^{11}_0)^2f''(q^{11}_0) + +\frac{(r^{10}-qr^{00}_\mathrm d)^2}{1-q^{11}_0}f'(1) + \\ + &\qquad+\frac{1-q^2}{1-q^{11}_0}+\left( + (r^{01})^2-\frac{r^{11}_\mathrm d-r^{11}_0}{1-q^{11}_0} + \left(2qr^{01}-\frac{(1-q^2)r^{11}_0-(q^{11}_0-q^2)r^{11}_\mathrm d}{1-q^{11}_0}\right) + \right)\big(f'(1)-f'(q_{22}^{(0)})\big) \\ + &\qquad + -\frac1{f'(1)}\frac{f'(1)^2-f'(q)^2}{f'(1)-f'(q^{11}_0)} + +\frac{r^{11}_\mathrm d-r^{11}_0}{1-q^{11}_0}\big(r^{11}_\mathrm df'(1)-r^{11}_0f'(q^{11}_0)\big) + +\log\left(\frac{1-q_{11}^0}{f'(1)-f'(q_{11}^0)}\right) + \Bigg]\Bigg\} + \end{aligned} \end{equation} -with radius $\mu_\text m=\sqrt{4f''(1)}$. Since the Hessian differs from the -matrix of partial derivatives by adding the constant diagonal matrix $\omega -I$, it follows that the spectrum of the Hessian is a Wigner semicircle shifted -by $\omega$, or $\rho(\lambda+\omega)$. +It is possible to further extremize this expression over all the other +variables but $q_0^{11}$, for which the saddle point conditions have a unique +solution. However, the resulting expression is quite complicated and provides +no insight. In practice, the complexity can be calculated in two ways. First, +the extremal problem can be done numerically, initializing from $q=0$ where the +problem reduces to that of the single-point complexity of points with energy +$E_1$ and stability $\mu_1$, and then taking small steps in $q$ or other +parameters to trace out the solution. This is how the data in all the plots of +this paper was produced. Second, the complexity can be calculated in the near +neighborhood of a reference point by expanding in small $1-q$. This is what we +describe in the next subsection. -The average over factors depending on the Hessian alone can be made separately -from those depending on the gradient or energy, since for random Gaussian -fields the Hessian is independent of these \cite{Bray_2007_Statistics}. In -principle the fact that we have conditioned the Hessian to belong to stationary -points of certain energy, stability, and proximity to another stationary point -will modify its statistics, but these changes will only appear at subleading -order in $N$ \cite{Ros_2019_Complexity}. At leading order, the various expectations factorize, each yielding +If there is no overlap gap between the reference point and its nearest +neighbors, their complexity can be calculated by an expansion in $1-q$. First, +we'll use this method to describe the most common type of stationary point in +the close vicinity of a reference point. The most common neighbors of a +reference point are given by further maximizing the two-point complexity over +the energy $E_1$ and stability $\mu_1$ of the nearby points. This gives the +conditions +\begin{align} + \hat\beta_1=0 && + \mu_1=2r^{11}_\mathrm df''(1) +\end{align} +where the second is only true for $\mu_1^2\leq\mu_\mathrm m^2$, i.e., when the +nearby points are saddle points or marginal minima. When these conditions are +inserted into the complexity, an expansion is made in small $1-q$, and the +saddle point in the remaining parameters is taken, the result is \begin{equation} - \overline{\big|\det\operatorname{Hess}H(\mathbf s,\omega)\big|\,\delta\big(N\mu-\operatorname{Tr}\operatorname{Hess}H(\mathbf s,\omega)\big)} - =e^{N\int d\lambda\,\rho(\lambda+\mu)\log|\lambda|}\delta(N\mu-N\omega) + \Sigma_{12} + =\frac{f'''(1)}{8f''(1)^2}(\mu_\mathrm m^2-\mu_0^2)\left(\sqrt{2+\frac{2f''(1)\big(f''(1)-f'(1)\big)}{f'''(1)f'(1)}}-1\right)(1-q) + +O\big((1-q)^2\big) \end{equation} -Therefore, all of the Lagrange multipliers are fixed to the stabilities $\mu$. We define the function +independent of $E_0$. To describe the properties of these most common +neighbors, it is convenient to first make a definition. The population of +stationary points that are most common at each energy (the blue line in +Fig.~\ref{fig:complexities}) have the relation \begin{equation} - \begin{aligned} - \mathcal D(\mu) - &=\int d\lambda\,\rho(\lambda+\mu)\log|\lambda| \\ - &=\begin{cases} - \frac12+\log\left(\frac12\mu_\text m\right)+\frac{\mu^2}{\mu_\text m^2} - & \mu^2\leq\mu_\text m^2 \\ - \frac12+\log\left(\frac12\mu_\text m\right)+\frac{\mu^2}{\mu_\text m^2} - -\left|\frac{\mu}{\mu_\text m}\right|\sqrt{\big(\frac\mu{\mu_\text m}\big)^2-1} - -\log\left(\left|\frac{\mu}{\mu_\text m}\right|-\sqrt{\big(\frac\mu{\mu_\text m}\big)^2-1}\right) & \mu^2>\mu_\text m^2 - \end{cases} - \end{aligned} + E_\mathrm{dom}(\mu_0)=-\frac{f'(1)^2+f(1)\big(f''(1)-f'(1)\big)}{2f''(1)f'(1)}\mu_0 \end{equation} -and the full factor due to the Hessians is +between $E_0$ and $\mu_0$ for $\mu_0^2\leq\mu_\mathrm m^2$. Using this +definition, the energy and stability of the most common neighbors at small +$\Delta q$ are +\begin{align} \label{eq:expansion.E.1} + E_1&=E_0+\frac12\frac{v_f}{u_f}\big(E_0-E_\mathrm{dom}(\mu_0)\big)(1-q)^2+O\big((1-q)^3\big) \\ + \label{eq:expansion.mu.1} + \mu_1&=\mu_0-\frac{v_f}{u_f}\big(E_0-E_\mathrm{dom}(\mu_0)\big)(1-q)+O\big((1-q)^2\big) +\end{align} +The most common neighboring saddles to a reference saddle are much nearer to +the reference in energy ($\Delta q^2$) than in stability ($\Delta q$). In fact, +this scaling also holds for the entire range of neighbors to a reference +saddle, with the limits in energy scaling like $\Delta q^2$ and those of +stability scaling like $\Delta q$. + +Because both expressions are proportional to $E_0-E_\mathrm{dom}(\mu_0)$, +whether the energy and stability of nearby points increases or decreases from +that of the reference point depends only on whether the energy of the reference +point is above or below that of the most common population at the same +stability. In particular, since $E_\mathrm{dom}(\mu_\mathrm m)=E_\mathrm{th}$, +the threshold energy is also the pivot around which the points asymptotically +nearby marginal minima change their properties. + +To examine better the population of marginal points, it is necessary to look at +the next term in the series of the complexity with $\Delta q$, since the linear +coefficient becomes zero at the marginal line. This tells us something +intuitive: stable minima have an effective repulsion between points, and one +always finds a sufficiently small $\Delta q$ that no stationary points are +point any nearer. For the marginal minima, it is not clear that the same should be true. + +When $\mu=\mu_\mathrm m$, the linear term above vanishes. Under these conditions, the quadratic term in the expansion is \begin{equation} - e^{Nm\mathcal D(\mu_0)+Nn\mathcal D(\mu_1)}\left[\prod_a^m\delta(N\mu_0-N\varsigma_a)\right]\left[\prod_a^n\delta(N\mu_1-N\omega_a)\right] + \Sigma_{12} + =\frac12\frac{f'''(1)v_f}{f''(1)^{3/2}u_f} + \left(\sqrt{\frac{2\big[f'(1)(f'''(1)-f''(1))+f''(1)^2\big]}{f'(1)f'''(1)}}-1\right)\big(E_0-E_\textrm{th}\big)(1-q)^2+O\big((1-q)^3\big) \end{equation} +Note that this expression is only true for $\mu=\mu_\mathrm m$. Therefore, +among marginal minima, when $E_0$ is greater than the threshold one finds +neighbors at arbitrarily close distance. When $E_0$ is less than the threshold, +the complexity of nearby points is negative, and there is a desert where none +are found. -\subsection{The other factors} +\begin{figure} + \centering + \includegraphics{figs/expansion_energy.pdf} + \hspace{1em} + \includegraphics{figs/expansion_stability.pdf} -Having integrated over the Lagrange multipliers using the $\delta$-functions -resulting from the average of the Hessians, any $\delta$-functions in the -remaining integrand we Fourier transform into their integral representation -over auxiliary fields. The resulting integrand has the form + \caption{ + Demonstration of the convergence of the $(1-q)$-expansion for marginal + reference minima. Solid lines and shaded region show are the same as in + Fig.~\ref{fig:marginal.prop.above} for $E_0-E_\mathrm{th}\simeq0.00667$. + The dotted lines show the expansion of most common neighbors, while the + dashed lines in both plots show the expansion for the minimum and maximum + energies and stabilities found at given $q$. + } \label{fig:expansion} +\end{figure} + +The properties of the nearby states above the threshold can be +further quantified. The most common points are still given by +\eqref{eq:expansion.E.1} and \eqref{eq:expansion.mu.1}, but the range of +available points can also be computed, and one finds that the stability lies in +the range +$\mu_1=\mu_\mathrm m+\delta\mu_1(1-q)\pm\delta\mu_2(1-q)^{3/2}+O\big((1-q)^2\big)$ +where $\delta\mu_1$ is given by the coefficient in \eqref{eq:expansion.mu.1} +and \begin{equation} - e^{ - Nm\hat\beta_0E_0+Nn\hat\beta_1E_1 - -\sum_a^m\left[(\pmb\sigma_a\cdot\hat{\pmb\sigma}_a)\mu_0 - -\frac12\hat\mu_0(N-\pmb\sigma_a\cdot\pmb\sigma_a) - \right] - -\sum_a^n\left[(\mathbf s_a\cdot\hat{\mathbf s}_a)\mu_1 - -\frac12\hat\mu_1(N-\mathbf s_a\cdot\mathbf s_a) - -\frac12\hat\mu_{12}(Nq-\pmb\sigma_1\cdot\mathbf s_a) - \right] - +\int d\mathbf t\,\mathcal O(\mathbf t)H(\mathbf t) + \delta\mu_2=\frac{v_f}{f'(1)f''(1)^{3/4}}\sqrt{ + \frac{E_0-E_\mathrm{th}}2\frac{2f''(1)\big(f''(1)-f'(1)\big)+f'(1)f'''(1)}{u_f} } \end{equation} -where we have introduced the linear operator +Similarly, one finds that the energy lies in the range $E_1=E_0+\delta +E_1(1-q)^2\pm\delta E_2(1-q)^{5/2}+O\big((1-q)^3\big)$ for $\delta E_1$ given +by the coefficient in \eqref{eq:expansion.E.1} and \begin{equation} - \mathcal O(\mathbf t) - =\sum_a^m\delta(\mathbf t-\pmb\sigma_a)\left( - i\hat{\pmb\sigma}_a\cdot\partial_{\mathbf t}-\hat\beta_0 - \right) - + - \sum_a^n\delta(\mathbf t-\mathbf s_a)\left( - i\hat{\mathbf s}_a\cdot\partial_{\mathbf t}-\hat\beta_1 - \right) + \begin{aligned} + \delta E_2 + &=\frac{\sqrt{E_0-E_\mathrm{th}}}{4f'(1)f''(1)^{3/4}}\bigg( + \frac{ + v_f + }{3u_f} + \big[ + f'(1)(2f''(1)-(2-(2-\delta q_0)\delta q_0)f'''(1))-2f''(1)^2 + \big] + \\ + &\hspace{12pc}\times + \big[f'(1)\big(6f''(1)+(18-(6-\delta q_0)\delta q_0)f'''(1)\big)-6f''(1)^2 + \big] + \bigg)^\frac12 + \end{aligned} \end{equation} -Here the $\hat\beta$s are the fields auxiliary to the energy constraints, the -$\hat\mu$s are auxiliary to the spherical and overlap constraints, and the -$\hat{\pmb\sigma}$s and $\hat{\mathbf s}$s are auxiliary to the constraint that -the gradient be zero. -We have written the $H$-dependent terms in this strange form for the ease of taking the average over $H$: since it is Gaussian-correlated, it follows that +and $\delta q_0$ is the coefficient in the expansion $q_0=1-\delta q_0(1-q)+O((1-q)^2)$ and is given by the real root to the quintic equation \begin{equation} - \overline{e^{\int d\mathbf t\,\mathcal O(\mathbf t)H(\mathbf t)}} - =e^{\frac12\int d\mathbf t\,d\mathbf t'\,\mathcal O(\mathbf t)\mathcal O(\mathbf t')\overline{H(\mathbf t)H(\mathbf t')}} - =e^{N\frac12\int d\mathbf t\,d\mathbf t'\,\mathcal O(\mathbf t)\mathcal O(\mathbf t')f\big(\frac{\mathbf t\cdot\mathbf t'}N\big)} + 0=((16-(6-\delta q_0)\delta q_0)\delta q_0-12)f'(1)f'''(1)-2\delta q_0(f''(1)-f'(1))f''(1) \end{equation} -It remains only to apply the doubled operators to $f$ and then evaluate the simple integrals over the $\delta$ measures. We do not include these details, which are standard. - -\subsection{Hubbard--Stratonovich} +These predictions from the small $1-q$ expansion are compared with numeric +saddle points for the complexity of marginal minima in +Fig.~\ref{fig:expansion}, and the results agree well at small $1-q$. -Having expanded this expression, we are left with an argument in the exponential which is a function of scalar products between the fields $\mathbf s$, $\hat{\mathbf s}$, $\pmb\sigma$, and $\hat{\pmb\sigma}$. We will change integration coordinates from these fields to matrix fields given by their scalar products, defined as -\begin{equation} \label{eq:fields} - \begin{aligned} - C^{00}_{ab}=\frac1N\pmb\sigma_a\cdot\pmb\sigma_b && - R^{00}_{ab}=-i\frac1N{\pmb\sigma}_a\cdot\hat{\pmb\sigma}_b && - D^{00}_{ab}=\frac1N\hat{\pmb\sigma}_a\cdot\hat{\pmb\sigma}_b \\ - C^{01}_{ab}=\frac1N\pmb\sigma_a\cdot\mathbf s_b && - R^{01}_{ab}=-i\frac1N{\pmb\sigma}_a\cdot\hat{\mathbf s}_b && - R^{10}_{ab}=-i\frac1N\hat{\pmb\sigma}_a\cdot{\mathbf s}_b && - D^{01}_{ab}=\frac1N\hat{\pmb\sigma}_a\cdot\hat{\mathbf s}_b \\ - C^{11}_{ab}=\frac1N\mathbf s_a\cdot\mathbf s_b && - R^{11}_{ab}=-i\frac1N{\mathbf s}_a\cdot\hat{\mathbf s}_b && - D^{11}_{ab}=\frac1N\hat{\mathbf s}_a\cdot\hat{\mathbf s}_b - \end{aligned} -\end{equation} -We insert into the integral the product of $\delta$-functions enforcing these -definitions, integrated over the new matrix fields, which is equivalent to -multiplying by one. Once this is done, the many scalar products appearing -throughout can be replaced by the matrix fields, and the original vector fields -can be integrated over. Conjugate matrix field integrals created when the -$\delta$-functions are promoted to exponentials can be evaluated by saddle -point in the standard way, yielding an effective action depending on the above -matrix fields alone. -\subsection{Saddle point} +\section{Isolated eigenvalue} +\label{sec:eigenvalue} -We will always assume that the square matrices $C^{00}$, $R^{00}$, $D^{00}$, -$C^{11}$, $R^{11}$, and $D^{11}$ are hierarchical matrices, with each set of -three sharing the same hierarchical structure. In particular, we immediately -define $c_\mathrm d^{00}$, $r_\mathrm d^{00}$, $d_\mathrm d^{00}$, $c_\mathrm d^{11}$, $r_\mathrm d^{11}$, and -$d_\mathrm d^{11}$ as the value of the diagonal elements of these matrices, -respectively. Note that $c_\mathrm d^{00}=c_\mathrm d^{11}=1$ due to the spherical constraint. +The two-point complexity depends on the spectrum at both stationary points +through the determinant of their Hessians, but only on the bulk of the +distribution. As we saw, this bulk is unaffected by the conditions of energy +and proximity. However, these conditions give rise to small-rank perturbations +to the Hessian, which can lead a subextensive number of eigenvalues leaving the +bulk. We study the possibility of \emph{one} stray eigenvalue. -Defining the `block' fields $\mathcal Q_{00}=(\hat\beta_0, \hat\mu_0, C^{00}, -R^{00}, D^{00})$, $\mathcal Q_{11}=(\hat\beta_1, \hat\mu_1, C^{11}, R^{11}, -D^{11})$, and $\mathcal Q_{01}=(\hat\mu_{01},C^{01},R^{01},R^{10},D^{01})$ -the resulting complexity is +We use a technique recently developed to find the smallest eigenvalue of a +random matrix \cite{Ikeda_2023_Bose-Einstein-like}. One defines a quadratic +statistical mechanics model with configurations defined on the sphere, whose +interaction tensor is given by the matrix of interest. By construction, the +ground state is located in the direction of the eigenvector associated with the +smallest eigenvalue, and the ground state energy is proportional to that +eigenvalue. + +\begin{figure} + \centering + \begin{tikzpicture} + \def\R{4 } % sphere radius + \def\Rt{2 } % tangent plane radius + \def\angEl{15} % elevation angle + \def\angsa{-160} % azimuth of s_1 + \def\angq{40} % elevation of constraint circle + \filldraw[ball color=white] (0,0) circle (\R); + % \filldraw[fill=white] (0,0) circle (\R); + + \foreach \t in {0,\angq} { \DrawLatitudeCircle[\R]{\t} } + %\foreach \t in {\angsa} { \DrawLongitudeCircle[\R]{\t} } + + \pgfmathsetmacro\H{\R*cos(\angEl)} % distance to north pole + \coordinate (O) at (0,0); + \node[circle,draw,black,scale=0.3] at (0,0) {}; + \coordinate (N) at (0,\H); + \draw node[right=10,below] at (0,\H){$\pmb\sigma_1$}; + \draw[thick, ->](O)--(N); + + \NewLatitudePlane[planeP]{\R}{\angEl}{\angq}; + \path[planeP] (\angsa:\R) coordinate (P); + \path[planeP] (0:1.5*\R) coordinate (Q); + \path[planeP] (0:\R) coordinate (Q2); + \draw[left] node at (P){$\mathbf s_1$}; + + \NewLatitudePlane[equator]{\R}{\angEl}{00}; + \path[equator] (-30:\R) coordinate (Pprime); + \path[equator] (0:{1.5*cos(\angq)*\R}) coordinate (Qe); + \path[equator] (0:\R) coordinate (Qe2); + \draw node[right=5,below] at (Pprime){$\pmb\sigma_c$}; + + \NewLatitudePlane[sbplane]{\R}{\angEl}{\angq}; + \path[sbplane] (20:\R) coordinate (sb); + \draw node[right=3,above=1] at (sb){$\mathbf s_b$}; + + \TangentPlane[tplane]{\R}{\angEl}{\angq}{\angsa}; + \draw[tplane,fill=gray,fill opacity=0.3] circle (\Rt); + \draw[tplane,->,thick] (0,0) -> ({\Rt*cos(160)},{\Rt*sin(160)}) node[above=1.5,right] {$\mathbf x_a$}; + \draw[tplane,->,thick] (0,0) -> ({\Rt*cos(250)},{\Rt*sin(250)}) node[above=1.5,left=0.1] {$\mathbf x_b$}; + + \draw[thick, ->] (O)->(P); + \draw[thick, ->] (O)->(Pprime); + \draw[thick, ->] (O)->(sb); + + \draw[dotted] (Qe) -- (Qe2); + \draw[dotted] (Q2) -- (Q); + \draw[decorate, decoration = {brace,raise=3}] (Q) -- (Qe) node[pos=0.5,right=7]{$q$}; + \end{tikzpicture} + \caption{ + A sketch of the vectors involved in the calculation of the isolated + eigenvalue. All replicas $\mathbf x$, which correspond with candidate + eigenvectors of the Hessian evaluated at $\mathbf s_1$, sit in an $N-2$ + sphere corresponding with the tangent plane (not to scale) of the first + $\mathbf s$ replica. All of the $\mathbf s$ replicas lie on the sphere, + constrained to be at fixed overlap $q$ with the first of the $\pmb\sigma$ + replicas, the reference configuration. All of the $\pmb\sigma$ replicas lie + on the sphere. + } \label{fig:sphere} +\end{figure} + +Our matrix of interest is the Hessian evaluated at a stationary point of the mixed spherical +model, conditioned on the relative position, energies, and stabilities +discussed above. We must restrict the artificial spherical model to lie in the +tangent plane of the `real' spherical configuration space at the point of +interest, to avoid our eigenvector pointing in a direction that violates the +spherical constraint. A sketch of the setup is shown in Fig.~\ref{fig:sphere}. The free energy of this model given a point $\mathbf s$ +and a specific realization of the disordered Hamiltonian is \begin{equation} - \Sigma_{01} - =\frac1N\lim_{n\to0}\lim_{m\to0}\frac\partial{\partial n}\int d\mathcal Q_{00}\,d\mathcal Q_{11}\,d\mathcal Q_{01}\,e^{Nm\mathcal S_0(\mathcal Q_{00})+Nn\mathcal S_1(\mathcal Q_{11},\mathcal Q_{01}\mid\mathcal Q_{00})} + \begin{aligned} + \beta F_H(\beta\mid\mathbf s,\omega) + &=-\frac1N\log\left(\int d\mathbf x\,\delta(\mathbf x\cdot\mathbf s)\delta(\|\mathbf x\|^2-N)\exp\left\{ + -\beta\frac12\mathbf x^T\operatorname{Hess}H(\mathbf s,\omega)\mathbf x + \right\}\right) \\ + &=-\lim_{\ell\to0}\frac1N\frac\partial{\partial\ell}\int\left[\prod_{\alpha=1}^\ell d\mathbf x_\alpha\,\delta(\mathbf x_\alpha^T\mathbf s)\delta(N-\mathbf x_\alpha^T\mathbf x_\alpha)\exp\left\{ + -\beta\frac12\mathbf x^T_\alpha\big(\partial\partial H(\mathbf s)+\omega I\big)\mathbf x_\alpha + \right\}\right] + \end{aligned} \end{equation} -where -\begin{equation} \label{eq:one-point.action} +where the first $\delta$-function keeps the configurations in the tangent +plane, and the second enforces the spherical constraint. We have anticipated +treating the logarithm with replicas. We are interested in points $\mathbf s$ +that have certain properties: they are stationary points of $H$ with given +energy density and stability, and fixed overlap from a reference configuration +$\pmb\sigma$. We therefore average the free energy above over such points, +giving +\begin{equation} \begin{aligned} - &\mathcal S_0(\mathcal Q_{00}) - =\hat\beta_0E_0-r^{00}_\mathrm d\mu_0-\frac12\hat\mu_0(1-c^{00}_\mathrm d)+\mathcal D(\mu_0)\\ - &\quad+\frac1m\bigg\{ - \frac12\sum_{ab}^m\left[ - \hat\beta_1^2f(C^{00}_{ab})+(2\hat\beta_1R^{00}_{ab}-D^{00}_{ab})f'(C^{00}_{ab})+(R_{ab}^{00})^2f''(C_{ab}^{00}) - \right]+\frac12\log\det\begin{bmatrix}C^{00}&R^{00}\\R^{00}&D^{00}\end{bmatrix} -\bigg\} -\end{aligned} + F_H(\beta\mid E_1,\mu_1,q,\pmb\sigma) + &=\int\frac{d\nu_H(\mathbf s,\omega\mid E_1,\mu_1)\delta(Nq-\pmb\sigma\cdot\mathbf s)}{\int d\nu_H(\mathbf s',\omega'\mid E_1,\mu_1)\delta(Nq-\pmb\sigma\cdot\mathbf s')}F_H(\beta\mid\mathbf s,\omega) \\ + &=\lim_{n\to0}\int\left[\prod_{a=1}^nd\nu_H(\mathbf s_a,\omega_a\mid E_1,\mu_1)\,\delta(Nq-\pmb\sigma\cdot\mathbf s_a)\right]F_H(\beta\mid\mathbf s_1,\omega_1) + \end{aligned} \end{equation} -is the action for the ordinary, one-point complexity, and remainder is given by -\begin{equation} \label{eq:two-point.action} +again anticipating the use of replicas. Finally, the reference configuration $\pmb\sigma$ should itself be a stationary point of $H$ with its own energy density and stability. Averaging over these conditions gives +\begin{equation} \begin{aligned} - &\mathcal S(\mathcal Q_{11},\mathcal Q_{01}\mid\mathcal Q_{00}) - =\hat\beta_1E_1-r^{11}_\mathrm d\mu_1-\frac12\hat\mu_1(1-c^{11}_\mathrm d)+\mathcal D(\mu_1) \\ - &\quad+\frac1n\sum_b^n\left\{-\frac12\hat\mu_{12}(q-C^{01}_{1b})+\sum_a^m\left[ - \hat\beta_0\hat\beta_1f(C^{01}_{ab})+(\hat\beta_0R^{01}_{ab}+\hat\beta_1R^{10}_{ab}-D^{01}_{ab})f'(C^{01}_{ab})+R^{01}_{ab}R^{10}_{ab}f''(C^{01}_{ab}) - \right]\right\} - \\ - &\quad+\frac1n\bigg\{ - \frac12\sum_{ab}^n\left[ - \hat\beta_1^2f(C^{11}_{ab})+(2\hat\beta_1R^{11}_{ab}-D^{11}_{ab})f'(C^{11}_{ab})+(R^{11}_{ab})^2f''(C^{11}_{ab}) - \right]\\ - &\quad+\frac12\log\det\left( - \begin{bmatrix} - C^{11}&iR^{11}\\iR^{11}&D^{11} - \end{bmatrix}- - \begin{bmatrix} - C^{01}&iR^{01}\\iR^{10}&D^{01} - \end{bmatrix}^T - \begin{bmatrix} - C^{00}&iR^{00}\\iR^{00}&D^{00} - \end{bmatrix}^{-1} - \begin{bmatrix} - C^{01}&iR^{01}\\iR^{10}&D^{01} - \end{bmatrix} - \right) - \bigg\} + F_H(\beta\mid E_1,\mu_1,E_2,\mu_2,q) + &=\int\frac{d\nu_H(\pmb\sigma,\varsigma\mid E_0,\mu_0)}{\int d\nu_H(\pmb\sigma',\varsigma'\mid E_0,\mu_0)}\,F_H(\beta\mid E_1,\mu_1,q,\pmb\sigma) \\ + &=\lim_{m\to0}\int\left[\prod_{a=1}^m d\nu_H(\pmb\sigma_a,\varsigma_a\mid E_0,\mu_0)\right]\,F_H(\beta\mid E_1,\mu_1,q,\pmb\sigma_1) \end{aligned} \end{equation} -Because of the structure of this problem in the twin limits of $m$ and $n$ to -zero, the parameters $\mathcal Q_{00}$ can be evaluated at a saddle point of -$\mathcal S_0$ alone. This means that these parameters will take the same value -they take when the ordinary, 1-point complexity is calculated. For a replica -symmetric complexity of the reference point, this results in -\begin{align} - \hat\beta_0 - &=-\frac{\mu_0f'(1)+E_0\big(f'(1)+f''(1)\big)}{u_f}\\ - r_\mathrm d^{00} - &=\frac{\mu_0f(1)+E_0f'(1)}{u_f} \\ - d_\mathrm d^{00} - &=\frac1{f'(1)} - -\left( - \frac{\mu_0f(1)+E_0f'(1)}{u_f} - \right)^2 -\end{align} -where we define for brevity (here and elsewhere) the constants -\begin{align} - u_f=f(1)\big(f'(1)+f''(1)\big)-f'(1)^2 - && - v_f=f'(1)\big(f''(1)+f'''(1)\big)-f''(1)^2 -\end{align} -Note that because the coefficients of $f$ must be nonnegative for $f$ to -be a sensible covariance, both $u_f$ and $v_f$ are strictly positive. Note also -that $u_f=v_f=0$ if $f$ is a homogeneous polynomial as in the pure models. -These expressions are invalid for the pure models because $\mu_0$ and $E_0$ -cannot be fixed independently; we would have done the equivalent of inserting -two identical $\delta$-functions. For the pure models, the terms $\hat\beta_0$ and -$\hat\beta_1$ must be set to zero in our prior formulae (as if the energy was -not constrained) and then the saddle point taken. - +This formidable expression is now ready to be averaged over the disordered Hamiltonians $H$. Once averaged, +the minimum eigenvalue of the conditioned Hessian is then given by twice the ground state energy, or +\begin{equation} + \lambda_\text{min}=2\lim_{\beta\to\infty}\overline{F_H(\beta\mid E_1,\mu_1,E_2,\mu_2,q)} +\end{equation} +For this calculation, there are three different sets of replicated variables. +Note that, as for the computation of the complexity, the $\pmb\sigma_1$ and +$\mathbf s_1$ replicas are \emph{special}. The first again is the only of the +$\pmb\sigma$ replicas constrained to lie at fixed overlap with \emph{all} the +$\mathbf s$ replicas, and the second is the only of the $\mathbf s$ replicas at +which the Hessian is evaluated. -In general, we except the $m\times n$ matrices $C^{01}$, $R^{01}$, $R^{10}$, -and $D^{01}$ to have constant \emph{rows} of length $n$, with blocks of rows -corresponding to the \textsc{rsb} structure of the single-point complexity. For -the scope of this paper, where we restrict ourselves to replica symmetric -complexities, they have the following form at the saddle point: -\begin{align} \label{eq:01.ansatz} - C^{01}= - \begin{subarray}{l} - \hphantom{[}\begin{array}{ccc}\leftarrow&n&\rightarrow\end{array}\hphantom{\Bigg]}\\ - \left[ - \begin{array}{ccc} - q&\cdots&q\\ - 0&\cdots&0\\ - \vdots&\ddots&\vdots\\ - 0&\cdots&0 - \end{array} - \right]\begin{array}{c} - \\\uparrow\\m-1\\\downarrow - \end{array} -\end{subarray} - && - R^{01} - =\begin{bmatrix} - r_{01}&\cdots&r_{01}\\ - 0&\cdots&0\\ - \vdots&\ddots&\vdots\\ - 0&\cdots&0 - \end{bmatrix} - && - R^{10} - =\begin{bmatrix} - r_{10}&\cdots&r_{10}\\ - 0&\cdots&0\\ - \vdots&\ddots&\vdots\\ - 0&\cdots&0 - \end{bmatrix} - && - D^{01} - =\begin{bmatrix} - d_{01}&\cdots&d_{01}\\ - 0&\cdots&0\\ - \vdots&\ddots&\vdots\\ - 0&\cdots&0 - \end{bmatrix} -\end{align} -where only the first row is nonzero. The other entries, which correspond to the -completely uncorrelated replicas in an \textsc{rsb} picture, are all zero -because uncorrelated vectors on the sphere are orthogonal. -The inverse of block hierarchical matrix is still a block hierarchical matrix, since +In this solution, we simultaneously find the smallest eigenvalue and information +about the orientation of its associated eigenvector: namely, its overlap with +the tangent vector that points directly toward the reference spin. This is +directly related to $x_0$. This tangent vector is $\mathbf x_{0\leftarrow +1}=\frac1{1-q}\big(\pmb\sigma_0-q\mathbf s_a\big)$, which is normalized and +lies strictly in the tangent plane of $\mathbf s_a$. Then \begin{equation} - \begin{bmatrix} - C^{00}&iR^{00}\\iR^{00}&D^{00} - \end{bmatrix}^{-1} - = - \begin{bmatrix} - (C^{00}D^{00}+R^{00}R^{00})^{-1}D^{00} & -i(C^{00}D^{00}+R^{00}R^{00})^{-1}R^{00} \\ - -i(C^{00}D^{00}+R^{00}R^{00})^{-1}R^{00} & (C^{00}D^{00}+R^{00}R^{00})^{-1}C^{00} - \end{bmatrix} + q_\textrm{min}=\frac{\mathbf x_{0\leftarrow 1}\cdot\mathbf x_\mathrm{min}}N + =\frac{x_0}{1-q} \end{equation} -Because of the structure of the 01 matrices, the volume element will depend -only on the diagonals of the matrices in this inverse block matrix. If we define -\begin{align} - \tilde c_\mathrm d^{00}&=[(C^{00}D^{00}+R^{00}R^{00})^{-1}C^{00}]_{\text d} \\ - \tilde r_\mathrm d^{00}&=[(C^{00}D^{00}+R^{00}R^{00})^{-1}R^{00}]_{\text d} \\ - \tilde d_\mathrm d^{00}&=[(C^{00}D^{00}+R^{00}R^{00})^{-1}D^{00}]_{\text d} -\end{align} -as the diagonals of the blocks of the inverse matrix, then the result of the product is -\begin{equation} - \begin{aligned} - & \begin{bmatrix} - C^{01}&iR^{01}\\iR^{10}&D^{01} - \end{bmatrix}^T - \begin{bmatrix} - C^{00}&iR^{00}\\iR^{00}&D^{00} - \end{bmatrix}^{-1} - \begin{bmatrix} - C^{01}&iR^{01}\\iR^{10}&D^{01} - \end{bmatrix} \\ - &\qquad=\begin{bmatrix} - q^2\tilde d_\mathrm d^{00}+2qr_{10}\tilde r^{00}_\mathrm d-r_{10}^2\tilde d^{00}_\mathrm d - & - i\left[d_{01}(r_{10}\tilde c^{00}_\mathrm d-q\tilde r^{00}_\mathrm d)+r_{01}(r_{10}\tilde r^{00}_\mathrm d+q\tilde d^{00}_\mathrm d)\right] - \\ - i\left[d_{01}(r_{10}\tilde c^{00}_\mathrm d-q\tilde r^{00}_\mathrm d)+r_{01}(r_{10}\tilde r^{00}_\mathrm d+q\tilde d^{00}_\mathrm d)\right] - & - d_{01}^2\tilde c^{00}_\mathrm d+2r_{01}d_{01}\tilde r^{00}_\mathrm d-r_{01}^2\tilde d^{00}_\mathrm d - \end{bmatrix} - \end{aligned} -\end{equation} -where each block is a constant $n\times n$ matrix. Because the matrices -$C^{00}$, $R^{00}$, and $D^{00}$ are diagonal in the replica symmetric case, -the diagonals of the blocks above take a simple form: -\begin{align} - \tilde c_\mathrm d^{00}=f'(1) && - \tilde r_\mathrm d^{00}=r^{00}_\mathrm df'(1) && - \tilde d_\mathrm d^{00}=d^{00}_\mathrm df'(1) -\end{align} -Once these expressions are inserted into the complexity, the limits of $n$ and -$m$ to zero can be taken, and the parameters from $D^{01}$ and $D^{11}$ can be -extremized explicitly. The resulting expression for the complexity, which must -still be extremized over the parameters $\hat\beta_1$, $r^{01}$, -$r^{11}_\mathrm d$, $r^{11}_0$, and $q^{11}_0$, is -\begin{equation} - \begin{aligned} - &\Sigma_{12}(E_0,\mu_0,E_1,\mu_1,q) - =\mathop{\mathrm{extremum}}_{\hat\beta_1,r^{11}_\mathrm d,r^{11}_0,r^{01},q^{11}_0}\Bigg\{ - \mathcal D(\mu_1)-\frac12+\hat\beta_1E_1-r^{11}_\mathrm d\mu_1 - +\hat\beta_1\big(r^{11}_\mathrm df'(1)-r^{11}_0f'(q^{11}_0)\big)\\ - &\qquad+\hat\beta_0\hat\beta_1f(q)+(\hat\beta_0r^{01}+\hat\beta_1r^{10}+r^{00}_\mathrm d r^{01})f'(q) - +\frac{r^{11}_\mathrm d-r^{11}_0}{1-q^{11}_0}(r^{10}-qr^{00}_\mathrm d)f'(q)\\ - &+\frac12\Bigg[ - \hat\beta_1^2\big(f(1)-f(q^{11}_{0})\big) - +(r^{11}_\mathrm d)^2f''(1)+2r^{01}r^{10}f''(q)-(r^{11}_0)^2f''(q^{11}_0) - +\frac{(r^{10}-qr^{00}_\mathrm d)^2}{1-q^{11}_0}f'(1) - \\ - &\qquad+\frac{1-q^2}{1-q^{11}_0}+\left( - (r^{01})^2-\frac{r^{11}_\mathrm d-r^{11}_0}{1-q^{11}_0} - \left(2qr^{01}-\frac{(1-q^2)r^{11}_0-(q^{11}_0-q^2)r^{11}_\mathrm d}{1-q^{11}_0}\right) - \right)\big(f'(1)-f'(q_{22}^{(0)})\big) \\ - &\qquad - -\frac1{f'(1)}\frac{f'(1)^2-f'(q)^2}{f'(1)-f'(q^{11}_0)} - +\frac{r^{11}_\mathrm d-r^{11}_0}{1-q^{11}_0}\big(r^{11}_\mathrm df'(1)-r^{11}_0f'(q^{11}_0)\big) - +\log\left(\frac{1-q_{11}^0}{f'(1)-f'(q_{11}^0)}\right) - \Bigg]\Bigg\} - \end{aligned} -\end{equation} -It is possible to further extremize this expression over all the other -variables but $q_0^{11}$, for which the saddle point conditions have a unique -solution. However, the resulting expression is quite complicated and provides -no insight. In practice, the complexity can be calculated in two ways. First, -the extremal problem can be done numerically, initializing from $q=0$ where the -problem reduces to that of the single-point complexity of points with energy -$E_1$ and stability $\mu_1$, and then taking small steps in $q$ or other -parameters to trace out the solution. This is how the data in all the plots of -this paper was produced. Second, the complexity can be calculated in the near -neighborhood of a reference point by expanding in small $1-q$. This is what we -describe in the next subsection. +The emergence of an isolated eigenvalue and its associated eigenvector are +shown in Fig.~\ref{fig:isolated.eigenvalue}, for the same reference point +properties as in Fig.~\ref{fig:min.neighborhood}. -\subsection{Expansion in the near neighborhood} +\begin{figure} + \includegraphics{figs/isolated_eigenvalue.pdf} + \hfill + \includegraphics{figs/eigenvector_overlap.pdf} -If there is no overlap gap between the reference point and its nearest -neighbors, their complexity can be calculated by an expansion in $1-q$. First, -we'll use this method to describe the most common type of stationary point in -the close vicinity of a reference point. The most common neighbors of a -reference point are given by further maximizing the two-point complexity over -the energy $E_1$ and stability $\mu_1$ of the nearby points. This gives the -conditions -\begin{align} - \hat\beta_1=0 && - \mu_1=2r^{11}_\mathrm df''(1) -\end{align} -where the second is only true for $\mu_1^2\leq\mu_\mathrm m^2$, i.e., when the -nearby points are saddle points or marginal minima. When these conditions are -inserted into the complexity, an expansion is made in small $1-q$, and the -saddle point in the remaining parameters is taken, the result is -\begin{equation} - \Sigma_{12} - =\frac{f'''(1)}{8f''(1)^2}(\mu_\mathrm m^2-\mu_0^2)\left(\sqrt{2+\frac{2f''(1)\big(f''(1)-f'(1)\big)}{f'''(1)f'(1)}}-1\right)(1-q) - +O\big((1-q)^2\big) -\end{equation} -independent of $E_0$. To describe the properties of these most common -neighbors, it is convenient to first make a definition. The population of -stationary points that are most common at each energy (the blue line in -Fig.~\ref{fig:complexities}) have the relation -\begin{equation} - E_\mathrm{dom}(\mu_0)=-\frac{f'(1)^2+f(1)\big(f''(1)-f'(1)\big)}{2f''(1)f'(1)}\mu_0 -\end{equation} -between $E_0$ and $\mu_0$ for $\mu_0^2\leq\mu_\mathrm m^2$. Using this -definition, the energy and stability of the most common neighbors at small -$\Delta q$ are -\begin{align} \label{eq:expansion.E.1} - E_1&=E_0+\frac12\frac{v_f}{u_f}\big(E_0-E_\mathrm{dom}(\mu_0)\big)(1-q)^2+O\big((1-q)^3\big) \\ - \label{eq:expansion.mu.1} - \mu_1&=\mu_0-\frac{v_f}{u_f}\big(E_0-E_\mathrm{dom}(\mu_0)\big)(1-q)+O\big((1-q)^2\big) -\end{align} -The most common neighboring saddles to a reference saddle are much nearer to -the reference in energy ($\Delta q^2$) than in stability ($\Delta q$). In fact, -this scaling also holds for the entire range of neighbors to a reference -saddle, with the limits in energy scaling like $\Delta q^2$ and those of -stability scaling like $\Delta q$. + \caption{ + Properties of the isolated eigenvalue and the overlap of its associated + eigenvector with the direction of the reference point. These curves + correspond with the lower solid curve in Fig.~\ref{fig:min.neighborhood}. + \textbf{Left:} The value of the minimum eigenvalue as a function of + overlap. The dashed line shows the continuation of the bottom of the + semicircle. Where the dashed line separates from the solid line, the + isolated eigenvalue has appeared. \textbf{Right:} The overlap between the + eigenvector associated with the minimum eigenvalue and the direction of the + reference point. The overlap is zero until an isolated eigenvalue appears, + and then it grows continuously until the nearest neighbor is reached. + } \label{fig:isolated.eigenvalue} +\end{figure} -Because both expressions are proportional to $E_0-E_\mathrm{dom}(\mu_0)$, -whether the energy and stability of nearby points increases or decreases from -that of the reference point depends only on whether the energy of the reference -point is above or below that of the most common population at the same -stability. In particular, since $E_\mathrm{dom}(\mu_\mathrm m)=E_\mathrm{th}$, -the threshold energy is also the pivot around which the points asymptotically -nearby marginal minima change their properties. +\section{Conclusion} +\label{sec:conclusion} -To examine better the population of marginal points, it is necessary to look at -the next term in the series of the complexity with $\Delta q$, since the linear -coefficient becomes zero at the marginal line. This tells us something -intuitive: stable minima have an effective repulsion between points, and one -always finds a sufficiently small $\Delta q$ that no stationary points are -point any nearer. For the marginal minima, it is not clear that the same should be true. +We have computed the complexity of neighboring stationary points for the mixed +spherical models. When we studied the neighborhoods of marginal minima, we +found something striking: only those at the threshold energy have other +marginal minima nearby. For the many marginal minima away from the threshold +(including the exponential majority), there is a gap in overlap between them. -When $\mu=\mu_\mathrm m$, the linear term above vanishes. Under these conditions, the quadratic term in the expansion is +This has implications for pictures of relaxation and aging. In most $p+s$ +models studied, quenches from infinite to zero temperature (gradient descent +starting from a random point) relax towards marginal states with energies above +the threshold energy \cite{Folena_2023_On}, while at least in some models a +quench to zero temperature from a temperature around the dynamic transition +relaxes towards marginal states with energies below the threshold energy +\cite{Folena_2020_Rethinking, Folena_2021_Gradient}. We found (see especially +Figs.~\ref{fig:marginal.prop.below} and \ref{fig:marginal.prop.above}) that the +neighborhoods of marginal states above and below the threshold are quite +different, and yet the emergent aging behaviors relaxing toward states above and +below the threshold seem to be the same. Therefore, this kind of dynamics +appears to be insensitive to the neighborhood of the marginal state being +approached. To understand something better about why certain states attract the +dynamics in certain situations, nonlocal information, like the +structure of their entire basin of attraction, seems vital. + +It is possible that replica symmetry breaking among the constrained stationary +points could change the details of the two-point complexity of very nearby +states. Indeed, it is difficult to rule out \textsc{rsb} in complexity +calculations. However, such corrections would not change the overarching +conclusions of this paper, namely that most marginal minima are separated from +each other by a macroscopic overlap gap and high barriers. This is because the +replica symmetric complexity bounds any \textsc{rsb} complexities from above, +and so \textsc{rsb} corrections can only decrease the complexity. Therefore, +the overlap gaps, which correspond to regions of negative complexity, cannot be +removed by a more detailed saddle point ansatz. + +Our calculation studied the neighborhood of typical reference points with the +given energy and stability. However, it is possible that marginal minima with +atypical neighborhoods actually attract the dynamics. To determine this, a +different type of calculation is needed. As our calculation is akin to the +quenched Franz--Parisi potential, study of atypical neighborhoods would entail +something like the annealed Franz--Parisi approach, i.e., \begin{equation} - \Sigma_{12} - =\frac12\frac{f'''(1)v_f}{f''(1)^{3/2}u_f} - \left(\sqrt{\frac{2\big[f'(1)(f'''(1)-f''(1))+f''(1)^2\big]}{f'(1)f'''(1)}}-1\right)\big(E_0-E_\textrm{th}\big)(1-q)^2+O\big((1-q)^3\big) + \Sigma^*(E_0,\mu_0,E_1,\mu_1,q)=\frac1N\overline{\log\left( + \int d\nu_H(\pmb\sigma,\varsigma\mid E_0,\mu_0)\,d\nu_H(\mathbf s,\omega\mid E_1,\mu_1)\,\delta(Nq-\pmb\sigma\cdot\mathbf s) + \right)} \end{equation} -Note that this expression is only true for $\mu=\mu_\mathrm m$. Therefore, -among marginal minima, when $E_0$ is greater than the threshold one finds -neighbors at arbitrarily close distance. When $E_0$ is less than the threshold, -the complexity of nearby points is negative, and there is a desert where none -are found. +which puts the two points on equal footing. This calculation and exploration of +the atypical neighborhoods it reveals is a clear future direction. -\begin{figure} - \centering - \includegraphics{figs/expansion_energy.pdf} - \hspace{1em} - \includegraphics{figs/expansion_stability.pdf} +The methods developed in this paper are straightforwardly (if not easily) +generalized to landscapes with replica symmetry broken complexities +\cite{Kent-Dobias_2023_How}. We suspect that many of the qualitative features +of this study would persist, with neighboring states being divided into +different clusters based on the \textsc{rsb} order but with the basic presence +or absence of overlap gaps and the nature of the stability of near-neighbors +remaining unchanged. Interesting structure might emerge in the arrangement of +marginal states in \textsc{frsb} systems, where the ground state itself is +marginal and coincides with the threshold. - \caption{ - Demonstration of the convergence of the $(1-q)$-expansion for marginal - reference minima. Solid lines and shaded region show are the same as in - Fig.~\ref{fig:marginal.prop.above} for $E_0-E_\mathrm{th}\simeq0.00667$. - The dotted lines show the expansion of most common neighbors, while the - dashed lines in both plots show the expansion for the minimum and maximum - energies and stabilities found at given $q$. - } \label{fig:expansion} -\end{figure} +\paragraph{Acknowledgements} -The properties of the nearby states above the threshold can be -further quantified. The most common points are still given by -\eqref{eq:expansion.E.1} and \eqref{eq:expansion.mu.1}, but the range of -available points can also be computed, and one finds that the stability lies in -the range -$\mu_1=\mu_\mathrm m+\delta\mu_1(1-q)\pm\delta\mu_2(1-q)^{3/2}+O\big((1-q)^2\big)$ -where $\delta\mu_1$ is given by the coefficient in \eqref{eq:expansion.mu.1} -and +The author would like to thank Valentina Ros, Giampaolo Folena, Chiara +Cammarota, and Jorge Kurchan for useful discussions related to this work. + +\paragraph{Funding information} + +JK-D is supported by a \textsc{DynSysMath} Specific Initiative by the +INFN. + +\appendix + +\section{Details of calculation for the two-point complexity} +\label{sec:complexity-details} + +\subsection{The Hessian factors} + +The double partial derivatives of the energy are Gaussian with the variance \begin{equation} - \delta\mu_2=\frac{v_f}{f'(1)f''(1)^{3/4}}\sqrt{ - \frac{E_0-E_\mathrm{th}}2\frac{2f''(1)\big(f''(1)-f'(1)\big)+f'(1)f'''(1)}{u_f} - } + \overline{(\partial_i\partial_jH(\mathbf s))^2}=\frac1Nf''(1) \end{equation} -Similarly, one finds that the energy lies in the range $E_1=E_0+\delta -E_1(1-q)^2\pm\delta E_2(1-q)^{5/2}+O\big((1-q)^3\big)$ for $\delta E_1$ given -by the coefficient in \eqref{eq:expansion.E.1} and +which means that the matrix of partial derivatives belongs to the GOE class. Its spectrum is given by the Wigner semicircle \begin{equation} - \begin{aligned} - \delta E_2 - &=\frac{\sqrt{E_0-E_\mathrm{th}}}{4f'(1)f''(1)^{3/4}}\bigg( - \frac{ - v_f - }{3u_f} - \big[ - f'(1)(2f''(1)-(2-(2-\delta q_0)\delta q_0)f'''(1))-2f''(1)^2 - \big] - \\ - &\hspace{12pc}\times - \big[f'(1)\big(6f''(1)+(18-(6-\delta q_0)\delta q_0)f'''(1)\big)-6f''(1)^2 - \big] - \bigg)^\frac12 - \end{aligned} + \rho(\lambda)=\begin{cases} + \frac2{\pi}\sqrt{1-\big(\frac{\lambda}{\mu_\text m}\big)^2} & \lambda^2\leq\mu_\text m^2 \\ + 0 & \text{otherwise} + \end{cases} \end{equation} -and $\delta q_0$ is the coefficient in the expansion $q_0=1-\delta q_0(1-q)+O((1-q)^2)$ and is given by the real root to the quintic equation +with radius $\mu_\text m=\sqrt{4f''(1)}$. Since the Hessian differs from the +matrix of partial derivatives by adding the constant diagonal matrix $\omega +I$, it follows that the spectrum of the Hessian is a Wigner semicircle shifted +by $\omega$, or $\rho(\lambda+\omega)$. + +The average over factors depending on the Hessian alone can be made separately +from those depending on the gradient or energy, since for random Gaussian +fields the Hessian is independent of these \cite{Bray_2007_Statistics}. In +principle the fact that we have conditioned the Hessian to belong to stationary +points of certain energy, stability, and proximity to another stationary point +will modify its statistics, but these changes will only appear at subleading +order in $N$ \cite{Ros_2019_Complexity}. At leading order, the various expectations factorize, each yielding \begin{equation} - 0=((16-(6-\delta q_0)\delta q_0)\delta q_0-12)f'(1)f'''(1)-2\delta q_0(f''(1)-f'(1))f''(1) + \overline{\big|\det\operatorname{Hess}H(\mathbf s,\omega)\big|\,\delta\big(N\mu-\operatorname{Tr}\operatorname{Hess}H(\mathbf s,\omega)\big)} + =e^{N\int d\lambda\,\rho(\lambda+\mu)\log|\lambda|}\delta(N\mu-N\omega) \end{equation} -These predictions from the small $1-q$ expansion are compared with numeric -saddle points for the complexity of marginal minima in -Fig.~\ref{fig:expansion}, and the results agree well at small $1-q$. - - -\section{Isolated eigenvalue} -\label{sec:eigenvalue} - -The two-point complexity depends on the spectrum at both stationary points -through the determinant of their Hessians, but only on the bulk of the -distribution. As we saw, this bulk is unaffected by the conditions of energy -and proximity. However, these conditions give rise to small-rank perturbations -to the Hessian, which can lead a subextensive number of eigenvalues leaving the -bulk. We study the possibility of \emph{one} stray eigenvalue. - -We use a technique recently developed to find the smallest eigenvalue of a -random matrix \cite{Ikeda_2023_Bose-Einstein-like}. One defines a quadratic -statistical mechanics model with configurations defined on the sphere, whose -interaction tensor is given by the matrix of interest. By construction, the -ground state is located in the direction of the eigenvector associated with the -smallest eigenvalue, and the ground state energy is proportional to that -eigenvalue. - -\begin{figure} - \centering - \begin{tikzpicture} - \def\R{4 } % sphere radius - \def\Rt{2 } % tangent plane radius - \def\angEl{15} % elevation angle - \def\angsa{-160} % azimuth of s_1 - \def\angq{40} % elevation of constraint circle - \filldraw[ball color=white] (0,0) circle (\R); - % \filldraw[fill=white] (0,0) circle (\R); - - \foreach \t in {0,\angq} { \DrawLatitudeCircle[\R]{\t} } - %\foreach \t in {\angsa} { \DrawLongitudeCircle[\R]{\t} } - - \pgfmathsetmacro\H{\R*cos(\angEl)} % distance to north pole - \coordinate (O) at (0,0); - \node[circle,draw,black,scale=0.3] at (0,0) {}; - \coordinate (N) at (0,\H); - \draw node[right=10,below] at (0,\H){$\pmb\sigma_1$}; - \draw[thick, ->](O)--(N); - - \NewLatitudePlane[planeP]{\R}{\angEl}{\angq}; - \path[planeP] (\angsa:\R) coordinate (P); - \path[planeP] (0:1.5*\R) coordinate (Q); - \path[planeP] (0:\R) coordinate (Q2); - \draw[left] node at (P){$\mathbf s_1$}; - - \NewLatitudePlane[equator]{\R}{\angEl}{00}; - \path[equator] (-30:\R) coordinate (Pprime); - \path[equator] (0:{1.5*cos(\angq)*\R}) coordinate (Qe); - \path[equator] (0:\R) coordinate (Qe2); - \draw node[right=5,below] at (Pprime){$\pmb\sigma_c$}; - - \NewLatitudePlane[sbplane]{\R}{\angEl}{\angq}; - \path[sbplane] (20:\R) coordinate (sb); - \draw node[right=3,above=1] at (sb){$\mathbf s_b$}; - - \TangentPlane[tplane]{\R}{\angEl}{\angq}{\angsa}; - \draw[tplane,fill=gray,fill opacity=0.3] circle (\Rt); - \draw[tplane,->,thick] (0,0) -> ({\Rt*cos(160)},{\Rt*sin(160)}) node[above=1.5,right] {$\mathbf x_a$}; - \draw[tplane,->,thick] (0,0) -> ({\Rt*cos(250)},{\Rt*sin(250)}) node[above=1.5,left=0.1] {$\mathbf x_b$}; - - \draw[thick, ->] (O)->(P); - \draw[thick, ->] (O)->(Pprime); - \draw[thick, ->] (O)->(sb); - - \draw[dotted] (Qe) -- (Qe2); - \draw[dotted] (Q2) -- (Q); - \draw[decorate, decoration = {brace,raise=3}] (Q) -- (Qe) node[pos=0.5,right=7]{$q$}; - \end{tikzpicture} - \caption{ - A sketch of the vectors involved in the calculation of the isolated - eigenvalue. All replicas $\mathbf x$, which correspond with candidate - eigenvectors of the Hessian evaluated at $\mathbf s_1$, sit in an $N-2$ - sphere corresponding with the tangent plane (not to scale) of the first - $\mathbf s$ replica. All of the $\mathbf s$ replicas lie on the sphere, - constrained to be at fixed overlap $q$ with the first of the $\pmb\sigma$ - replicas, the reference configuration. All of the $\pmb\sigma$ replicas lie - on the sphere. - } \label{fig:sphere} -\end{figure} - -Our matrix of interest is the Hessian evaluated at a stationary point of the mixed spherical -model, conditioned on the relative position, energies, and stabilities -discussed above. We must restrict the artificial spherical model to lie in the -tangent plane of the `real' spherical configuration space at the point of -interest, to avoid our eigenvector pointing in a direction that violates the -spherical constraint. A sketch of the setup is shown in Fig.~\ref{fig:sphere}. The free energy of this model given a point $\mathbf s$ -and a specific realization of the disordered Hamiltonian is +Therefore, all of the Lagrange multipliers are fixed to the stabilities $\mu$. We define the function \begin{equation} \begin{aligned} - \beta F_H(\beta\mid\mathbf s,\omega) - &=-\frac1N\log\left(\int d\mathbf x\,\delta(\mathbf x\cdot\mathbf s)\delta(\|\mathbf x\|^2-N)\exp\left\{ - -\beta\frac12\mathbf x^T\operatorname{Hess}H(\mathbf s,\omega)\mathbf x - \right\}\right) \\ - &=-\lim_{\ell\to0}\frac1N\frac\partial{\partial\ell}\int\left[\prod_{\alpha=1}^\ell d\mathbf x_\alpha\,\delta(\mathbf x_\alpha^T\mathbf s)\delta(N-\mathbf x_\alpha^T\mathbf x_\alpha)\exp\left\{ - -\beta\frac12\mathbf x^T_\alpha\big(\partial\partial H(\mathbf s)+\omega I\big)\mathbf x_\alpha - \right\}\right] + \mathcal D(\mu) + &=\int d\lambda\,\rho(\lambda+\mu)\log|\lambda| \\ + &=\begin{cases} + \frac12+\log\left(\frac12\mu_\text m\right)+\frac{\mu^2}{\mu_\text m^2} + & \mu^2\leq\mu_\text m^2 \\ + \frac12+\log\left(\frac12\mu_\text m\right)+\frac{\mu^2}{\mu_\text m^2} + -\left|\frac{\mu}{\mu_\text m}\right|\sqrt{\big(\frac\mu{\mu_\text m}\big)^2-1} + -\log\left(\left|\frac{\mu}{\mu_\text m}\right|-\sqrt{\big(\frac\mu{\mu_\text m}\big)^2-1}\right) & \mu^2>\mu_\text m^2 + \end{cases} \end{aligned} \end{equation} -where the first $\delta$-function keeps the configurations in the tangent -plane, and the second enforces the spherical constraint. We have anticipated -treating the logarithm with replicas. We are interested in points $\mathbf s$ -that have certain properties: they are stationary points of $H$ with given -energy density and stability, and fixed overlap from a reference configuration -$\pmb\sigma$. We therefore average the free energy above over such points, -giving +and the full factor due to the Hessians is \begin{equation} - \begin{aligned} - F_H(\beta\mid E_1,\mu_1,q,\pmb\sigma) - &=\int\frac{d\nu_H(\mathbf s,\omega\mid E_1,\mu_1)\delta(Nq-\pmb\sigma\cdot\mathbf s)}{\int d\nu_H(\mathbf s',\omega'\mid E_1,\mu_1)\delta(Nq-\pmb\sigma\cdot\mathbf s')}F_H(\beta\mid\mathbf s,\omega) \\ - &=\lim_{n\to0}\int\left[\prod_{a=1}^nd\nu_H(\mathbf s_a,\omega_a\mid E_1,\mu_1)\,\delta(Nq-\pmb\sigma\cdot\mathbf s_a)\right]F_H(\beta\mid\mathbf s_1,\omega_1) - \end{aligned} + e^{Nm\mathcal D(\mu_0)+Nn\mathcal D(\mu_1)}\left[\prod_a^m\delta(N\mu_0-N\varsigma_a)\right]\left[\prod_a^n\delta(N\mu_1-N\omega_a)\right] \end{equation} -again anticipating the use of replicas. Finally, the reference configuration $\pmb\sigma$ should itself be a stationary point of $H$ with its own energy density and stability. Averaging over these conditions gives + +\subsection{The other factors} + +Having integrated over the Lagrange multipliers using the $\delta$-functions +resulting from the average of the Hessians, any $\delta$-functions in the +remaining integrand we Fourier transform into their integral representation +over auxiliary fields. The resulting integrand has the form \begin{equation} - \begin{aligned} - F_H(\beta\mid E_1,\mu_1,E_2,\mu_2,q) - &=\int\frac{d\nu_H(\pmb\sigma,\varsigma\mid E_0,\mu_0)}{\int d\nu_H(\pmb\sigma',\varsigma'\mid E_0,\mu_0)}\,F_H(\beta\mid E_1,\mu_1,q,\pmb\sigma) \\ - &=\lim_{m\to0}\int\left[\prod_{a=1}^m d\nu_H(\pmb\sigma_a,\varsigma_a\mid E_0,\mu_0)\right]\,F_H(\beta\mid E_1,\mu_1,q,\pmb\sigma_1) - \end{aligned} + e^{ + Nm\hat\beta_0E_0+Nn\hat\beta_1E_1 + -\sum_a^m\left[(\pmb\sigma_a\cdot\hat{\pmb\sigma}_a)\mu_0 + -\frac12\hat\mu_0(N-\pmb\sigma_a\cdot\pmb\sigma_a) + \right] + -\sum_a^n\left[(\mathbf s_a\cdot\hat{\mathbf s}_a)\mu_1 + -\frac12\hat\mu_1(N-\mathbf s_a\cdot\mathbf s_a) + -\frac12\hat\mu_{12}(Nq-\pmb\sigma_1\cdot\mathbf s_a) + \right] + +\int d\mathbf t\,\mathcal O(\mathbf t)H(\mathbf t) + } \end{equation} -This formidable expression is now ready to be averaged over the disordered Hamiltonians $H$. Once averaged, -the minimum eigenvalue of the conditioned Hessian is then given by twice the ground state energy, or +where we have introduced the linear operator \begin{equation} - \lambda_\text{min}=2\lim_{\beta\to\infty}\overline{F_H(\beta\mid E_1,\mu_1,E_2,\mu_2,q)} + \mathcal O(\mathbf t) + =\sum_a^m\delta(\mathbf t-\pmb\sigma_a)\left( + i\hat{\pmb\sigma}_a\cdot\partial_{\mathbf t}-\hat\beta_0 + \right) + + + \sum_a^n\delta(\mathbf t-\mathbf s_a)\left( + i\hat{\mathbf s}_a\cdot\partial_{\mathbf t}-\hat\beta_1 + \right) \end{equation} -For this calculation, there are three different sets of replicated variables. -Note that, as for the computation of the complexity, the $\pmb\sigma_1$ and -$\mathbf s_1$ replicas are \emph{special}. The first again is the only of the -$\pmb\sigma$ replicas constrained to lie at fixed overlap with \emph{all} the -$\mathbf s$ replicas, and the second is the only of the $\mathbf s$ replicas at -which the Hessian is evaluated. - - -In this solution, we simultaneously find the smallest eigenvalue and information -about the orientation of its associated eigenvector: namely, its overlap with -the tangent vector that points directly toward the reference spin. This is -directly related to $x_0$. This tangent vector is $\mathbf x_{0\leftarrow -1}=\frac1{1-q}\big(\pmb\sigma_0-q\mathbf s_a\big)$, which is normalized and -lies strictly in the tangent plane of $\mathbf s_a$. Then +Here the $\hat\beta$s are the fields auxiliary to the energy constraints, the +$\hat\mu$s are auxiliary to the spherical and overlap constraints, and the +$\hat{\pmb\sigma}$s and $\hat{\mathbf s}$s are auxiliary to the constraint that +the gradient be zero. +We have written the $H$-dependent terms in this strange form for the ease of taking the average over $H$: since it is Gaussian-correlated, it follows that \begin{equation} - q_\textrm{min}=\frac{\mathbf x_{0\leftarrow 1}\cdot\mathbf x_\mathrm{min}}N - =\frac{x_0}{1-q} + \overline{e^{\int d\mathbf t\,\mathcal O(\mathbf t)H(\mathbf t)}} + =e^{\frac12\int d\mathbf t\,d\mathbf t'\,\mathcal O(\mathbf t)\mathcal O(\mathbf t')\overline{H(\mathbf t)H(\mathbf t')}} + =e^{N\frac12\int d\mathbf t\,d\mathbf t'\,\mathcal O(\mathbf t)\mathcal O(\mathbf t')f\big(\frac{\mathbf t\cdot\mathbf t'}N\big)} \end{equation} -The emergence of an isolated eigenvalue and its associated eigenvector are -shown in Fig.~\ref{fig:isolated.eigenvalue}, for the same reference point -properties as in Fig.~\ref{fig:min.neighborhood}. - -\begin{figure} - \includegraphics{figs/isolated_eigenvalue.pdf} - \hfill - \includegraphics{figs/eigenvector_overlap.pdf} - - \caption{ - Properties of the isolated eigenvalue and the overlap of its associated - eigenvector with the direction of the reference point. These curves - correspond with the lower solid curve in Fig.~\ref{fig:min.neighborhood}. - \textbf{Left:} The value of the minimum eigenvalue as a function of - overlap. The dashed line shows the continuation of the bottom of the - semicircle. Where the dashed line separates from the solid line, the - isolated eigenvalue has appeared. \textbf{Right:} The overlap between the - eigenvector associated with the minimum eigenvalue and the direction of the - reference point. The overlap is zero until an isolated eigenvalue appears, - and then it grows continuously until the nearest neighbor is reached. - } \label{fig:isolated.eigenvalue} -\end{figure} +It remains only to apply the doubled operators to $f$ and then evaluate the simple integrals over the $\delta$ measures. We do not include these details, which are standard. -\section{Conclusion} -\label{sec:conclusion} +\subsection{Hubbard--Stratonovich} -We have computed the complexity of neighboring stationary points for the mixed -spherical models. When we studied the neighborhoods of marginal minima, we -found something striking: only those at the threshold energy have other -marginal minima nearby. For the many marginal minima away from the threshold -(including the exponential majority), there is a gap in overlap between them. +Having expanded this expression, we are left with an argument in the exponential which is a function of scalar products between the fields $\mathbf s$, $\hat{\mathbf s}$, $\pmb\sigma$, and $\hat{\pmb\sigma}$. We will change integration coordinates from these fields to matrix fields given by their scalar products, defined as +\begin{equation} \label{eq:fields} + \begin{aligned} + C^{00}_{ab}=\frac1N\pmb\sigma_a\cdot\pmb\sigma_b && + R^{00}_{ab}=-i\frac1N{\pmb\sigma}_a\cdot\hat{\pmb\sigma}_b && + D^{00}_{ab}=\frac1N\hat{\pmb\sigma}_a\cdot\hat{\pmb\sigma}_b \\ + C^{01}_{ab}=\frac1N\pmb\sigma_a\cdot\mathbf s_b && + R^{01}_{ab}=-i\frac1N{\pmb\sigma}_a\cdot\hat{\mathbf s}_b && + R^{10}_{ab}=-i\frac1N\hat{\pmb\sigma}_a\cdot{\mathbf s}_b && + D^{01}_{ab}=\frac1N\hat{\pmb\sigma}_a\cdot\hat{\mathbf s}_b \\ + C^{11}_{ab}=\frac1N\mathbf s_a\cdot\mathbf s_b && + R^{11}_{ab}=-i\frac1N{\mathbf s}_a\cdot\hat{\mathbf s}_b && + D^{11}_{ab}=\frac1N\hat{\mathbf s}_a\cdot\hat{\mathbf s}_b + \end{aligned} +\end{equation} +We insert into the integral the product of $\delta$-functions enforcing these +definitions, integrated over the new matrix fields, which is equivalent to +multiplying by one. Once this is done, the many scalar products appearing +throughout can be replaced by the matrix fields, and the original vector fields +can be integrated over. Conjugate matrix field integrals created when the +$\delta$-functions are promoted to exponentials can be evaluated by saddle +point in the standard way, yielding an effective action depending on the above +matrix fields alone. -This has implications for pictures of relaxation and aging. In most $p+s$ -models studied, quenches from infinite to zero temperature (gradient descent -starting from a random point) relax towards marginal states with energies above -the threshold energy \cite{Folena_2023_On}, while at least in some models a -quench to zero temperature from a temperature around the dynamic transition -relaxes towards marginal states with energies below the threshold energy -\cite{Folena_2020_Rethinking, Folena_2021_Gradient}. We found (see especially -Figs.~\ref{fig:marginal.prop.below} and \ref{fig:marginal.prop.above}) that the -neighborhoods of marginal states above and below the threshold are quite -different, and yet the emergent aging behaviors relaxing toward states above and -below the threshold seem to be the same. Therefore, this kind of dynamics -appears to be insensitive to the neighborhood of the marginal state being -approached. To understand something better about why certain states attract the -dynamics in certain situations, nonlocal information, like the -structure of their entire basin of attraction, seems vital. +\subsection{Saddle point} -It is possible that replica symmetry breaking among the constrained stationary -points could change the details of the two-point complexity of very nearby -states. Indeed, it is difficult to rule out \textsc{rsb} in complexity -calculations. However, such corrections would not change the overarching -conclusions of this paper, namely that most marginal minima are separated from -each other by a macroscopic overlap gap and high barriers. This is because the -replica symmetric complexity bounds any \textsc{rsb} complexities from above, -and so \textsc{rsb} corrections can only decrease the complexity. Therefore, -the overlap gaps, which correspond to regions of negative complexity, cannot be -removed by a more detailed saddle point ansatz. +We will always assume that the square matrices $C^{00}$, $R^{00}$, $D^{00}$, +$C^{11}$, $R^{11}$, and $D^{11}$ are hierarchical matrices, with each set of +three sharing the same hierarchical structure. In particular, we immediately +define $c_\mathrm d^{00}$, $r_\mathrm d^{00}$, $d_\mathrm d^{00}$, $c_\mathrm d^{11}$, $r_\mathrm d^{11}$, and +$d_\mathrm d^{11}$ as the value of the diagonal elements of these matrices, +respectively. Note that $c_\mathrm d^{00}=c_\mathrm d^{11}=1$ due to the spherical constraint. -Our calculation studied the neighborhood of typical reference points with the -given energy and stability. However, it is possible that marginal minima with -atypical neighborhoods actually attract the dynamics. To determine this, a -different type of calculation is needed. As our calculation is akin to the -quenched Franz--Parisi potential, study of atypical neighborhoods would entail -something like the annealed Franz--Parisi approach, i.e., +Defining the `block' fields $\mathcal Q_{00}=(\hat\beta_0, \hat\mu_0, C^{00}, +R^{00}, D^{00})$, $\mathcal Q_{11}=(\hat\beta_1, \hat\mu_1, C^{11}, R^{11}, +D^{11})$, and $\mathcal Q_{01}=(\hat\mu_{01},C^{01},R^{01},R^{10},D^{01})$ +the resulting complexity is \begin{equation} - \Sigma^*(E_0,\mu_0,E_1,\mu_1,q)=\frac1N\overline{\log\left( - \int d\nu_H(\pmb\sigma,\varsigma\mid E_0,\mu_0)\,d\nu_H(\mathbf s,\omega\mid E_1,\mu_1)\,\delta(Nq-\pmb\sigma\cdot\mathbf s) - \right)} + \Sigma_{01} + =\frac1N\lim_{n\to0}\lim_{m\to0}\frac\partial{\partial n}\int d\mathcal Q_{00}\,d\mathcal Q_{11}\,d\mathcal Q_{01}\,e^{Nm\mathcal S_0(\mathcal Q_{00})+Nn\mathcal S_1(\mathcal Q_{11},\mathcal Q_{01}\mid\mathcal Q_{00})} \end{equation} -which puts the two points on equal footing. This calculation and exploration of -the atypical neighborhoods it reveals is a clear future direction. - -The methods developed in this paper are straightforwardly (if not easily) -generalized to landscapes with replica symmetry broken complexities -\cite{Kent-Dobias_2023_How}. We suspect that many of the qualitative features -of this study would persist, with neighboring states being divided into -different clusters based on the \textsc{rsb} order but with the basic presence -or absence of overlap gaps and the nature of the stability of near-neighbors -remaining unchanged. Interesting structure might emerge in the arrangement of -marginal states in \textsc{frsb} systems, where the ground state itself is -marginal and coincides with the threshold. - -\paragraph{Acknowledgements} - -The author would like to thank Valentina Ros, Giampaolo Folena, Chiara -Cammarota, and Jorge Kurchan for useful discussions related to this work. - -\paragraph{Funding information} +where +\begin{equation} \label{eq:one-point.action} + \begin{aligned} + &\mathcal S_0(\mathcal Q_{00}) + =\hat\beta_0E_0-r^{00}_\mathrm d\mu_0-\frac12\hat\mu_0(1-c^{00}_\mathrm d)+\mathcal D(\mu_0)\\ + &\quad+\frac1m\bigg\{ + \frac12\sum_{ab}^m\left[ + \hat\beta_1^2f(C^{00}_{ab})+(2\hat\beta_1R^{00}_{ab}-D^{00}_{ab})f'(C^{00}_{ab})+(R_{ab}^{00})^2f''(C_{ab}^{00}) + \right]+\frac12\log\det\begin{bmatrix}C^{00}&R^{00}\\R^{00}&D^{00}\end{bmatrix} +\bigg\} +\end{aligned} +\end{equation} +is the action for the ordinary, one-point complexity, and remainder is given by +\begin{equation} \label{eq:two-point.action} + \begin{aligned} + &\mathcal S(\mathcal Q_{11},\mathcal Q_{01}\mid\mathcal Q_{00}) + =\hat\beta_1E_1-r^{11}_\mathrm d\mu_1-\frac12\hat\mu_1(1-c^{11}_\mathrm d)+\mathcal D(\mu_1) \\ + &\quad+\frac1n\sum_b^n\left\{-\frac12\hat\mu_{12}(q-C^{01}_{1b})+\sum_a^m\left[ + \hat\beta_0\hat\beta_1f(C^{01}_{ab})+(\hat\beta_0R^{01}_{ab}+\hat\beta_1R^{10}_{ab}-D^{01}_{ab})f'(C^{01}_{ab})+R^{01}_{ab}R^{10}_{ab}f''(C^{01}_{ab}) + \right]\right\} + \\ + &\quad+\frac1n\bigg\{ + \frac12\sum_{ab}^n\left[ + \hat\beta_1^2f(C^{11}_{ab})+(2\hat\beta_1R^{11}_{ab}-D^{11}_{ab})f'(C^{11}_{ab})+(R^{11}_{ab})^2f''(C^{11}_{ab}) + \right]\\ + &\quad+\frac12\log\det\left( + \begin{bmatrix} + C^{11}&iR^{11}\\iR^{11}&D^{11} + \end{bmatrix}- + \begin{bmatrix} + C^{01}&iR^{01}\\iR^{10}&D^{01} + \end{bmatrix}^T + \begin{bmatrix} + C^{00}&iR^{00}\\iR^{00}&D^{00} + \end{bmatrix}^{-1} + \begin{bmatrix} + C^{01}&iR^{01}\\iR^{10}&D^{01} + \end{bmatrix} + \right) + \bigg\} + \end{aligned} +\end{equation} +Because of the structure of this problem in the twin limits of $m$ and $n$ to +zero, the parameters $\mathcal Q_{00}$ can be evaluated at a saddle point of +$\mathcal S_0$ alone. This means that these parameters will take the same value +they take when the ordinary, 1-point complexity is calculated. For a replica +symmetric complexity of the reference point, this results in +\begin{align} + \hat\beta_0 + &=-\frac{\mu_0f'(1)+E_0\big(f'(1)+f''(1)\big)}{u_f}\\ + r_\mathrm d^{00} + &=\frac{\mu_0f(1)+E_0f'(1)}{u_f} \\ + d_\mathrm d^{00} + &=\frac1{f'(1)} + -\left( + \frac{\mu_0f(1)+E_0f'(1)}{u_f} + \right)^2 +\end{align} +where we define for brevity (here and elsewhere) the constants +\begin{align} + u_f=f(1)\big(f'(1)+f''(1)\big)-f'(1)^2 + && + v_f=f'(1)\big(f''(1)+f'''(1)\big)-f''(1)^2 +\end{align} +Note that because the coefficients of $f$ must be nonnegative for $f$ to +be a sensible covariance, both $u_f$ and $v_f$ are strictly positive. Note also +that $u_f=v_f=0$ if $f$ is a homogeneous polynomial as in the pure models. +These expressions are invalid for the pure models because $\mu_0$ and $E_0$ +cannot be fixed independently; we would have done the equivalent of inserting +two identical $\delta$-functions. For the pure models, the terms $\hat\beta_0$ and +$\hat\beta_1$ must be set to zero in our prior formulae (as if the energy was +not constrained) and then the saddle point taken. -JK-D is supported by a \textsc{DynSysMath} Specific Initiative by the -INFN. -\appendix +In general, we except the $m\times n$ matrices $C^{01}$, $R^{01}$, $R^{10}$, +and $D^{01}$ to have constant \emph{rows} of length $n$, with blocks of rows +corresponding to the \textsc{rsb} structure of the single-point complexity. For +the scope of this paper, where we restrict ourselves to replica symmetric +complexities, they have the following form at the saddle point: +\begin{align} \label{eq:01.ansatz} + C^{01}= + \begin{subarray}{l} + \hphantom{[}\begin{array}{ccc}\leftarrow&n&\rightarrow\end{array}\hphantom{\Bigg]}\\ + \left[ + \begin{array}{ccc} + q&\cdots&q\\ + 0&\cdots&0\\ + \vdots&\ddots&\vdots\\ + 0&\cdots&0 + \end{array} + \right]\begin{array}{c} + \\\uparrow\\m-1\\\downarrow + \end{array} +\end{subarray} + && + R^{01} + =\begin{bmatrix} + r_{01}&\cdots&r_{01}\\ + 0&\cdots&0\\ + \vdots&\ddots&\vdots\\ + 0&\cdots&0 + \end{bmatrix} + && + R^{10} + =\begin{bmatrix} + r_{10}&\cdots&r_{10}\\ + 0&\cdots&0\\ + \vdots&\ddots&\vdots\\ + 0&\cdots&0 + \end{bmatrix} + && + D^{01} + =\begin{bmatrix} + d_{01}&\cdots&d_{01}\\ + 0&\cdots&0\\ + \vdots&\ddots&\vdots\\ + 0&\cdots&0 + \end{bmatrix} +\end{align} +where only the first row is nonzero. The other entries, which correspond to the +completely uncorrelated replicas in an \textsc{rsb} picture, are all zero +because uncorrelated vectors on the sphere are orthogonal. +The inverse of block hierarchical matrix is still a block hierarchical matrix, since +\begin{equation} + \begin{bmatrix} + C^{00}&iR^{00}\\iR^{00}&D^{00} + \end{bmatrix}^{-1} + = + \begin{bmatrix} + (C^{00}D^{00}+R^{00}R^{00})^{-1}D^{00} & -i(C^{00}D^{00}+R^{00}R^{00})^{-1}R^{00} \\ + -i(C^{00}D^{00}+R^{00}R^{00})^{-1}R^{00} & (C^{00}D^{00}+R^{00}R^{00})^{-1}C^{00} + \end{bmatrix} +\end{equation} +Because of the structure of the 01 matrices, the volume element will depend +only on the diagonals of the matrices in this inverse block matrix. If we define +\begin{align} + \tilde c_\mathrm d^{00}&=[(C^{00}D^{00}+R^{00}R^{00})^{-1}C^{00}]_{\text d} \\ + \tilde r_\mathrm d^{00}&=[(C^{00}D^{00}+R^{00}R^{00})^{-1}R^{00}]_{\text d} \\ + \tilde d_\mathrm d^{00}&=[(C^{00}D^{00}+R^{00}R^{00})^{-1}D^{00}]_{\text d} +\end{align} +as the diagonals of the blocks of the inverse matrix, then the result of the product is +\begin{equation} + \begin{aligned} + & \begin{bmatrix} + C^{01}&iR^{01}\\iR^{10}&D^{01} + \end{bmatrix}^T + \begin{bmatrix} + C^{00}&iR^{00}\\iR^{00}&D^{00} + \end{bmatrix}^{-1} + \begin{bmatrix} + C^{01}&iR^{01}\\iR^{10}&D^{01} + \end{bmatrix} \\ + &\qquad=\begin{bmatrix} + q^2\tilde d_\mathrm d^{00}+2qr_{10}\tilde r^{00}_\mathrm d-r_{10}^2\tilde d^{00}_\mathrm d + & + i\left[d_{01}(r_{10}\tilde c^{00}_\mathrm d-q\tilde r^{00}_\mathrm d)+r_{01}(r_{10}\tilde r^{00}_\mathrm d+q\tilde d^{00}_\mathrm d)\right] + \\ + i\left[d_{01}(r_{10}\tilde c^{00}_\mathrm d-q\tilde r^{00}_\mathrm d)+r_{01}(r_{10}\tilde r^{00}_\mathrm d+q\tilde d^{00}_\mathrm d)\right] + & + d_{01}^2\tilde c^{00}_\mathrm d+2r_{01}d_{01}\tilde r^{00}_\mathrm d-r_{01}^2\tilde d^{00}_\mathrm d + \end{bmatrix} + \end{aligned} +\end{equation} +where each block is a constant $n\times n$ matrix. Because the matrices +$C^{00}$, $R^{00}$, and $D^{00}$ are diagonal in the replica symmetric case, +the diagonals of the blocks above take a simple form: +\begin{align} + \tilde c_\mathrm d^{00}=f'(1) && + \tilde r_\mathrm d^{00}=r^{00}_\mathrm df'(1) && + \tilde d_\mathrm d^{00}=d^{00}_\mathrm df'(1) +\end{align} +Once these expressions are inserted into the complexity, the limits of $n$ and +$m$ to zero can be taken, and the parameters from $D^{01}$ and $D^{11}$ can be +extremized explicitly. \section{Details of calculation for the isolated eigenvalue} \label{sec:eigenvalue-details} -- cgit v1.2.3-70-g09d2