prompt
stringlengths
0
256
answer
stringlengths
1
256
:81a} found that such states associated with a trivial $P(q)=\delta(q)$ should not exist below a critical energy $E_c$ and for the Ising case $E_c=-0.672$. States with an energy close to $-0.73$ would be expected to be have a $P(q)$ rather simila
r to those for full replica symmetry breaking, but those generated in the quench have a trivial $P(q)$. There is no paradox as the states generated in the quench have more than one-spin flip stability \cite{yan:15}. This results in a distribu
tion of local fields behaving at small fields so that $p(h) \sim h$, very different from that expected from the study of the $p(h)$ of one-spin flip stable states \cite{roberts:81} for which $p(0)$ is finite, and instead similar to what i
s found in the true ground state -- the state which is stable against flipping an arbitrary number of spins. It is by that means that the theorem of Newman and Stein \cite{newman:99} that in a quench from a random initial state the final $P(q)$ shou
ld be trivial is realized, despite the quenched energy being in the region where one would expect the $P(q)$ of one spin flip stable states to be non-trivial. The change in the form of $p(h)$ means that the true $E_c$ is not at $-0.672$, but
instead is at least closer to $-0.73$. For the vector SK spin glasses in zero field we have studied the energy reached in a quench from infinite temperature by putting the spins parallel to their local fields. In Figs. \ref{fig:SKXYEc} and \ref{fig:SKH
eisEc} we have plotted our estimates of this energy as a function of $1/N^{2/3}$, the form commonly used for the energy size dependence of the SK model \cite{boettcher:03, billoire:08}. For $m=2$, the extrapolated energy per spin component is $\approx -0.
870$, whereas its $E_c=-0.866$ according to the analysis in Ref. \onlinecite{bm1981}; for $m=3$ the extrapolated energy per spin component is $\approx -0.915$ whereas its $E_c =-0.914$ \cite{bm1981}. Minima whose energies lie below the critical energy $E_c
$, are associated with non-trivial (i.e. RSB) form for their $P(q)$, calculated from the overlaps of the minima at the same energy \cite{bm1981, bm:81a}. We found just as for the Ising SK model that the energy reached in the quench varied little when the
greedy algorithm was used instead of the sequential algorithm \cite{parisi:95}. \begin{figure} \includegraphics[width=\columnwidth]{fig_XY_SK_energy.eps} \caption{(Color online) The average energy per site and spin component for the XY SK spin glas
s model ($m =2$) with $h_r=0$ plotted against $1/N^{2/3}$ in order to estimate the infinite system value of the energy obtained from a quench from infinite temperature. For $m=2$, $E_c=-0.866$ \cite{bm1981}.} \label{fig:SKXYEc} \end{figure} \begin{figur
e} \includegraphics[width=\columnwidth]{fig_Heisenberg_SK_energy.eps} \caption{(Color online) The average energy per site and spin component for the Heisenberg SK spin glass ($m =3$) with $h_r=0$ plotted against $1/N^{2/3}$ in order to estimate the in
finite system value of the energy obtained from a quench from infinite temperature. For $m =3$, $E_c=-0.914$ \cite{bm1981}. } \label{fig:SKHeisEc} \end{figure} As the energy of the quenched state is remarkably close to the critical energies cal
culated by Bray and Moore \cite{bm1981,bm:81a} for $m=2$ and $m=3$, this suggests that the state reached in the quench is well-described by the calculations in Ref. \onlinecite{bm1981}, whereas for the Ising case the equiv
alent calculation which enumerates the number of one-spin flip stable states does not give the resulting $p(h)$ of the quenched states with much accuracy and so does not produce an accurate estimate of $E_c$. One knows a lot about behavior at $E_c$ a
t least for Ising spins in zero random field \cite{bm:81a}. For states of energy per spin $\varepsilon > E_c$, the annealed and quenched averages agree with each other, but for energies $\varepsilon < E_c$, the two calculations differ. As $\
varepsilon$ approaches $E_c$, behavior is as at a critical point, with growing length scales etc. and massless modes \cite{bm:81a}. For the Ising case the properties of these modes were discussed in Ref. \onlinecite{bm:81a}. We intend to ret
urn to this topic in a future publication for the case of vector spin glasses. When one sets an Ising spin parallel to its local field in the course of the quench, spin avalanches may be triggered. If the number of neighbors $z$ is of order $N$ th
en the avalanches can be on all size scales \cite{boettcher:08, andresen:13}. Thus the Ising SK model is an example of a system with marginal stability as discussed by M\"{u}ller and Wyart \cite{muller:15}. It was argued in Ref.
\onlinecite{muller:15} that as the quench progresses the system will reach the marginal manifold which separates stable from unstable configurations. As this point is approached the dynamics slows and eventually freezes near the marginal manifold.
The VB model with $z =6$ does not have large scale avalanches \cite{andresen:13} and does not have any marginal features; a first study of avalanches in the undiluted one-dimensional long-range models can be found in \cite{boettcher:08}. While the
Ising VB model does not have large scale avalanches, there certainly will be an energy $E_c$ below which the minima will have non-trivial overlaps. What is not clear is whether it is the large avalanches which ensures that the states generated
in a quench are close to this energy. We also do not know what difference the existence of a finite temperature phase might make to the properties of the quenched state. For example, are there features of the quenched states of one a
nd two dimensional Ising spin glasses, where there is no finite temperature spin glass transition, which differ significantly from those of the three dimensional spin glass, where there is a finite temperature phase transition? We also do no
t know what features might arise if there is a phase transition to a state with full replica symmetry breaking, as opposed to a state with just replica symmetry. For systems for which the excitations are not discrete, such as in vector spin glas
ses, marginality takes a different form, and seems related to the development of negative eigenvalues in the Hessian \cite{muller:15,sharma2014avalanches}. Such eigenvalue instabilities might be triggered in a quench where one puts spins paralle
l to their local fields. On the other hand, one could imagine a steepest descent procedure starting from the initial spin orientation and smoothly proceeding to a minimum. Does that result in a final state whose properties differ from those ge
nerated by putting spins parallel to their local fields? There are many topics which should be studied! We believe that the proximity of the quenched energy to the calculated critical energy $E_c$, at least for the cases of $m =2$ and $m =3$ will prov
ide valuable analytical insights concerning marginal stability. One of our motivations for the analytic work in the next section was to calculate $E_c(h_r)$ in the presence of a non-zero random vector field, but, as we shall see, algebraic difficulties pre
vented us from achieving this goal. But it would be good to know how general is the result that the energy obtained in a quench coincides with the energy at which the overlaps of the minima display replica symmetry breaking features. \section{Metastable
states in the SK model in the presence of a random field} \label{sec:SKanalytic} In this section we follow the method of Ref. \onlinecite{bm1981} to study the complexity and Hessian properties of the minima for the SK model but in the presence of a rando
m vector field. We begin by writing down the first steps in the formalism following Ref. \onlinecite{bm1981}. In subsection \ref{SKannealed} we show that within the annealed approximation, where one averages $N_S(\varepsilon)$ itself over the bonds $J_{ij
}$ and the random fields $h_i^{ex}$ analytical progress is fairly straightforward. Fortunately the annealed approximation is also exact for fields $h_r > h_{AT}$. In subsection \ref{sec:quenched} we describe our attempts to solve the quenched case. We beli
eve that our approach based on replica symmetry assumptions should be good down to its limit of stability which would be at $E_c(h_r)$, but algebraic difficulties prevented us from actually determining $E_c(h_r)$. We find it convenient to write the Ha
miltonian for the $m$-vector spin glass in an $m$-component external field as \begin{equation} \mathcal{H}=-\frac{m}{2}\sum_{i,j} J_{ij}\bm{S}_i\cdot\bm{S}_j-m\sum_i\bm{h}^{\rm ex}_i\cdot\bm{S}_i, \label{Hamil} \end{equation} where the $m$-component spi
ns $\bm{S}_i=\{S^\alpha_i\}$, ($\alpha=1,\cdots,m$, $i=1,\cdots,N$) have a unit length $S_i=1$. The interactions $J_{ij}$ are chosen from a Gaussian distribution with zero mean and the variance $1/N$. In this section, for convenience, we use the notation
$\bm{h}^{\rm ex}_i=\bm{h}_i/\sqrt{m}$ for the random Gaussian external fields with zero mean and the variance \begin{equation} \langle h^{{\rm ex},\alpha}_i h^{{\rm ex},\beta}_j\rangle =\frac{h^2_r}{m}\delta_{ij}\delta^{\alpha\beta}. \end{equation} At
zero temperature, the spins are aligned in the direction of the local internal field $\bm{H}_i$, i.e. \begin{equation} \bm{S}_i=\hat{\bm{H}}_i\equiv\frac{\bm{H}_i}{H_i}, \label{eq:par} \end{equation} where \begin{equation} \bm{H}_i=\sum_j J_{ij}\bm{S}_j
+\bm{h}^{\rm ex}_i. \label{eq:hdef} \end{equation} In terms of the local fields, the ground state energy $E$ can be written as \begin{equation} E=-\frac{m}{2}\sum_i(H_i+\hat{\bm{H}}_i\cdot \bm{h}^{\rm ex}_i). \end{equation} The number of metastable stat
es with energy $\varepsilon$ per site and per spin component is given by \begin{align} &N_S(\varepsilon)=\int\prod_{i,\alpha}dH_i^{\alpha}\int\prod_{i,\alpha}dS_i^{\alpha} \prod_{i,\alpha}\delta(S_i^{\alpha}-\hat{H}_i^{\alpha}) \nonumber \\ &~~~~~~\times
\prod_{i,\alpha}\delta\left(H_i^{\alpha}-\sum_j J_{ij}S_j^{\alpha}-h^{{\rm ex},\alpha}_{i}\right) |\det\mathsf{M}\{J_{ij}\}| \nonumber \\ &~~~~~~\times \delta\left(Nm\varepsilon+\frac 1 2 m \sum_i (H_i + \hat{\bm{H}}_i\cdot\bm{h}^{\rm ex}_i ) \right),
\label{NS} \end{align} where \begin{equation} M^{\alpha\beta}_{ij}=\frac{\partial}{\partial S^\beta_j}(S^\alpha_i-\hat{H}^\alpha_i) =\delta_{ij}\delta^{\alpha\beta}-J_{ij}\frac{P^{\alpha\beta}_i}{H_i} \end{equation} with $P^{\alpha\beta}_i\equiv \delta
^{\alpha\beta}-\hat{H}^\alpha_i\hat{H}^\beta_i$ is the projection matrix. \subsection{Annealed Approximation} \label{SKannealed} We now calculate the average of $N_S(\varepsilon)$ over the random couplings and the random external fields. As we will see b
elow, the direct evaluation of the quenched average $\langle \ln N_S(\varepsilon) \rangle$ is very complicated. Here we first present the annealed approximation, where we evaluate the annealed complexity $g_A(\varepsilon)=\ln \langle N_S(\varepsilon) \ran
gle/N$. The whole calculation is very similar to those in Appendix 2 of Ref.~\onlinecite{bm1981} except for the part involving the average over the random field. Below we sketch the calculation. The first delta functions in Eq.~(\ref{NS}) can be integrat
ed away. We use the integral representations for the second and third delta functions using the variables $x_i^{\alpha}$ and $u$, respectively, along the imaginary axis. The average over the random couplings can be done in an exactly the same way as in Re
f.~\onlinecite{bm1981}. We briefly summarize the results below. The random couplings appear in the factor \begin{align} & \left\langle \exp\Big[-\sum_{i<j}J_{ij}\sum_{i,\alpha} (x^\alpha_{i}\hat{H}^\alpha_{j} +x^\alpha_{j}\hat{H}^\alpha_{i})\Big] |\det
\mathsf{M} \{J_{ij}\}| \right\rangle_J \nonumber \\ =&\exp \Big[ \frac 1 {2N} \sum_{i<j}\Big\{ \sum_{\alpha}(x^\alpha_{i}\hat{H}^\alpha_{j} +x^\alpha_{j}\hat{H}^\alpha_{i}) \Big\}^2\Big] \nonumber \\ &~~~~~~~~~~~~~\times \left\langle |\det\mathsf{M} \{
J_{ij}-O(\frac 1 N)\}| \right\rangle_J . \end{align} After neglecting the $O(1/N)$ term, we evaluate the average of the determinant as \cite{bm1981} \begin{equation} \left\langle |\det\mathsf{M} \{J_{ij}\}| \right\rangle_J=\exp(\frac 1 2 Nm\bar{\chi})
\prod_i \left(1-\frac{\bar{\chi}}{H_i}\right)^{m-1}, \label{det} \end{equation} where the susceptibility $\bar{\chi}$ satisfies the self-consistency equation \cite{bm1981} \begin{equation} \bar{\chi}=(1-\frac 1 m)\frac 1 N \sum_i \frac 1 {H_i-\bar{\chi}} \
label{chi} \end{equation} with the condition $H_i\ge\bar{\chi}$. Using the rotational invariance and the Hubbard-Stratonovich transformation, we can rewrite the exponential factor in front of the determinant as \begin{align} &\exp[\frac 1{2m}\sum_{i,\alp
ha}(x^\alpha_{i})^2 ] \\ &~\times \int\frac{dv}{(2\pi/Nm)^{1/2}}\; \exp [ -\frac{Nm}{2} v^2+ v \sum_{i,\alpha}x^\alpha_{i}\hat{H}^\alpha_{i} ]. \nonumber \end{align} In the present case, we have to average over the random field. Collecting the rele
vant terms, we have \begin{align} & \left\langle \exp\Big[-\sum_{i,\alpha} (x^\alpha_{i}+\frac 1 2 u m \hat{H}^\alpha_{i})h^{{\rm ex},\alpha}_{i}\Big] \right\rangle_{\bm{h}^{\rm ex}} \\ =&\exp\Big[ \frac{h^2_r}{2m}\sum_{i,\alpha}(x^\alpha_{i})^2 +\fra
c{h^2_r}{2} u \sum_{i,\alpha}x^\alpha_{i}\hat{H}^\alpha_{i} +Nm\frac{h^2_r }{8}u^2 \Big] . \nonumber \end{align} All the site indices are now decoupled. We express the condition Eq.~(\ref{chi}) using the integral representation of the delta function
with the variable $\lambda$ running along the imaginary axis. Putting all the terms together, we have \begin{align} &\langle [N_S(\varepsilon)] \rangle_{J,h^{\rm ex}}= \int \frac{du}{2\pi i} \int \frac{dv}{\sqrt{2\pi/Nm}} \int d\bar{\chi} \int \frac
{d\lambda}{2\pi i} \nonumber \\ & \times \exp\Big[ Nm \lambda \bar{\chi} + \frac{Nm}{2} \bar{\chi}^2 -Nm\varepsilon u -\frac{Nm}{2} v^2 \nonumber \\ &~~~~~~~~~~~~~~~~~~~~~~~+Nm\frac{h^2_r}{8}u^2 +N\ln I^\prime \Big], \label{NS_ann} \end{align} wh
ere \begin{align} I^\prime=&\int_{H\ge\bar{\chi}}\prod_{\alpha}dH^\alpha \int\prod_{\alpha}\frac{dx^\alpha}{2\pi i} \left(1-\frac{\bar{\chi}}{H}\right)^{m-1} \nonumber \\ &\times \exp\Bigg[ \frac{1+h^2_r}{2m}\sum_{\alpha}(x^\alpha)^2 + (v+\frac{h^2_
r}{2}u)\sum_\alpha x^\alpha\hat{H}^\alpha \nonumber \\ & +\sum_{\alpha}x^\alpha H^\alpha -(m-1) \lambda (H-\bar{\chi})^{-1} -\frac{m}{2} u H \Bigg] \end{align} The Gaussian integral over $x^\alpha$ can be done analytically. The integrals i
n Eq.~(\ref{NS_ann}) are evaluated via the saddle point method in the $N\to\infty$ limit. Following the procedure described in Ref.~\onlinecite{bm1981}, we introduce new variables $\bm{h}\equiv (H-\bar{\chi})\hat{\bm{H}}$ and $\Delta=-v-\bar{\chi}$ and us
e the saddle point condition for $\bar{\chi}$, which is \begin{equation} \lambda-\Delta-\frac u 2=0. \end{equation} We finally have an expression for the annealed complexity $g_A(\varepsilon)\equiv N^{-1}\ln\langle N_S(\varepsilon)\rangle$ as \begin{equa
tion} g_A(\varepsilon) = m (-\frac{\Delta^2}{2} -\varepsilon u +\frac{h^2_r}{8}u^2) +\ln I, \end{equation} where \begin{align} &I= \left(\frac{m}{2\pi(1+h^2_r)}\right)^{m/2}S_m\int^\infty_0 dh\; h^{m-1} \\ &\times\exp\Big[ -\frac{m}{2(1+h^2
_r)}(h-\Delta+\frac{h^2_r}{2}u)^2 \nonumber \\ &~~~~~~~~~~~~~~~ -\frac{(m-1)}{h}(\Delta+\frac u 2) -\frac{m}{2}uh\Big] \nonumber \end{align} with the surface area of the $m$-dimensional unit sphere $S_m=2\pi^{m/2}/\Gamma(m/2)$. The parameters $\Delta$ an
d $u$ are determined variationally as $\partial g_A/\partial \Delta =\partial g_A/\partial u =0$. We focus on the total number of metastable states, which are obtained by integrating $\exp(Ng_A(\varepsilon))$ over $\varepsilon$, or equivalently by setti
ng $u=0$. Thus we are effectively focussing on the most numerous states, those at the top of the band where $g_A(\varepsilon)$ is largest. In this case, $g_A=-(m/2)\Delta^2+\ln I_0$, where \begin{align} I_0 =& S_m \left(\frac{m}{2\pi(1+h^2_r)}\right)^{m/
2}\int_0^\infty dh\; h^{m-1} \nonumber \\ &\times \exp\left[ -(m-1)\frac{\Delta}{h}-\frac{m(h-\Delta)^2}{2(1+h^2_r)}\right] \label{I0}. \end{align} The parameter $\Delta$ is determined by the saddle point equation \begin{equation} \Delta=\frac{1}{2+h^2_r
}\langle h \rangle - \left(1-\frac{1}{m}\right)\left(\frac{1+h^2_r}{2+h^2_r}\right) \left\langle\frac{1}{h}\right\rangle, \label{Delta} \end{equation} where the average is calculated with respect to the probability distribution for the internal field gi
ven by the integrand of $I_0$ in Eq.~(\ref{I0}). Using $\langle h \rangle =\Delta + \langle h-\Delta \rangle$, we can rewrite Eq.~(\ref{I0}) as \begin{equation} \Delta \left[ 1- \left(1-\frac 1 m \right)\left\langle \frac 1 {h^2} \right\rangle \right]=0.
\label{Delta1} \end{equation} For various values of the external field $h_r$, we solve numerically Eq.~(\ref{Delta}). For $m=3$, we find that when $h_r>h_{AT}=1$ there is only a trivial solution, $\Delta=0$. (Note that the Almeida-Thouless field $h_{AT}
$ at $T=0$ is $h_{AT}=1/\sqrt{m-2}$ ~\cite{sharma2010almeida}). From Eq.~(\ref{I0}), we see that in this case $I_0=1$ and the complexity $g$ vanishes above the AT field. For $h_r<h_{AT}$, a nontrivial solution, $\Delta\neq 0$ exists. We find that the val
ues of $\Delta$ and $g_A$ increase as the external field $h_r$ decreases from $h_{AT}$, and approach the known values, 0.170 and 0.00839 at zero external field \cite{bm1981}. For $h_r$ smaller than but very close to $h_{AT}$, $\Delta$ is very small. We ma
y obtain an analytic expression for $g_A$ in this case. By expanding everything in Eq.~(\ref{Delta1}) in powers of $\Delta$, we find for $m=3$ that \begin{equation} g_A=\frac 3 2 (h^2_{AT}-h^2_r)\tilde{\Delta}^2+8\sqrt{\frac{3}{2\pi}} \tilde{\Delta}^3\
ln\tilde{\Delta}+O(\tilde{\Delta}^3), \label{gA0} \end{equation} where $\tilde{\Delta}=\Delta/\sqrt{1+h^2_r}$. The fact that $g_A$ must be stationary with respect to $\tilde{\Delta}$, enables one to determine how the complexity vanishes as $h_r \to h_{AT}
$ and the value of $\tilde{\Delta}$ in this limit. Using the distribution for the internal field $H$ (or $h$), we first calculate the spin glass susceptibility $\chi_{SG}\equiv (Nm)^{-1}\mathrm{Tr} \mathbf{\chi}^2$ with the susceptibility matrix $\mathb
f{\chi}=\chi_{ij}^{\alpha\beta}$ \cite{bm1981}. Note that the susceptibility in Eq.~(\ref{chi}) is just $\bar{\chi}=(Nm)^{-1}\mathrm{Tr}\mathbf{\chi}$. The spin glass susceptibility is given by \cite{bm1981} $\chi_{SG}=(1-\lambda_R)/\lambda_R$, where \be
gin{equation} \lambda_R=1-(1-\frac 1 m)\frac 1 N \sum_i \frac 1{(H_i-\bar{\chi})^2}. \label{lambda_R} \end{equation} This quantity is exactly the one in the square bracket in Eq.~(\ref{Delta1}). Therefore, since $\Delta\neq 0$ for $h_r<h_{AT}$, $\lambda_R
$ vanishes and consequently $\chi_{SG}$ diverges. Above the AT field, there is only a trivial solution $\Delta=0$. In this case the integrals are just Gaussians and we can evaluate explicitly $\frac{1}{N} \sum_i \frac 1{(H_i-\bar{\chi})^2}$, with the resul
t that $\lambda_R=(h_r^2-1/(m-2))/(1+h_r^2)$, so the spin glass susceptibility as a function of the external random field for $h_r>h_{AT}$ is given by \begin{equation} \chi_{SG}=\frac{1+h_{AT}^2}{h_r^2-h_{AT}^2}, \label{chisgexact} \end{equation} provid
ed $h_r > h_{AT}$ and $m > 2$. The simple divergence of $\chi_{SG}$ as $h_r \to h_{AT}$ is a feature of the SK limit and is not found in the Viana-Bray model at least amongst the quenched states of our numerical studies, see Sec. \ref{sec:spinglasssuscepti
bility} We now calculate the eigenvalue spectrum of the Hessian matrix $\mathsf{A}$. The calculation closely follows the steps in Ref.~\onlinecite{bm1982} for the case of zero external field. We consider (transverse) fluctuations around the $T=0$ sol
ution $\bm{S}^{0}_i\equiv\hat{\bm{H}}_i$ by writing $\bm{S}_i=\bm{S}^{0}_i+\bm{\epsilon}_i$, where $\bm{\epsilon}_i=\sum_\alpha \epsilon^\alpha_i \hat{\bm{e}}_\alpha(i)$ with the $(m-1)$ orthonormal vectors $\hat{e}_\alpha(i)$, $\alpha=1,\cdots,m-1$ satisf
ying $\bm{S}^{0}_i\cdot \hat{\bm{e}}_\alpha(i)=0$. Inserting this into Eq.~(\ref{Hamil}), we have the Hessian matrix as \begin{equation} A^{\alpha\beta}_{ij}\equiv\frac{\partial(\mathcal{H}/m)}{\partial\epsilon^\alpha_i \partial\epsilon^\beta_j} =H_i\delta
_{ij}\delta^{\alpha\beta}-J_{ij} \hat{\bm{e}}_\alpha(i)\cdot\hat{\bm{e}}_\beta(j). \end{equation} The eigenvalue spectrum $\rho(\lambda)$ can be calculated from the resolvent $\mathsf{G}=(\lambda\mathsf{I}-\mathsf{A})^{-1}$ as \begin{equation} \rho(\lambd
a)=\frac{1}{N(m-1)\pi}\mathrm{Im} \; \mathrm{Tr} \mathsf{G}(\lambda-i\delta), \label{eqn:rho} \end{equation} where $\mathsf{I}$ is the $(m-1)N$-dimensional unit matrix and $\delta$ is an infinitesimal positive number. The locator expansion method \cite{bm1
979} is used to evaluate $\rho(\lambda)$, which yields the following self-consistent equation for $\bar{G}(\lambda)\equiv ((m-1)N)^{-1} \mathrm{Tr} \mathsf{G}(\lambda)$: \begin{equation} \bar{G}(\lambda)=\left\langle \frac 1{\lambda-H-(1-\frac 1 m)\bar{G}
(\lambda)}\right\rangle, \label{Gbar} \end{equation} where $\langle ~\rangle$ denotes the average over the distribution for $h$ given in the integrand in Eq.~(\ref{I0}). Note that $H=h+\bar{\chi}$ and $\bar{\chi}=(1-1/m)\langle 1/h\rangle$ from Eq.~(\ref{
chi}). We first separate $\bar{G}=\bar{G}^\prime+i\bar{G}^{\prime\prime}$ into real and imaginary parts and solve Eq.~(\ref{Gbar}) numerically for $\bar{G}^\prime(\lambda)$ and $\bar{G}^{\prime\prime}(\lambda)$ as a function of $\lambda$. The eigenvalue
spectrum is just $\rho(\lambda)=\pi^{-1}\bar{G}^{\prime\prime}(\lambda)$. \begin{figure} \includegraphics[width=\columnwidth]{rho_SK.eps} \caption{(Color online) The eigenvalue spectrum of the Hessian at zero temperature for the vector spin glass
with $m=3$ in the SK limit.The various lines correspond to different values of $h_r$, the external random field.} \label{fig:rho} \end{figure} \vspace{0.2cm} \begin{figure} \includegraphics[width=\columnwidth]{rho_mag_SK.eps} \caption{(Color o
nline) The magnified view of the same figure as Fig.~\ref{fig:rho} but for the small eigenvalues.} \label{fig:rho_mag} \end{figure} As we can see from Figs.~\ref{fig:rho} and \ref{fig:rho_mag}, $\rho(\lambda)$ does not change very much as we increase $
h_r$ from zero up to $h_{AT}=1$. For the external field larger than the AT field, however, Fig.~\ref{fig:rho_mag}\ clearly shows that the eigenvalue spectrum develops a gap. The gap increases with the increasing external field. By directly working on Eq.~
(\ref{Gbar}) in the small-$\lambda$ limit, we find that for small eigenvalues \begin{equation} \rho(\lambda)\simeq \frac{1}{\pi(1-1/m)}\frac{1}{\sqrt{s}}\sqrt{\lambda-\lambda_0}, \end{equation} where $s=(1-m^{-1})\langle 1/h^3 \rangle$ and $\lambda_0=\la
mbda^2_R/4s$ with $\lambda_R$ defined in Eq.~(\ref{lambda_R}). Our numerical solution of the equations for $G(\lambda)$ confirms that there is no gap below $h_{AT}$ which is consistent with the previous observation that $\lambda_R$ vanishes there. However,
the integral by which $s$ is defined diverges for $h_r > h_{AT}$ when $m <3$ and we no longer see a square root singularity at the band-edge. In the case of $m =3$ our numerical solution shown in Fig. \ref{fig:rho_mag} suggests instead of the square root
dependence there is a roughly linear dependence as $\lambda$ approaches the numerically determined band-edge $\lambda_0$, but unfortunately we have not been able to derive its form analytically. Fig. \ref{fig:rho} shows that away from $\lambda_0$ the dens
ity of states is rather as if it had the square root form. As $h_r \to h_{AT}$ this square root form works all the way to zero. \subsection{Quenched Average} \label{sec:quenched} In this subsection, we attempt to evaluate the quenched complexity $g(\v
arepsilon)=N^{-1}\langle\ln N_S(\varepsilon)\rangle$. The calculations are quite complicated and some of the details are sketched in the Appendix. In order to calculate $\langle \ln N_S(\varepsilon) \rangle$, we consider an aver
age of the replicated quantity $\langle [N_S(\varepsilon)]^n \rangle_{J,h^{\rm ex}}$. We then have an expression similar to Eq.~(\ref{NS_ann}), where the integrals are now over replicated variables, $u^\eta$, $v^\eta$, $\bar{\chi}^\eta$
and $\lambda^\eta$ with the replica indices $\eta,\mu=1,\cdots,n$. In addition to these, the expression also involves the integrals over the variables carrying off-diagonal replica indices, which are denoted by $A_{\eta\nu}$, $A^*_{\et
a\nu}$, $B_{\eta\nu}$ and $B^*_{\eta\nu}$ with $\eta<\nu$. In the absence of external field, it can be shown \cite{bm1981} that $A_{\eta\nu}=A^*_{\eta\nu} =B_{\eta\nu}=B^*_{\eta\nu}=0$ is always a solution to the saddle point equati
ons. It is shown to be stable for $\varepsilon> E_c$ for the $E_c$, for which the quenched average coincides with the annealed one. For $h_r\neq 0$, however, we find that this is no longer the case. $A_{\eta\nu}=A^*_{\eta\nu}=
B_{\eta\nu}=B^*_{\eta\nu}=0$ is not a solution to saddle point equations. The saddle point solutions involve nonvanishing off-diagonal variables in replica indices. We find that in general the saddle point equations are too complicated to
allow explicit solutions. (See the Appendix for details.) The quenched average is different from the annealed one for a finite external field when $h_r< h_{AT}$. When $h_r> h_{AT}$ the annealed and quenched averages are identi
cal in every way for the SK model, which has vanishing complexity in this region. We doubt whether the same statement is true for any model such as the Viana-Bray model which has non-zero complexity for $h_r > h_{AT}$. We also do not kno
w for sure whether our replica symmetric solution for $A_{\eta \nu}$ etc. is stable. It is possible that even at $u=0$ there is a need to go to full replica symmetry breaking. Unfortunately algebraic complexities have prevented us from even findin
g a solution of the replica symmetric equations, so determining their stability looks very challenging. However, the results of the numerical work reported on the form of $P(q)$ in Sec. \ref{sec:metastability} for the Viana-Bray model in a field sugges
ts that the states reached in the quench have replica symmetry. We look for the saddle points in the replica symmetric form, \begin{align} &A_{\eta\nu}=A,~~A^*_{\eta\nu}=A^*,~~B_{\eta\nu}=B^*_{\eta\nu}=B, \nonumber \\ &u^\eta=u,~~v^\eta=v,~~\bar{\chi}^\et
a=\bar{\chi},~~\lambda^\eta=\lambda. \end{align} After a lengthy calculation (see Appendix), we arrive at the expression for the quenched complexity as follows. \begin{align} g(\varepsilon)=& m \Big\{ -\frac{\Delta^2}{2} -\varepsilon u - \frac{A}{2m} +\f
rac{1}{2} (AA^*+B^2)\Big\} \label{geps} \\ +& \int \frac{d^m \bm{w}}{(2\pi)^{m/2}} \int \frac{d^m \bm{y}}{(2\pi)^{m/2}} \int \frac{d^m \bm{z} d^m\bm{z}^*}{(2\pi)^m}\; \nonumber \\ &\times \exp[-\frac 1 2 \sum^m_\alpha (w^2_\alpha+y^2_\alpha+z_\alpha z^*_
\alpha)] \; \ln K(\bm{w},\bm{y},\bm{z},\bm{z}^*), \nonumber \end{align} where \begin{align} K=& \int d^m\bm{h} \int^{i\infty}_{-i\infty} \frac{d^m \bm{x}}{2\pi i} \;\exp\Bigg[ \frac{1-mA^*}{2m}\bm{x}^2 \nonumber \\ &+ (h-\Delta-B)\bm{x}\cdot\hat{\b
m{h}} -(m-1)\frac{\Delta + u/2}{h} -\frac{m}{2}uh \nonumber \\ &+ \sqrt{A^*+\frac{h^2_r}{m}} \;\bm{w}\cdot\bm{x} +\sqrt{A+\frac{mh^2_r}{4}u^2 } \; \bm{y}\cdot \hat{\bm{h}} \nonumber \\ &+ \sqrt{\frac{1}{2}(B+\frac{h^2_r}{2}u) } \left(\bm{z}\cdot\bm
{x}+\bm{z}^* \cdot \hat{\bm{h}}\right) \Bigg]. \end{align} All the parameters, $\Delta$, $A$, $A^*$, $B$ and $u$ are to be determined in a variational way. We found, however, that it is very difficult to solve the saddle point equations and obtain the qu
enched complexity, even numerically. For the total number of metastable states, $u=0$, we can find a simple solution to saddle point equations at $\Delta=A=B=0$ and $A^*=1/m$. In this case, $K=1$ and the complexity $g$ vanishes. This solution must desc
ribe the case where $h_r>h_{AT}$ and it is identical to the annealed average. For the external field $h_r$ just below $h_{AT}$, $\Delta$, $A$, $B$ and $C\equiv 1/m-A^*$ are expected to be very small, and we may expand the integrals in Eq.~(\ref{geps}) in t
hese variables. We find after a very lengthy calculation that \begin{equation} g\simeq \frac{m}{1+h^2_r} (h^2_{AT}-h^2_r)\Big[ \frac{\Delta^2}{2} + \frac{AC}{2} - \frac{B^2}{2} \Big] . \end{equation} Note that from Eq.~(\ref{saddle}), we expect $B$
is pure imaginary. In order to determine how these variables behave near $h_{AT}$, we need higher order terms. Unfortunately, the complicated nature of these equations, however, has prevented us from going beyond the quadratic orders. It seems natural t
o expect that the $\Delta$ sector is decoupled from the off-diagonal variables, and so will have the same $\Delta^3\ln\Delta$ behavior as in Eq.~(\ref{gA0}). But the effort to obtain a full solution is so large that we abandoned further work on it. \se
ction{Hessian studies} \label{sec:hessian} In this section we write down the Hessian for the $m=3$ Heisenberg spin glass in a form which is convenient for numerical work. The Hessian is of interest as it describes the nature of the energy of the spin glas
s in the vicinity of the minima. It is also closely related to the matrices needed to describe the spin waves in the system \cite{bm1981}. We follow the approach used in the paper of Beton and Moore~\cite{beton1984electron} to find the elements of the Hes
sian matrix $T$ corresponding to directions transverse to each spin subject to the above metastability condition. We first define the site-dependent two-dimensional orthogonal unit vectors $\hat{e}_{x}(i)$ and $\hat{e}_{y}(i)$ such that \begin{align} \hat{
e}_{m}(i)\cdot\mathbf{S}_{i}^{0} &= 0\\ \hat{e}_{m}(i)\cdot\hat{e}_{n}(i) &= \delta^{mn}, \end{align} where $m,n = x,y$ denotes the directions perpendicular to the spin at the $i$th site, which is deemed in the $\lq \lq z"$ direction. The linear combinatio
ns $\hat{e}_{i}^{\pm}=\frac{1}{\sqrt{2}}(\hat{e}_{x}(i)\pm i\hat{e}_{y}(i))$ turn out to be particularly useful. Expanding $\mathbf{S}_{i}$ about $\mathbf{S}_{i}^{0}$, subject to the condition that the length of the spins remains unchanged yields, upto sec
ond-order: \begin{align} \mathbf{S}_{i} = \mathbf{S}_{i}^{0}+\Gamma_{i}^{x} \hat{e}_{x}(i)+\Gamma_{i}^{y} \hat{e}_{y}(i)-\frac{1}{2}[(\Gamma_{i}^{x})^{2}+(\Gamma_{i}^{y})^{2}]\mathbf{S}_{i}^{0}. \end{align} Equivalently, \begin{align} \mathbf{S}_{i} = \ma
thbf{S}_{i}^{0}+Z_{i}^{-}\hat{e}_{i}^{+}+Z_{i}^{+}\hat{e}_{i}^{-}-Z_{i}^{-}Z_{i}^{+}\mathbf{S}_{i}^{0}, \end{align} where $Z_{i}^{\pm} = \frac{1}{\sqrt{2}}(\Gamma_{i}^{x}\pm i\Gamma_{i}^{y})$, and $(Z_{i}^{+})^{*}=Z_{i}^{-}$. Defining the $2N$-dimensional
vector \begin{align} |Z\rangle = \begin{pmatrix} Z_{i}^{-}\\ Z_{i}^{+} \end{pmatrix}, \end{align} the change in energy per spin component degree of freedom $\frac{\delta E}{3}$ due to a change in spin orientations $|Z\rangle$, is given by: \begin{align} \
frac{\delta E}{3} = \frac{1}{2}\langle Z|T|Z\rangle, \end{align} where $T$ is the $2N \times 2N$ Hessian matrix given by \begin{align*} \begin{aligned} T = \frac{1}{3} \begin{pmatrix} |\mathbf{H}_{i}|\delta_{ij}+A_{ij}^{*} & B_{ij}^{*}\\ B_{ij} & |\m
athbf{H}_{i}|\delta_{ij}+A_{ij} \end{pmatrix} \end{aligned}, \end{align*} where the matrix elements are \begin{align*} A_{ij} = A_{ji}^{*} = -3J_{ij}\hat{e}_{i}^{+}\cdot\hat{e}_{j}^{-}\\ B_{ij} = B_{ji}^{*} = -3J_{ij}\hat{e}_{i}^{+}\cdot\hat{e}_{j}^{+}. \
end{align*} Converting to spherical coordinates, the matrix elements are \begin{widetext} \begin{align} A_{ij}^{*} &= - \frac{3J_{ij}}{2}[(\cos(\theta_{i})\cos(\theta_{j})+1)\cos(\phi_{i}-\phi_{j})+i(\cos(\theta_{i})+\cos(\theta_{j}))\sin(\phi_{i}-\phi_{
j})+\sin(\theta_{i})\sin(\theta_{j})]\nonumber\nonumber\\ B_{ij}^{*} &= - \frac{3J_{ij}}{2}[(\cos(\theta_{i})\cos(\theta_{j})-1)\cos(\phi_{i}-\phi_{j})-i(\cos(\theta_{i})-\cos(\theta_{j}))\sin(\phi_{i}-\phi_{j})+\sin(\theta_{i})\sin(\theta_{j})]\nonumber\\
B_{ij} &= - \frac{3J_{ij}}{2}[(\cos(\theta_{i})\cos(\theta_{j})-1)\cos(\phi_{i}-\phi_{j})+i(\cos(\theta_{i})-\cos(\theta_{j}))\sin(\phi_{i}-\phi_{j})+\sin(\theta_{i})\sin(\theta_{j})]\nonumber\\ A_{ij} &= - \frac{3J_{ij}}{2}[(\cos(\theta_{i})\cos(\theta_{
j})+1)\cos(\phi_{i}-\phi_{j})-i(\cos(\theta_{i})+\cos(\theta_{j}))\sin(\phi_{i}-\phi_{j})+\sin(\theta_{i})\sin(\theta_{j})]\nonumber\\ \end{align}. \end{widetext} \begin{figure} \includegraphics[width=\columnwidth]{fig_SKchi.eps} \caption{(Color onl
ine) The inverse of the spin glass susceptibility $\chi_{SG}^{-1}$ versus $h_r^2$ for a range of system sizes of the Heisenberg SK model. The analytic curve is the result of Eq. (\ref{chisgexact}). For $h_r \le 1$, one expects that $\chi_{SG}^{-1} =0$, but
finite size effects make it non-zero.} \label{fig1SK} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{fig_VBchi.eps} \caption{(Color online) The inverse of the spin glass susceptibility $\chi_{SG}^{-1}$ versus $h_r^2$ for a range
of system sizes for the VB model with $z = 6$.} \label{fig1VB} \end{figure} In the next subsection we use the Hessian to numerically calculate the spin glass susceptibility of both the SK model and VB model in a range of random fields for the Heisenberg
spin glass. \subsection{Spin Glass Susceptibility} \label{sec:spinglasssusceptibility} The spin glass susceptibility for the metastable states can be computed from the inverse of the Hessian matrix using the relation~\cite{bm1981} \begin{equation} \chi
_{SG} = \frac{1}{N}\Tr{(T^{-1})^{2}}. \end{equation} For the SK model and $h_r> h_{AT}=1$, we have calculated $\chi_{SG}$ analytically and Fig. \ref{fig1SK} shows that our numerical work is approaching the analytical solution, but finite size effects are s
till very considerable at the sizes we can study. Notice that for the SK model there is (weak) numerical evidence that $\chi_{SG}$ diverges below the AT field. For the VB model, the plot of $\chi_{SG}$ in Fig. \ref{fig1VB} obtained from our metastable stat
es which lie above the true ground state energy provides no evidence that an AT field has much relevance for these states. \begin{figure} \includegraphics[width=\columnwidth]{fig_dos.eps} \caption{(Color online) The averaged density of states of the
Hessian matrix of the metastable states obtained after a quench to $T=0$ starting from spins with random orientations i.e. $T= \infty$ for the SK model ($\sigma =0, z =N-1$ of the diluted model). Data shown here for the special case of $h_{r} = 0.8$, for w
hich the system is in the spin glass phase, just below $h_{AT} = 1$. The analytical curve is that calculated from Eqs. (\ref{eqn:rho}) and (\ref{Gbar})) for metastable states at the top of the band within the annealed approximation. The numerical results a
re strikingly similar to the analytical results, despite the fact that they refer to Hessians for quite different situations! } \label{fig2} \end{figure} \subsection{Density of States} \label{sec:density} The density of states of the eigenvalues of t
he Hessian matrix has been obtained numerically for the minima obtained in a quench from infinite temperature to zero temperature. The results have remarkable agreement with the analytical calculation performed on the Heisenberg SK model as shown in Fig.~
\ref{fig2}. The analytical calculation itself is not for the same set of metastable states. It applies to the states corresponding to $u=0$ (i.e. those with the largest complexity within the annealed approximation). In Fig.~\ref{fig2}, data are shown for
$h_r =0.8 h_{AT}$, where no gap is present. The agreement between the analytical curve which is obtained for the thermodynamic limit, and the data for a $N=1024$ size system from numerical simulations, is striking. Notice that the $\sqrt{\lambda}$ form p
redicted from the annealed study (see Sec. \ref{SKannealed}) seems to hold as $\lambda \to 0$, despite there being no Goldstone theorem in the presence of a random field to ensure the existence of massless modes. We have also studied the density of states
and quantities like the inverse participation ratios for the quenched state minima in models like the VB model and the one-dimensional long range models. Basically the results seem similar to those reported in Refs. \onlinecite{baity2015soft} for the thre
e dimensional Heisenberg spin glass model in a random field. But it requires large systems to get accurate results for the density of states at small values of $\lambda$ and we are leaving these issues to a future publication. \section{Conclusions} \labe
l{sec:conclusions} We believe that the most interesting feature which has turned up in our studies is the discovery for the SK model in zero external fields that the quenched states reached for $m=2$ and $m =3$ are quite close to the critical energies $E_c
$ at which the overlap of the states would acquire features associated with a $P(q)$ with broken replica symmetry. In the Ising SK model the local fields after the quench are so different from those used in the analytical calculations of $E_c$ that the con
nection of the quenched state to being just at the edge of the states with broken replica symmetry was not recognized. Thus in systems with marginal stability this means that features normally associated with continuous phase transitions, in particular div
erging length scales, could be studied as in Ref. \onlinecite{bm:81a}. We have noticed too that the energy of the states reached from the quench have zero overlap with each other. This behavior was predicted for the Ising case in Ref.~\onlinecite{newman:
99} by Newman and Stein who proved that after a quench from infinite temperature for Ising systems the states which are reached have a characteristic energy and a trivial $P(q)$. It would be good to extend their theorems to vector spin systems both in zer
o field and also in the presence of random fields. In Sec. \ref{sec:SKanalytic} we attempted to extend the old calculations of Bray and Moore \cite{bm1981} which were for zero random field to non-zero random fields. For fields $h_r >h_{AT}$ where the comp
lexity is zero, the annealed approximation is exact and we were able to obtain the exact form for the behaviour of the density of states of the Hessian matrix. There was found to be a gap in the spectrum which went to zero in the limit $h_r \to h_{AT}$. Wh
en $h_r< h_{AT}$ one needs to study the quenched average in order to get results pertinent to typical minima, but we were not able to overcome the algebraic complexities (see Sec. \ref{sec:quenched} and the Appendix), although the only difficulty is that o
f solving the equations which we have obtained. If that could be done then one could investigate the limit of stability of the replica symmetric solution and determine $E_c(h_r)$. Then one could investigate whether a quench in a field $h_r$ takes one to t
he limit of stability towards full replica symmetry breaking i.e. $E_c(h_r)$, just as we found for $h_r=0$. The annealed approximation is tractable but alas it is only an approximation. Nevertheless the studies in Sec. \ref{sec:density} shows that it gi
ves good results for the density of states of the Hessian for the SK model for $h_r < h_{AT}$. The VB model is a mean-field model and one could hope that it too could be understood analytically, but we do not know how this might be achieved. Our numerica
l studies of the density of states of its Hessian indicates that this is very different from that of the SK model. This is probably because for the SK model all the eigenstates are extended, whereas for the VB model, eigenvectors can also be localized. In
fact our results for the VB model are quite similar to those reported for the three dimensional Heisenberg spin glass in a field \cite{baity2015soft}. There seems to be localized states lying in the gap region, all the way down to $\lambda=0$. But underst
anding the VB model analytically is very challenging. \acknowledgements We should like to thank the authors of Ref.~\onlinecite{lupo:16} for an advance copy of their paper and helpful discussions. One of us (MAM) would like to thank Dan Stein for dis
cussions on quenches in Ising systems. AS acknowledges support from the DST-INSPIRE Faculty Award [DST/INSPIRE/04/2014/002461]. JY was supported by Basic
Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2014R1A1A2053362). \begin{widetext}
\section{Introduction} \label{sec:intro} The production of hadrons and jets at a future Electron Ion Collider (EIC) will play a central role in understanding the structure of the protons and nuclei which comprise the visible matter in the universe. Measu
rements of inclusive jet and hadron production with transversely polarized protons probe novel phenomena within the proton such as the Sivers function~\cite{Kang:2011jw}, and address fundamental questions concerning the validity of QCD factorization. Even
t shapes in jet production can give insight into the nuclear medium and its effect on particle propagation~\cite{Kang:2012zr}. The precision study of these processes at a future EIC will provide a much sharper image of proton and nucleus structure than is
currently available. Progress is needed on both the experimental and theoretical fronts to achieve this goal. Currently, much of our knowledge of proton spin phenomena, such as the global fit to helicity-dependent structure functions~\cite{deFlorian:200
8mr}, comes from comparison to predictions at the next-to-leading order (NLO) in the strong coupling constant. Theoretical predictions at the NLO level for jet and hadron production in DIS suffer from large theoretical uncertainties from uncalculated high
er-order QCD corrections~\cite{Hinderer:2015hra} that will eventually hinder the precision determination of proton structure. In some cases even NLO is unknown, and an LO analysis fails to describe the available data~\cite{Gamberg:2014eia}. Given the hig
h luminosity and expected precision possible with an EIC, it is desirable to extend the theoretical precision beyond what is currently available. For many observables, a prediction to next-to-next-to-leading order (NNLO) in the perturbative QCD expansion
will ultimately be needed. An important step toward improving the achievable precision for jet production in electron-nucleon collisions was taken in Ref.~\cite{Hinderer:2015hra}, where the full NLO ${\cal O}(\alpha^2\alpha_s)$ corrections to unpolarized
$lN \to jX$ and $lN \to hX$ scattering were obtained. Focusing on single-inclusive jet production for this discussion, it was pointed out that two distinct processes contribute: the deep-inelastic scattering (DIS) process $lN \to ljX$, where the final-sta
te lepton is resolved, and $\gamma N \to jX$, where the initial photon is almost on-shell and the final-state lepton is emitted collinear to the initial-state beam direction. Both processes were found to contribute for expected EIC parameters, and the shi
ft of the leading-order prediction was found to be both large and dependent on the final-state jet kinematics. Our goal in this manuscript is to present the full ${\cal O}(\alpha^2\alpha_s^2)$ NNLO contributions to single-inclusive jet production in elect
ron-nucleon collisions, including all the relevant partonic processes discussed above. Achieving NNLO precision for jet and hadron production is a formidable task. The relevant Feynman diagrams which give rise to the NNLO corrections consist of two-loop
virtual corrections, one-loop real-emission diagrams, and double-real emission contributions. Since these three pieces are separately infrared divergent, some way of regularizing and canceling these divergences must be found. However, theoretical techniq
ues for achieving this cancellation in the presence of final-state jets have seen great recent progress. The introduction of the $N$-jettiness subtraction scheme for higher order QCD calculations~\cite{Boughezal:2015dva,Gaunt:2015pea} has lead to the firs
t complete NNLO descriptions of jet production processes in hadronic collisions. During the past year several NNLO predictions for processes with final-state jets have become available due to this theoretical advance~\cite{Boughezal:2015dva,Boughezal:2015
aha,Boughezal:2015ded,Boughezal:2016dtm,Boughezal:2016isb,Boughezal:2016yfp}. In some cases the NNLO corrections were critical in explaining the observed data~\cite{Boughezal:2016yfp}. We discuss here the application of the $N$-jettiness subtraction sche
me to inclusive jet production in electron-proton collisions. Our result includes both the DIS and photon-initiated contributions, and allows arbitrary selection cuts to be imposed on the final state. Upon integration of the DIS terms over the final-sta
te hadronic phase space we compare our result against the known NNLO prediction for the inclusive structure function, and we find complete agreement. We present phenomenological results for proposed EIC parameters. We find that all partonic channels, inc
luding new ones that first appear at this order, contribute in a non-trivial way to give the complete NNLO correction. We note that the NNLO corrections to similar processes, massive charm-quark production in deeply inelastic scattering and dijet producti
on, were recently obtained~\cite{Berger:2016inr,Currie:2016ytq}. \section{Lower-order results}\label{sec:low} We begin by discussing our notation for the hadronic and partonic cross sections, and outlining the expressions for the LO and NLO cross sectio
ns. We will express the hadronic cross section in the following notation: \begin{equation} {\rm d}\sigma = {\rm d}\sigma_{\text{LO}}+{\rm d}\sigma_{\rm NLO}+{\rm d}\sigma_{\rm NNLO}+\ldots \,, \end{equation} where the ellipsis denotes neglected higher-ord
er terms. The LO subscript refers to the ${\cal O}(\alpha^2)$ term, the NLO subscript denotes the ${\cal O}(\alpha^2\alpha_s)$ correction, while the NNLO subscript indicates the ${\cal O}(\alpha^2\alpha_s^2)$ contribution. For the partonic cross sections
, we introduce superscripts that denote the powers of both $\alpha$ and $\alpha_s$ that appear. For example, the leading quark-lepton scattering process is expanded as \begin{equation} {\rm d}\hat{\sigma}_{ql}= {\rm d}\hat{\sigma}_{ql}^{(2,0)}+{\rm d}\ha
t{\sigma}_{ql}^{(2,1)}+{\rm d}\hat{\sigma}_{ql}^{(2,2)}+\ldots \,. \end{equation} Here, the ${\rm d}\hat{\sigma}_{ql}^{(2,0)}$ denotes the ${\cal O}(\alpha^2)$ correction, while ${\rm d}\hat{\sigma}_{ql}^{(2,1)}$ indicates the ${\cal O}(\alpha^2\alpha_s)$
term. The leading-order hadronic cross section can be written as a convolution of parton distribution functions with a partonic cross section, \begin{eqnarray} \label{eq:sigLO} {\rm d}\sigma_{\text{LO}} &=& \int \frac{{\rm d} \xi_1}{\xi_1} \frac{{\rm d
} \xi_2}{\xi_2} \sum_q \left[ f_{q/H}(\xi_1) f_{l/l}(\xi_2) {\rm d}\hat{\sigma}_{ql}^{(2,0)} \right. \\ \nonumber &+& \left. f_{\bar{q}/H}(\xi_1) f_{l/l}(\xi_2) {\rm d}\hat{\sigma}_{\bar{q}l}^{(2,0)}\right]. \end{eqnarray} Here, $f_{q/H}(\xi_1)$ is the
usual parton distribution function that describes the distributions of a quark $q$ in the hadron $H$ carrying a fraction $\xi_1$ of the hadron momentum. $f_{l/l}(\xi_2)$ is the distribution for finding a lepton with momentum fraction $\xi_2$ inside the or
iginal lepton. At leading order this is just $f_{l/l}(\xi_2)=\delta(1-\xi_2)$, but it is modified at higher orders in the electromagnetic coupling by photon emission. $d\hat{\sigma}_{ql}^{(2,0)}$ is the differential partonic cross section. At leading o
rder only the partonic channel $q(p_1)+l(p_2) \to q(p_3)+l(p_4)$ and the same process with anti-quarks instead contribute. The relevant Feynman diagram is shown in Fig.~\ref{fig:LOdiag}. It is straightforward to obtain these terms. \begin{figure}[h] \ce
ntering \includegraphics[width=1.5in]{LO.pdf}% \caption{Feynman diagram for the leading-order process $q(p_1)+l(p_2) \to q_(p_3)+l(p_4)$. We have colored the photon line red, the lepton lines green and the quark lines black.} \label{fig:LOdiag} \end{figur
e} At the next-to-leading order level several new contributions first occur. The quark-lepton scattering channel that appears at LO receives both virtual and real-emission corrections that are separately infrared divergent. We use the antennae subtract
ion method~\cite{Kosower:1997zr} to regularize and cancel these divergences. Initial-state collinear divergences are handled as usual by absorbing them into the PDFs via mass factorization. A gluon-lepton scattering channel also contributes at this orde
r. The collinear divergences that appear in these contributions are removed by mass factorization. Example Feynman diagrams for these processes are shown in Fig.~\ref{fig:NLOlepdiag}. \begin{figure}[h] \centering \includegraphics[width=3.0in]{NLOlep.pdf
}% \caption{Representative Feynman diagrams contributing to the following perturbative QCD corrections at NLO: virtual corrections to the $q(p_1)+l(p_2) \to q(p_3)+l(p_4)$ process (left); real emission correction $q(p_1)+l(p_2) \to q(p_3)+l(p_4)+g(p_5)$ (
middle); the process $g(p_1)+l(p_2) \to q_(p_3) +l(p_4)+\bar{q}(p_5)$ (right). We have colored the photon line red, the lepton lines green, the gluon lines blue and the quark lines black.} \label{fig:NLOlepdiag} \end{figure} The processes discussed abov
e exhaust the possible NLO contributions when the final-state lepton is observed. However, for single-inclusive jet production a kinematic configuration is allowed where the $t$-channel photon is nearly on-shell, and the final-state lepton travels down t
he beam pipe. The transverse momentum of the leading jet is balanced by the additional jet present in these diagrams. This kinematic configuration leads to a QED collinear divergence for vanishing lepton mass, since the photon can become exactly on-shell
in this limit. While it is regulated by the lepton mass, it is more convenient to obtain these corrections by introducing a photon distribution function in analogy with the usual parton distribution function. The collinear divergences that appear in the